阅读说明：本技术 利用人工智能的时间序列数据学习及分析方法 (Time series data learning and analyzing method using artificial intelligence ) 是由 吴圭三 田恩周 权纯焕 孙炯官 尹用根 金珉洙 吕泫珠 于 2019-03-27 设计创作，主要内容包括：本发明提供利用人工智能的时间序列数据的学习及分析方法。根据本发明一实施例的时间序列数据分析方法,由计算装置执行,其可以包括：对于上述时间序列数据在时间轴上被分割的多个单位中的各个单位,将各个上述单位的特征(feature)输入到中间人工神经网络的步骤；从上述中间人工神经网络获得m(m为2以上的自然数)维的中间输出数据的步骤；将时间上紧邻的多个单元的上述中间输出数据输入到最终人工神经网络,获得从上述最终人工神经网络输出的最终输出数据的步骤；及利用上述最终输出数据生成上述时间序列数据的分析结果的步骤。(The invention provides a learning and analyzing method using artificial intelligence time series data. The time series data analysis method according to an embodiment of the present invention is executed by a computing device, and may include: inputting a feature (feature) of each of a plurality of units into an intermediate artificial neural network, the units being obtained by dividing the time-series data on a time axis; obtaining intermediate output data of m (m is a natural number of 2 or more) dimensions from the intermediate artificial neural network; inputting the intermediate output data of a plurality of temporally adjacent units to a final artificial neural network to obtain final output data output from the final artificial neural network; and generating an analysis result of the time-series data by using the final output data.)

1. A time-series data analysis method executed by a computing device, comprising:

a step of inputting, for each of a plurality of cells into which the time-series data is divided on a time axis, a feature of each of the cells into an intermediate artificial neural network;

a step of obtaining m-dimensional intermediate output data from the intermediate artificial neural network, wherein m is a natural number of 2 or more;

a step of inputting the intermediate output data of a plurality of units that are temporally adjacent to each other into a final artificial neural network, and obtaining final output data output from the final artificial neural network; and

and generating an analysis result of the time-series data by using the final output data.

2. The time-series data analysis method according to claim 1,

the intermediate artificial neural network and the final artificial neural network are recurrent neural networks.

3. The time-series data analysis method according to claim 1,

and m nodes of the input layer of the final artificial neural network are provided.

4. The time-series data analysis method according to claim 1,

the step of obtaining the intermediate output data comprises:

a step in which a first computing device inputs features of a first cell included in the plurality of cells to the intermediate artificial neural network embodied in the first computing device, and obtains m-dimensional intermediate output data from the intermediate artificial neural network; and

a step in which a second computing device inputs features of a second cell included in the plurality of cells to an intermediate artificial neural network embodied in the second computing device, and obtains m-dimensional intermediate output data from the intermediate artificial neural network.

5. The time-series data analysis method according to claim 4,

the step of obtaining said final output data comprises,

and a step in which a third computing device receives the intermediate output data of the first unit from the first computing device, receives the intermediate output data of the second unit from the second computing device, and sequentially inputs the intermediate output data of the first unit and the intermediate output data of the second unit to the final artificial neural network, wherein the first unit and the second unit are immediately adjacent in time.

6. The time-series data analysis method according to claim 1,

the intermediate artificial neural network consists of k levels, wherein k is a natural number more than 2,

the step of inputting to the intermediate artificial neural network comprises:

inputting the characteristics of each of the plurality of cells into a level 1 intermediate artificial neural network;

inputting level i output data obtained from a level i intermediate artificial neural network into a level i +1 intermediate artificial neural network, wherein i is a natural number of 1 or more; and

a step of obtaining output data of the level i +1 from the intermediate artificial neural network of the level i +1,

wherein the intermediate output data is data output from a level k intermediate artificial neural network.

7. The time-series data analysis method according to claim 1,

the time series data is electrocardiogram data.

8. The time-series data analysis method according to claim 7,

the analysis result is an electrocardiogram diagnosis result.

9. The time-series data analysis method according to claim 7,

the n units are beat units divided by an R peak value.

10. The time-series data analysis method according to claim 9,

the n divided beat units are,

after the time-series data are segmented, the similarity between the segmented beats is compared, and a beat having a similarity to an existing specific beat equal to or greater than a preset value is substituted as the feature unit.

11. A time-series data learning method executed by a computing device, comprising:

a step of inputting, for each of a plurality of cells into which the time-series data is divided on a time axis, a feature of each of the cells into an intermediate artificial neural network;

a step of obtaining m-dimensional intermediate output data from the intermediate artificial neural network, wherein m is a natural number of 2 or more;

a step of inputting the intermediate output data of a plurality of units that are temporally adjacent to each other into a final artificial neural network, and obtaining final output data output from the final artificial neural network; and

a step of comparing the final output data with a tag tagged to the time-series data.

12. The time-series data learning method according to claim 11,

the intermediate artificial neural network and the final artificial neural network are recurrent neural networks.

13. The time-series data learning method according to claim 11,

and m nodes of the input layer of the final artificial neural network are provided.

14. The time-series data learning method according to claim 11,

the step of comparing includes the step of comparing the final output data with tags tagged to the time series data and calculating an error,

after the step of comparing, the time series data learning method further includes a step of back-propagating the error, and adjusting the weight of the intermediate artificial neural network and the weight of the final artificial neural network.

15. The time-series data learning method according to claim 14,

the step of adjusting the weight values may include,

and setting the learning rate of the intermediate artificial neural network to be 1/n times of the final artificial neural network to adjust the weight of the intermediate artificial neural network.

16. The time-series data learning method according to claim 11,

the intermediate artificial neural network consists of k levels, wherein k is a natural number more than 2,

the step of inputting to the intermediate artificial neural network comprises:

inputting the characteristics of each of the plurality of cells into a level 1 intermediate artificial neural network;

inputting level i output data obtained from a level i intermediate artificial neural network into a level i +1 intermediate artificial neural network, wherein i is a natural number of 1 or more; and

a step of obtaining output data of the level i +1 from the intermediate artificial neural network of the level i +1,

wherein the intermediate output data is data output from a level k intermediate artificial neural network.

17. The time-series data learning method according to claim 11,

the time series data is electrocardiogram data.

18. The time-series data learning method according to claim 17,

the n units are beat units divided by an R peak value.

19. A time-series data analysis apparatus, comprising:

a network interface that receives the time series data;

one or more processors;

a memory to which a computer program executed by the processor is loaded; and

a memory storing the computer program,

the computer program includes:

a command for inputting, for each of a plurality of cells into which the time-series data is divided on a time axis, a feature of each of the cells into an intermediate artificial neural network, and obtaining intermediate output data in m-dimensions from the intermediate artificial neural network, wherein m is a natural number of 2 or more;

a command to input the intermediate output data of a plurality of cells that are temporally proximate to a final artificial neural network, obtaining final output data output from the final artificial neural network; and

and a command for generating an analysis result of the time-series data using the final output data.

20. A time-series data learning apparatus characterized by comprising:

a network interface that receives the time series data;

one or more processors;

a memory to which a computer program executed by the processor is loaded; and

a memory storing the computer program,

the computer program includes:

a command for inputting, for each of a plurality of cells into which the time-series data is divided on a time axis, a feature of each of the cells into an intermediate artificial neural network, and obtaining intermediate output data in m-dimensions from the intermediate artificial neural network, wherein m is a natural number of 2 or more;

a command to input the intermediate output data of a plurality of cells that are temporally proximate to a final artificial neural network, obtaining final output data output from the final artificial neural network; and

a command to compare the final output data with tags tagged to the time series data.

Technical Field

The invention relates to a learning and analyzing method of time series data by using artificial intelligence. More particularly, the present invention relates to a method of performing learning and analysis of time series data using a plurality of artificial neural networks.

Background

As information digitization and data storage technologies have been developed to accumulate a large amount of data, machine learning techniques have been introduced and applied in various fields. Machine learning is a technique that analyzes a large amount of input data, classifies objects according to probability, or predicts values within a specific range. Machine learning does not derive a result value by a specific rule, but operates by empirically analyzing a large amount of input data and deriving a result value according to probability.

However, learning and analyzing time-series data using artificial intelligence has a problem in that the time-series data needs to be compared with image data or the like, and the data length of processing may be long. Taking electrocardiographic data as an example, it is necessary to comprehensively learn and analyze electrocardiographic data measured for 24 hours or more to derive a diagnosis result, which requires a lot of learning and analysis time, and initial input data is gradually diluted and blurred, so that there is a problem in accuracy in learning and analysis of long-term data.

Disclosure of Invention

Problems to be solved by the invention

The technical problem to be solved by the present invention is to provide a method and an apparatus for reducing the computation amount of learning and analyzing time series data by artificial intelligence.

Another technical problem to be solved by the present invention is to provide a method and an apparatus that can improve the accuracy of time series data learning and analysis using artificial intelligence.

The technical problem to be solved by the present invention is to provide a method and apparatus for reducing the amount of computation for learning and analyzing an Electrocardiogram (ECG) using artificial intelligence.

The technical problem to be solved by the present invention is to provide a method and apparatus for improving the accuracy of learning and analysis by artificial intelligence Electrocardiogram (ECG).

The technical problems of the present invention are not limited to the above-mentioned technical problems, and other technical problems not mentioned can be clearly understood by those skilled in the art through the following description.

Means for solving the problems

In order to solve the above technical problem, a time series data analysis method according to an embodiment of the present invention is executed by a computing device, and may include: inputting a feature (feature) of each of a plurality of cells into an intermediate artificial neural network, the plurality of cells being obtained by dividing the time-series data on a time axis; obtaining intermediate output data of m (m is a natural number of 2 or more) dimensions from the intermediate artificial neural network; inputting the intermediate output data of a plurality of temporally adjacent units to a final artificial neural network to obtain final output data output from the final artificial neural network; and generating an analysis result of the time-series data by using the final output data.

In an embodiment, the intermediate artificial Neural Network and the final artificial Neural Network may be a Recurrent Neural Network (RNN).

In one embodiment, the number of neurons in the input layer of the final artificial neural network may be m.

In an embodiment, the step of obtaining the intermediate output data may include: a step in which a first computing device inputs features of a first cell included in the plurality of cells into the intermediate artificial neural network embodied in the first computing device, and obtains m-dimensional intermediate output data from the intermediate artificial neural network; and a step in which the second computing device inputs the characteristics of a second cell included in the plurality of cells into an intermediate artificial neural network embodied in the second computing device, and obtains m-dimensional intermediate output data from the intermediate artificial neural network.

In one embodiment, the step of obtaining the final output data may include a step of receiving, by a third computing device, intermediate output data of the first unit from the first computing device, receiving intermediate output data of the second unit from the second computing device, and sequentially inputting the intermediate output data of the first unit and the intermediate output data of the second unit to the final artificial neural network, wherein the first unit and the second unit are temporally adjacent units.

In one embodiment, the intermediate artificial neural network is composed of k (k is a natural number of 2 or more) levels, and the step of inputting the intermediate artificial neural network includes: inputting the feature (feature) of each of the plurality of cells into a level 1 intermediate artificial neural network; inputting level i output data obtained from a level i (i is a natural number of 1 or more) intermediate artificial neural network into a level i +1 intermediate artificial neural network; and obtaining output data of the level i +1 from the intermediate artificial neural network of the level i +1, wherein the intermediate output data can be the output data of the intermediate artificial neural network of the level k.

In an embodiment, the time-series data may be electrocardiogram data.

In one embodiment, the analysis result may be an electrocardiogram diagnosis result.

In one embodiment, the n cells may be beat cells divided by an R-peak (R-peak).

In one embodiment, the n divided beats may be a beat in which the similarity to the current specific beat is equal to or greater than a predetermined value, and the feature unit is replaced with a beat in which the similarity to the current specific beat is equal to or greater than a predetermined value, after the time-series data is divided.

In order to solve the above technical problem, a time-series data learning method according to another embodiment of the present invention is executed by a computing device, and may include: inputting a feature (feature) of each of a plurality of cells into an intermediate artificial neural network, the plurality of cells being obtained by dividing the time-series data on a time axis; obtaining m (m is a natural number of 2 or more) intermediate output data from the intermediate artificial neural network; inputting the intermediate output data of a plurality of temporally adjacent units to a final artificial neural network to obtain final output data output from the final artificial neural network; and comparing the final output data with the tag marked on the time series data.

In an embodiment, the intermediate artificial Neural Network and the final artificial Neural Network may be a Recurrent Neural Network (RNN).

In one embodiment, the comparing step includes comparing the final output data with a tag marked on the time-series data and calculating an error, and the time-series data learning method may further include a step of adjusting a weight of the intermediate artificial neural network and a weight of the final artificial neural network by back-propagating the error after the comparing step.

In an embodiment, the step of adjusting the weight may include the step of adjusting the weight of the intermediate artificial neural network by setting a learning rate (learning rate) of the intermediate artificial neural network to be 1/n times of the final artificial neural network.

In one embodiment, the intermediate artificial neural network is composed of k (k is a natural number of 2 or more) levels, and the step of inputting the intermediate artificial neural network includes: inputting the characteristics of each unit in the plurality of units into a level 1 intermediate artificial neural network; inputting level i output data obtained from a level i (i is a natural number of 1 or more) intermediate artificial neural network into a level i +1 intermediate artificial neural network; and obtaining output data of the level i +1 from the intermediate artificial neural network of the level i +1, wherein the intermediate output data can be the output data of the intermediate artificial neural network of the level k.

In an embodiment, the time-series data may be electrocardiogram data.

In one embodiment, the n units may be pulse units divided by R peak values.

A time-series data analysis apparatus according to another embodiment of the present invention to solve the above-described technical problem includes: a network interface that receives the time series data; one or more processors; a memory (memory) to which a computer program executed by the processor is loaded; and a memory (storage) storing the computer program, the computer program may include: a command for inputting a feature (feature) of each of a plurality of cells into an intermediate artificial neural network, the plurality of cells being obtained by dividing the time-series data on a time axis, and obtaining intermediate output data in m (m is a natural number equal to or greater than 2) dimensions from the intermediate artificial neural network; a command for inputting the intermediate output data of the plurality of units temporally adjacent to each other to a final artificial neural network and obtaining final output data output from the final artificial neural network; and a command for generating an analysis result of the time-series data by using the final output data.

A time-series data learning apparatus according to another embodiment of the present invention to solve the above-described technical problem includes: a network interface that receives the time series data; one or more processors; a memory (memory) to which a computer program executed by the processor is loaded; and a memory (storage) storing the computer program, the computer program may include: a command for inputting a feature (feature) of each of a plurality of cells into an intermediate artificial neural network, the plurality of cells being obtained by dividing the time-series data on a time axis, and obtaining intermediate output data in m (m is a natural number equal to or greater than 2) dimensions from the intermediate artificial neural network; a command for inputting the intermediate output data of the plurality of units adjacent in time to a final artificial neural network to obtain final output data output from the final artificial neural network; and a command for comparing the final output data with the tag marked in the time-series data.

A recording medium of another embodiment of the present invention to solve the above technical problem may be a computer-readable recording medium storing a computer program for executing: inputting a feature (feature) of each of a plurality of cells into an intermediate artificial neural network, the cells being divided in time series data on a time axis; obtaining intermediate output data of m (m is a natural number of 2 or more) dimensions from the intermediate artificial neural network; inputting the intermediate output data of a plurality of temporally adjacent units to a final artificial neural network to obtain final output data output from the final artificial neural network; and generating an analysis result of the time-series data by using the final output data.

Effects of the invention

According to the various embodiments of the present invention, when time series data learning and analysis using artificial intelligence is performed, the amount of calculation can be reduced and the accuracy can be improved.

According to the various embodiments of the present invention, when learning and analyzing electrocardiogram data using artificial intelligence is performed, the amount of calculation can be reduced and the accuracy can be improved.

The effects of the present invention are not limited to the above-mentioned effects, and other effects not mentioned can be clearly understood by the ordinary skilled person through the following description.

Drawings

Fig. 1a and 1b are diagrams illustrating a wearable electrocardiogram measuring apparatus and an attaching method thereof.

Fig. 2 is a diagram showing an electrocardiogram analysis system including an electrocardiogram measuring apparatus, a network 210, and a server.

Fig. 3a and 3b are diagrams illustrating a conventional RNN-based electrocardiographic analysis method.

Fig. 4 is a flowchart illustrating an electrocardiogram analysis method using artificial intelligence according to an embodiment of the present invention.

Fig. 5 is a diagram showing an electrocardiographic signal segmentation method according to an embodiment of the present invention.

Fig. 6a to 6c are diagrams illustrating a similar beat processing method according to an embodiment of the present invention.

Fig. 7a and 7b are diagrams illustrating an intermediate artificial neural network and a final artificial neural network according to an embodiment of the present invention.

Fig. 8 to 10 are diagrams illustrating a time-series data analysis method using artificial intelligence according to an embodiment of the present invention.

Fig. 11 is a flowchart illustrating a time-series data analysis method using artificial intelligence according to an embodiment of the present invention.

Fig. 12 is a diagram illustrating when an intermediate artificial neural network is embodied in a plurality of levels in an artificial neural network structure according to an embodiment of the present invention.

Fig. 13 is a flowchart illustrating a time-series data analysis method when an intermediate artificial neural network is embodied in a plurality of levels in the structure of the artificial neural network according to an embodiment of the present invention.

Fig. 14 is a flowchart illustrating an electrocardiogram learning method using artificial intelligence according to an embodiment of the present invention.

Fig. 15 is a diagram illustrating learning data used in an electrocardiogram learning method using artificial intelligence according to an embodiment of the present invention.

Fig. 16 is a diagram illustrating a time-series data learning method using artificial intelligence according to an embodiment of the present invention.

Fig. 17 is a flowchart illustrating a time-series data learning method using artificial intelligence according to an embodiment of the present invention.

Fig. 18 is a hardware configuration diagram of an exemplary computing device that can embody a time-series data analysis method or a time-series data learning method using artificial intelligence according to an embodiment of the present invention.

Detailed Description

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. The advantages and features of the invention, and the methods of embodying them, may be ascertained by reference to the following detailed description of the embodiments taken in conjunction with the accompanying drawings. The present invention is not limited to the embodiments disclosed below, but can be embodied in various ways different from each other. However, the present embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art, and the present invention will only be defined by the appended claims. Like reference numerals refer to like elements throughout the specification.

Unless otherwise defined, all terms (including technical and scientific terms) used herein may be used in the same sense as commonly understood by one of ordinary skill in the art to which this invention belongs. And terms commonly used and defined in dictionaries should not be interpreted abnormally or excessively unless they are explicitly and specifically defined. The terms used in the present specification are used only for illustrating the embodiments and do not limit the present invention. In this specification, the singular forms also include the plural forms unless otherwise specifically mentioned in a sentence.

When the terms "comprises" and/or "comprising" are used in this specification, the presence of structural elements, steps, actions, and/or elements may not preclude the presence or addition of one or more other structural elements, steps, actions, and/or elements.

Before describing the present specification, several terms used in the present specification are clarified.

The invention belongs to the technical field of 'artificial neural network' and relates to a diagram structure consisting of multiple layers and multiple nodes forming each layer, wherein the multiple layers are formed by simulating a biological neural network (particularly human visual/auditory cortex). The artificial neural network comprises an input layer, more than one hidden layer and an output layer.

The 'input layer' in the technical field of the present invention refers to a layer that receives data to be analyzed/learned in a layer structure of an artificial neural network.

The 'output layer' in the technical field of the invention refers to a layer for outputting a result value in a layer structure of an artificial neural network.

The 'hidden layer' in the technical field of the present invention refers to all layers except an input layer and an output layer in a layer structure of an artificial neural network. The artificial neural network is composed of successive layers of neurons, and the neurons of each layer are connected to the neurons of the next layer. If the input layer and the output layer are not directly connected through the hidden layer, each input may independently contribute to the output independently of the other inputs, thereby making it difficult to obtain an accurate result. In fact, the input data affects the output in a complex structure by being interdependent and inter-combined, so by adding a hidden layer, it is possible to capture the subtle interactions between the neurons of the hidden layer affecting the input of the final output. That is, the high-level features or attributes of the data can be treated as a hidden layer.

The present invention belongs to the technical field of 'neuron (neuron)' which is the concept of corresponding neurons in a biological neural network to an artificial neural network. 'neurons' are also referred to as 'nodes'.

The 'weight' in the technical field of the present invention is a concept of corresponding a degree of enhancing synapse (synapse) connection through repeated signal transmission in a biological neural network to an artificial neural network. In the biological neural network, when a signal is frequently transmitted from the neuron 1 to the neuron 2, in order to improve efficiency of signal transmission, a path from the neuron 1 to the neuron 2, that is, synaptic connection is enhanced. This is considered a learning or memory process, which in artificial neural networks is interpreted as 'weights'.

The 'back propagation' in the technical field of the invention is a word derived from 'back propagation of errors', and refers to adjusting the weight of the artificial neural network by the difference between an output value and an actual value, i.e. the back propagation of errors.

The 'gradient descent' method in the technical field of the invention is a method for determining a weight of a learning model during reverse propagation, and is a method for searching a weight which minimizes a difference between an output value of a feedforward function value including the weight and an actual value with a label (label). As an example, when the shape of the loss function shows a parabolic shape, the lowest point can be found to calculate the weight, and the process of finding the lowest point is similar to the pattern of a walking down slope, so the method is named as a gradient descent method.

The learning rate in the field of the present invention refers to the step size (step size) of finding the minimum value in the gradient descent algorithm for adjusting the weights of the artificial neural network. When the learning rate is large, the adjustment speed is high but the accuracy is low, and when the learning rate is small, the accuracy is high but the adjustment speed is low.

The invention belongs to the technical field of 'Convolutional Neural Networks (CNN)' which is a deep learning model combining an artificial Neural network with a filtering technology and enables the artificial Neural network to be optimized to better master the characteristics of input data. Convolutional neural networks show excellent performance in the field of computer vision (computer vision).

The present invention relates to a 'Recurrent Neural Network (RNN)' which is a deep learning model suitable for learning time-varying data (e.g., time series data), and relates to an artificial Neural Network constructed by connecting a Network to a reference time point t and a next time point t + 1. The recurrent neural network is characterized in that the connections between nodes have a recurrent structure, and this structure allows the artificial neural network to efficiently process time-varying features. The recurrent neural network can efficiently process an input in the form of a sequence and exhibits good performance in processing time-series data having time-varying characteristics (e.g., handwriting recognition and speech recognition).

The artificial intelligence related art and terms described above are described in detail in websites such as wikipedia and YouTube, and thus detailed description thereof is omitted. For example, Recurrent neural networks are described in detail in the literature by https:// en.

The "time-series data" in the present specification means data arranged in time series. It is not necessarily continuous in time, and even discrete (discrete) data can be referred to as "time-series data" in this specification as long as they are arranged in time series.

In the present specification, "cell" refers to each individual when time-series data is divided (split) into 2 or more individuals on the time axis.

The 'feature' in the present specification refers to a feature of time-series data or a unit extracted from time-series data or the unit.

The 'intermediate artificial neural network' in the present specification refers to an artificial neural network of an intermediate step (receiving input data and outputting intermediate output data). The intermediate output data is not a value that is finally obtained by using artificial intelligence, unlike the output value of a general artificial neural network.

The 'final artificial neural network' in the present specification refers to an artificial neural network of a final step of receiving intermediate output data from the intermediate artificial neural network and outputting final output data. The final artificial neural network is an artificial neural network located at the tip (terminal) of an artificial intelligence system (composed of a plurality of artificial neural networks according to an embodiment of the present invention), and the final output data output by the final artificial neural network is a value finally obtained by using the artificial intelligence.

The present invention described below relates to a method for learning and analyzing time-series data using artificial intelligence. Specifically, the invention discloses a time series data learning and analyzing method for performing learning and analyzing of time series data by using a plurality of artificial neural networks.

In the present specification, an electrocardiogram is described as an example of time-series data to specifically describe the present invention. Specifically, a method for learning and analyzing time-series data using artificial intelligence according to an embodiment of the present invention will be described by describing a process of using a plurality of artificial neural networks according to an embodiment of the present invention when performing electrocardiogram learning and analysis.

The present invention will be described in detail below with reference to the accompanying drawings.

An electrocardiogram (electrocardiograph) refers to a graph in which the electrical activity of the heart is analyzed and recorded in wavelength form. The electrocardiogram is a graph in which potentials related to cardiac pulsation are graphically recorded on the body surface, and is the most widely used test for diagnosing circulatory diseases.

Electrocardiograms are often used to diagnose heart arrhythmias and coronary artery disease. When arrhythmia occurs intermittently, arrhythmia cannot be diagnosed by electrocardiographic examination only once, and thus it is necessary to obtain an electrocardiogram recorded in daily life. Since it may not be possible to detect an infrequent arrhythmia from a short-time electrocardiogram recording, if an arrhythmia is suspected, an electrocardiogram is measured and recorded for a long time to make an accurate diagnosis.

The conventional electrocardiographic measurement device is a method of measuring by attaching a plurality of electrodes to the body. For example, a Holter monitor (Holter monitor) is generally used in which 5 to 7 electrodes are attached to the chest and a detector weighing about 500g is worn at the waist to measure an electrocardiogram. The prior electrocardiogram measuring device has the following problems: it is large in size and heavy in weight, and has many electrodes, which cause discomfort when attached to the body, hindering free daily life.

To solve these problems, wearable electrocardiographic measuring devices under recent development have reduced the size and weight of their devices and minimized the number of electrodes. The reason for minimizing the number of electrodes is to take the convenience of the user in daily life and the battery performance of the electrocardiogram measuring apparatus into consideration.

Fig. 1a and 1b are diagrams illustrating a wearable electrocardiogram measuring apparatus and an attaching method thereof.

Fig. 1a shows a wearable electrocardiographic measuring device 100, which has 2 electrodes as shown in fig. 1 a.

Fig. 1b shows a method of attaching the wearable electrocardiographic measuring device 100 to the body to measure an electrocardiogram, and as shown in fig. 1b, attaching electrodes around the heart to measure an electrocardiogram.

The wearable electrocardiographic measurement device 100 is more portable than conventional electrocardiographic measurement devices, but the conventional electrocardiographic measurement devices can measure and diagnose a relatively accurate electrocardiogram by analyzing multichannel (multi-channel) electrocardiographic signals with a plurality of electrodes, and compared to this, the wearable electrocardiographic measurement device 100 can only measure 1-channel electrocardiographic signals, and thus has a problem of low accuracy of electrocardiographic measurement and diagnosis.

The wearable electrocardiographic measurement device 100 cannot use an electrocardiographic analysis method used in an electrocardiographic measurement device (e.g., Holter detector) that is conventionally collected by a plurality of channels as it is, and has a difficulty in diagnosing arrhythmia and the like by only 1-channel electrocardiographic signals. Nevertheless, since the wearable electrocardiographic measurement device 100 has advantages such as portability, many studies are being conducted on a method of diagnosing arrhythmia and the like using only 1-channel electrocardiographic signals measured by the wearable electrocardiographic measurement device 100. In particular, with the development of deep learning techniques in the field of artificial intelligence, there is a possibility that a high-accuracy arrhythmia diagnosis can be sufficiently achieved by only one signal.

However, the apparatus for measuring an electrocardiogram of a person to be examined in the present invention is not limited to the wearable electrocardiogram measuring apparatus 100, and any form of electrocardiogram measuring apparatus 100 may be used as long as it can be attached to a body to measure an electrocardiogram. Hereinafter, for convenience of explanation, the present invention will be referred to as an "electrocardiogram measuring apparatus 100" in which an apparatus for measuring an electrocardiogram of a subject is used.

Fig. 2 is a diagram showing an electrocardiogram analysis system including an electrocardiogram measuring apparatus, a network, and a server.

The electrocardiogram measuring apparatus 100 measures only the electrocardiogram of the person to be inspected, and does not provide a self-analysis function in many cases. Therefore, it is preferable that the electrocardiographic measuring device 100 performs only electrocardiographic measurement, and electrocardiographic analysis is performed on the server 200 side. Specifically, it is preferable that the electrocardiogram measuring apparatus 100 measures an electrocardiogram and transmits it to the server 200, and electrocardiogram data is analyzed using artificial intelligence or the like at the server 200 side.

As shown in fig. 2, an electrocardiogram analysis system according to an embodiment of the present invention may include: electrocardiogram measuring apparatus 100, network 210 and server 200.

The person to be examined can perform daily life with the electrocardiographic measuring device 100 attached. The electrocardiographic data measured by the electrocardiographic measuring device 100 is transmitted to the server 200 via the network 210. In this case, the server 200 may be a server of a hospital, a server of a medical service provider, a server of an IT company, a server of a government agency such as a health and welfare department, or the like. Since the electrocardiogram data of an individual is sensitive medical information, the server 200 is preferably a server of a unit permitted to collect and analyze medical information.

The server 200 performs learning and analysis using the electrocardiogram data received from the electrocardiogram measuring apparatus 100. Server 200 may be enabled to learn electrocardiogram data in which electrocardiogram diagnostic results are tagged with tags by medical professionals, such as cardiologists (cardiologists), while learning the electrocardiogram signals.

In the nature of artificial intelligence, learning of artificial neural networks requires a longer time and higher computational power, whereas for applications of artificial neural networks, a shorter time and lower computational power are sufficient. Therefore, the learning of the electrocardiographic data can be performed at the server 200 side, and the analysis of the electrocardiographic data can be performed at the electrocardiographic measuring device 100 itself, at an additional electrocardiographic analyzing apparatus wired/wirelessly connected to the electrocardiographic measuring device 100, and at a smartphone, a tablet computer, a notebook computer, and the like, which are wired/wirelessly connected to the electrocardiographic measuring device 100.

In the present specification, the method for learning and analyzing electrocardiographic data according to an embodiment of the present invention will be described on the premise that the server 200 can learn and analyze electrocardiographic data. However, these descriptions do not exclude that the analysis of the electrocardiographic data is performed in the electrocardiographic measuring device 100 itself, in an additional electrocardiographic analyzing apparatus wired/wirelessly connected to the electrocardiographic measuring device 100, in a smart phone, a tablet computer, a notebook computer, and the like wired/wirelessly connected to the electrocardiographic measuring device 100.

The learning and analysis methods using artificial intelligence electrocardiographic data studied so far are based on Convolutional Neural Networks (CNN). CNN is a deep learning model in which an artificial neural network is combined with a filtering technique, and exhibits excellent performance in the field of computer vision (computer vision) such as image analysis. However, unlike images, electrocardiograms are time-series data having sequential content (sequential context), and thus have limitations in analyzing electrocardiograms using CNN techniques dedicated to image analysis.

The CNN-based electrocardiogram analysis method does not read sequential contents of all electrocardiographic signals, but grasps characteristics of local electrocardiographic signals and combines the characteristics to derive a total electrocardiogram analysis result. Since the ecg analysis method based on CNN analyzes ecg data from a microscopic perspective, not from a macroscopic perspective, when a local ecg signal includes noise, the characteristics of the local ecg signal cannot be well grasped, which adversely affects the accuracy of the determination of the entire ecg signal. Also, when a plurality of features that conflict with each other exist in one signal, the filtered features are not taken into consideration in diagnosis, so there is a problem that the determination accuracy is significantly reduced.

In order to overcome the above problems, studies are currently being made on electrocardiographic analysis methods using RNN.

Fig. 3a and 3b are diagrams illustrating a conventional RNN-based electrocardiographic analysis method.

Fig. 3a shows the electrocardiographic signal measured by the electrocardiographic measuring device 100 for a predetermined time.

In the cardiac signal shown in fig. 3a, the portion of the sharp protrusion is the R peak (R-peak). The R peak value is described in detail in https:// en. wikipedia. org/wiki/QRS _ complete, etc., and thus, the detailed description thereof is omitted.

Fig. 3b shows that the electrocardiographic signal measured by the electrocardiographic measuring device 100 for a predetermined time is input to the RNN in units of 3 beats (beat: means heart beat unit).

The electrocardiographic signal measured in a predetermined time is divided (split) by the beat (beat) unit using the R peak value. The RNN inputs 3 beats in close temporal proximity together because the relationship between the preceding beat, the beat to be determined, and the subsequent beat is important for determining arrhythmia from the electrocardiographic signal. That is, the occurrence interval and the height difference of the preliminary beat, the beat to be determined, and the postnatal beat are important factors for determining the arrhythmia.

When the cardiac signal is divided into 9 beats, and each beat is named A, B, C, D, E, F, G, H and I, A + B + C, B + C + D, C + D + E, D + E + F, E + F + G, F + G + H and G + H + I are input to RNN.

As shown in FIG. 3B, A + B + C, B + C + D, C + D + E, D + E + F, E + F + G, F + G + H and G + H + I are sequentially input to an RNN.

When one beat is sampled at 250Hz, if the electrocardiographic signal of fig. 3a is analyzed by the conventional RNN, it is necessary to process an input having a length of 7 (the number of beats to be determined) × 3 (3 beats adjacent to each other are input) × 250 (the number of samples of each beat) × 5250.

Moreover, since arrhythmia does not always occur, it is necessary to collect and analyze the electrocardiographic signals of the person to be examined for a long time. The healthy adult has a heartbeat rate of 60 to 100 times per minute, which is converted into 86400 to 144000 times per 24 hours. Assuming that the number of beats per minute is 80, the number of beats per 24 hours is 115200. This means that the number of beats measured for 1 minute is 80, and the number of beats measured for 24 hours is 115200.

In order to determine 115200 beats individually, it is necessary to consider both the preceding beat and the following beat of each beat, and therefore, when sampling one beat at 250Hz, the RNN needs to process an input having a length of about 115200 (the number of determination target beats) × 3 (3 beats adjacent to each other are input together) × 250 (the number of samples of each beat) × 86400000.

Compared with electrocardiogram analysis based on RNN, electrocardiogram analysis based on CNN is fast, but because sequential content of electrocardiogram signals cannot be read, analysis accuracy is low. Conversely, the sequential contents of the electrocardiographic signals can be read by the electrocardiographic analysis based on RNN, but it is difficult to ensure the real-time performance of the electrocardiographic analysis because the input length to be processed is too long.

Fig. 4 is a flowchart illustrating an electrocardiogram analysis method using artificial intelligence according to an embodiment of the present invention.

As shown in fig. 4, the electrocardiogram analysis method using artificial intelligence according to an embodiment of the present invention may include: an electrocardiographic signal collection step S411; an electrocardiographic signal preprocessing step S412; r peak detection step S413; an electrocardiographic signal dividing step S421; a significance of beat (significance) determination step S422; a similar beat processing step S430; a feature (feature) extraction step S440; an intermediate artificial neural network analyzing step S450; a final artificial intelligence network analysis step S460; and a step of reporting (report) the judgment result.

The electrocardiographic signal collection step S411 is a step in which the server 200 receives and collects electrocardiographic signals from the electrocardiographic measurement device 100.

The electrocardiographic signal preprocessing step S412 is a step of processing noise, signal interruption, signal interference, and the like mixed in the electrocardiographic signal. Specifically, the electrocardiographic signal received from the electrocardiographic measuring device 100 may be preprocessed by a conventionally known method such as a Band Pass Filter (Band Pass Filter).

The R peak detection step S413 is a step of detecting an R peak from the electrocardiographic signal. The detection of the R peak is not limited to a specific detection algorithm, and various detection algorithms known in the art, such as Pan-Tompkins algorithm (Pan-Tompkins algorithm), Hilbert transform (Hilbert transform), Kathirvel et al (a high efficiency R peak detection based on a new nonlinear transform and a first order Gaussian differentiator, cardiovascular Engineering and Technology (and interference R-peak detection based on a nonlinear transform and a first order Gaussian differentiator), etc., may be used.

The electrocardiographic signal dividing step S421 is a step of dividing the electrocardiographic signal into pulse units by using the detected R peak value.

The beat significance determination step S422 is a step of determining whether or not the divided beats correspond to the actual heart rate using the R peak value. For example, when sudden noise is erroneously determined as an R peak, a beat divided based on the erroneously determined R peak does not correspond to the actual heart rate.

Human heart beats have the property of maintaining a similar length, so beats that are too short or too long can be considered to have been falsely detected in the R peak detection step. Specifically, when the beat divided by the R peak is statistically too long (e.g., more than 99 percentile) or too short (e.g., less than 1 percentile), it may be considered that an erroneous detection has occurred in the R peak detection step, and at this time, the R peak needs to be re-detected. For the algorithm used for detecting the R peak, another algorithm different from the previously used algorithm may be used, or parameters of the previously used algorithm may be modified and used, and for the beats that are statistically too long or statistically too short, different algorithms may be applied respectively. The reason for re-detecting the R peak is that the R peak detection algorithm cannot guarantee 100% accuracy, and especially the wearable device may generate much noise when detecting and transmitting data.

When the divided beats are not statistically significant in the beat significance determination step S422 ('no' in step S422), another suitable R peak detection algorithm is selected (step S423), and the R peak is redetected using the selected R peak detection algorithm (step S424). The electrocardiographic signal is divided into beat units again by the R peak value detected by the selected R peak value detection algorithm. The selection of the R-peak detection algorithm and the R-peak detection process are repeated until the divided beats are determined to be statistically significant.

When the divided beats are statistically significant in the beat significance determination step S422 ('yes' in step S422), the similar beat processing step S430 is executed. The similar beat processing step S430 is a step of replacing the current beat with the previously occurring beat when there is a beat similar to the previously occurring beat among the previously occurring beats. Although the similar beat processing step S430 is not essential, the amount of computation by the computer can be reduced by the similar beat processing step S430.

The feature extraction step S440 is a step of extracting features of each beat. The extracted features may be obtained by using a variety of methods such as Fast Fourier Transform (FFT), Mel Frequency Cepstral Coefficient (MFCC), Filter Bank (Filter Bank), Wavelet (Wavelet), and the like, which are widely used in signal processing.

The intermediate artificial neural network analyzing step S450 is a step of inputting the characteristics of each beat to the intermediate artificial neural network and obtaining intermediate output data. The intermediate output data is not a value that is finally obtained by using artificial intelligence, unlike the output value of a general artificial neural network. The intermediate artificial neural network and the intermediate artificial neural network analyzing step S450 will be described in detail with reference to fig. 7a to 12.

The final artificial neural network analyzing step S460 is a step of receiving intermediate output data from the intermediate artificial neural network, thereby obtaining final output data. The final artificial neural network is an artificial neural network located at the tip (terminal) of an artificial intelligence system (composed of a plurality of artificial neural networks according to an embodiment of the present invention), and the final output data output by the final artificial neural network is a value finally obtained by using the artificial intelligence. The final artificial neural network and the final artificial neural network analyzing step S460 will be described in detail later with reference to fig. 7a to 12.

The judgment result reporting step S470 is a step of generating and reporting a diagnosis result using the final output data obtained in the final artificial neural network analyzing step S460. Specifically, in the step S470, the judgment result is reported after the electrocardiogram signal is analyzed to judge whether the person to be examined has arrhythmia.

Fig. 5 is a diagram showing an electrocardiographic signal segmentation method according to an embodiment of the present invention.

As shown in fig. 5, the electrocardiographic signal can be divided into beat units based on the R peak value. Specifically, the left side of the R peak value may be divided in the cardiac signal, and the right side of the R peak value may be divided. When the electrocardiographic signal is divided based on the R peak value, the position of the divided electrocardiographic signal may be an arbitrary position. The waveform of a typical electrocardiogram is classified into P-QRS-T, which is reflected in fig. 5, and the electrocardiogram signal is divided in such a manner that the R peak is located at the center of each beat.

Fig. 6a to 6c are diagrams illustrating a similar beat processing method according to an embodiment of the present invention.

In order to determine arrhythmia using an electrocardiographic signal, it is necessary to collect data for a long time. Arrhythmia is characterized by an irregular occurrence and therefore its point in time of occurrence cannot be predicted. Therefore, in order to measure an electrocardiogram for a long time (several days), the person to be examined is attached with the electrocardiogram measuring apparatus 100 in the form of a wearable apparatus.

When the electrocardiographic signals measured by the electrocardiographic measurement device 100 are transmitted to the server 200 and the server 200 analyzes the electrocardiographic signals, if similar beats exist among the already-generated beats before data is input to the artificial neural network, only the previously-generated beats can be analyzed and the results can be used together, thereby increasing the analysis speed.

As shown in fig. 6a, the similarity is calculated by comparing the current beat K with the previously generated beats A, B, C, D, E, F, G, H, I and J. Calculating the similarity between the current pulse to be judged and N pre-generated pulses, and selecting the most similar pulse as a substitute pulse from the pre-generated pulses with the similarity being more than a critical value.

If the most similar beat is the beat G among the previously generated beats having a similarity to the current beat K equal to or greater than the threshold value (see fig. 6b), the beat G is placed at the position where the beat K is placed instead of the beat K. When there is no similar beat, beat K is maintained.

Since only N pieces of previous beat data are held, the situation as shown in fig. 6c occurs when the similarity of the next beat is determined.

Using the thus generated beat connection information based on the similarity, the analysis result of the beat G is shared without performing additional analysis on the beat K. The connection information generated in this step is used to recover the result of the beat K after the diagnosis is completed. Any one of known similarity calculation methods such as L1, L2, and DTW may be used as the similarity calculation algorithm. Since the electrocardiogram of a human tends to have a strong tendency to repeat similar signals, the amount of calculation can be greatly reduced by the beat filtering in the similarity method as described above.

Fig. 7a and 7b are diagrams illustrating an intermediate artificial neural network and a final artificial neural network according to an embodiment of the present invention.

As described above, according to the electrocardiographic analysis method of the embodiment of the present invention, after the electrocardiographic signal is divided into the beat units by the R peak value, the features of each beat are extracted, and the features of each beat are input to the intermediate artificial neural network.

FIG. 7a is a diagram illustrating an intermediary artificial neural network, according to an embodiment of the present invention. As shown in fig. 7a, when feature extraction of t0, t1, …, tv is performed for any one beat, for example, for beat a, v +1 features of beat a are input to v +1 nodes located at the input layer of the intermediate artificial neural network.

The data input into the intermediate artificial neural network is output to m nodes positioned in an output layer through a hidden layer. The values output to the m nodes located at the output layer of the intermediate artificial neural network may represent the characteristics of beat a. Note that the values output to the m nodes in the output layer of the intermediate artificial neural network are not values that are ultimately obtained by comprehensively analyzing the cardiac electrical signal using artificial intelligence, but rather values that are output to any one beat, for example, the features of beat a are displayed in m dimensions. The intermediate artificial neural network may learn beat a by directing the model, at which time the values output to the m nodes at the output layer of the intermediate artificial neural network may be some predicted value or classification value for beat a. The intermediate artificial neural network may also learn beat a through a non-guided model, and at this time, the values output to the m nodes located in the output layer of the intermediate artificial neural network may be values for clustering (clustering) beat a. In any case, the values output to the m nodes located at the output layer of the intermediate artificial neural network simply show the features of beat a as m dimensions, rather than the values that would ultimately be obtained using artificial intelligence. At this time, the number of extracted features is preferably 2 or more. That is, m is preferably a natural number of 2 or more.

In one aspect, fig. 7a shows that the intermediate artificial neural network is an RNN structure, but the intermediate artificial neural network of fig. 7a is not limited to the RNN structure, and may be a CNN structure or other artificial neural network structures.

FIG. 7b is a diagram illustrating a final artificial neural network, according to an embodiment of the present invention. As shown in fig. 7b, the values output by the m nodes of the output layer of the intermediate artificial neural network are input to the m nodes located at the input layer of the final artificial neural network.

The final artificial neural network may be an RNN, and in this case, in order to reflect the time-varying characteristic of the continuous beats, the m-dimensional characteristic value of the beat to be determined, and the m-dimensional characteristic value of the beat to be determined may be sequentially input to the input layer of the final artificial neural network.

As shown in fig. 3a and 3B, when the cardiac electric signal is divided into 9 beats (beats), each beat is named A, B, C, D, E, F, G, H and I, and the beat to be determined, the previous beat, and the next beat are input together and analyzed, the m-dimensional feature value of the beat a, the m-dimensional feature value of the beat B, and the m-dimensional feature value of the beat C can be sequentially and continuously input to the final artificial neural network. When the entire cardiac signal is divided into 9 beats (beats), and each beat is named A, B, C, D, E, F, G, H and I, the data can be input to the final artificial neural network in the following order.

(1) A feature value for the m-dimension of beat a, a feature value for the m-dimension of beat B, and a feature value for the m-dimension of beat C.

(2) A feature value for the m-dimension of beat B, a feature value for the m-dimension of beat C, and a feature value for the m-dimension of beat D.

(3) A feature value for the m-dimension of beat C, a feature value for the m-dimension of beat D, and a feature value for the m-dimension of beat E.

(4) A feature value for the m-dimension of the beat D, a feature value for the m-dimension of the beat E, and a feature value for the m-dimension of the beat F.

(5) A feature value for the m-dimension of the beat E, a feature value for the m-dimension of the beat F, and a feature value for the m-dimension of the beat G.

(6) A feature value for the m-dimension of the beat F, a feature value for the m-dimension of the beat G, and a feature value for the m-dimension of the beat H.

(7) A feature value for the m-dimension of the beat G, a feature value for the m-dimension of the beat H, and a feature value for the m-dimension of the beat I.

And the data input into the final artificial neural network is output to c nodes positioned in an output layer through the hidden layer.

Finally, the artificial neural network can learn all electrocardiosignals through the guide model. At this time, the values output to the c nodes located in the output layer of the final artificial neural network may be values for classifying all the electrocardiographic signals into c classifications.

As shown in fig. 7b, the final artificial neural network preferably reflects the time-varying characteristic RNN structure of the data well. The RNN structure has a feature that the connection between nodes has a cyclic structure, and can effectively process time-varying features, and the final artificial neural network is preferably configured to sequentially receive m-dimensional features of consecutive beats and classify all electrocardiographic signals (as a set of consecutive beats) into c pieces. However, the final artificial neural network is not limited to the RNN structure, and may be another artificial neural network structure that reflects the time-varying characteristics of the time-series data. That is, the final artificial neural network is not limited to the RNN structure.

Fig. 8 to 10 are diagrams illustrating a time-series data analysis method using artificial intelligence according to an embodiment of the present invention.

Fig. 8 shows analysis of an electrocardiographic signal using the time-series data analysis method using artificial intelligence according to an embodiment of the present invention, but as described above, this is merely for convenience of explanation, and the time-series data analysis method using artificial intelligence according to an embodiment of the present invention is not limited to electrocardiographic signal analysis.

A process of analyzing time-series data (e.g., electrocardiographic information) using artificial intelligence according to an embodiment of the present invention is described with reference to fig. 8.

When the electrocardiographic signal 810 is input, the electrocardiographic signal is divided into a plurality of beats 821, 822, 823 … according to a predetermined criterion. For example, after detecting the R peak from the electrocardiographic signal 810, the electrocardiographic signal can be divided into a plurality of beats 821, 822, 823 … by the R peak.

This is illustrated generically as time series data, which can be described as: when time-series data is input, the time-series data is divided into a plurality of units (units) according to a predetermined criterion.

Then, an operation of extracting features 831, 832, 833 … of the respective beats is performed for the plurality of beats 821, 822, 823 …, respectively. The extracted features may be obtained by using a variety of methods such as Fast Fourier Transform (FFT), Mel Frequency Cepstral Coefficient (MFCC), Filter Bank (Filter Bank), Wavelet (Wavelet), and the like, which are widely used in signal processing.

This is illustrated generically as time series data, which can be described as: when time series data is divided into a plurality of units (units), features of the respective units are extracted for the plurality of units.

Then, an operation of inputting the features 831, 832, 833 … of the respective beats to the intermediate artificial neural networks 841, 842, 843 … is performed. If the electrocardiographic signal is divided into n beats, a total of n features are extracted, and in this case, the number of the intermediate artificial neural networks is preferably n. The features of each of the n beats are input to an input layer of each of n intermediate artificial neural networks, thereby obtaining m-dimensional feature values for each of the n beats from an output layer of each of the n intermediate artificial neural networks. At this time, the value output by the intermediate artificial neural network is referred to as 'intermediate output data'.

The operation of extracting the features 831, 832, 833 … from the beats 821, 822, 823 … uses a known algorithm such as Fast Fourier Transform (FFT), Mel cepstral coefficient (MFCC), Filter Bank (Filter Bank), Wavelet (Wavelet), etc., but the operation of obtaining the m-dimensional feature values from the features 831, 832, 833 … of the beats is a result of analysis using artificial intelligence, which is a difference between the two.

This is illustrated generically as time series data, which can be described as: when the time-series data is divided into n units (units), and the features of the respective units are extracted, the features of the respective units are input to n intermediate artificial neural networks. The features of each of the n cells are input to an input layer of each of n intermediate artificial neural networks, whereby an output layer of each of the n intermediate artificial neural networks obtains m-dimensional feature values for each of the n cells.

Then, the m-dimensional feature value for each of the n beats output by the n intermediate artificial neural networks is input to the final artificial neural network 850 according to the order of the individual beats on the time axis. The final artificial neural network may be an RNN, and in this case, in order to reflect the time-varying characteristic of the continuous beats, the m-dimensional characteristic value of the beat to be determined, and the m-dimensional characteristic value of the beat to be determined next are sequentially input to the input layer of the final artificial neural network.

The data input into the final artificial neural network is output to c nodes located in the output layer through the hidden layer, and when the final artificial neural network is a classification model, the values output to the c nodes can be values for classifying all the electrocardiosignals into c classifications. When the final artificial neural network is a predictive model, the values output to the c nodes may be values for predicting a result within a specific range from all the electrocardiographic signals. The value output by the final artificial neural network is referred to as 'final output data'.

The final artificial neural network receives the m-dimensional feature values for the respective beats in turn, and therefore, the number of nodes of the input layer of the final artificial neural network is preferably m. Of course, the number of nodes of the input layer of the final artificial neural network may be m or more, so that other values or parameters than the m-dimensional characteristic values about the respective beats can be received.

As generally used in the art to which the present invention pertains, 'classification model' in this specification refers to an artificial intelligence model with the purpose of finding which group input data belongs to, and 'prediction model' refers to an artificial intelligence model whose result value may be any value within a range of learning data.

Describing this as generalization into time series data, when the time series data is divided into n cells (units), a feature value in m-dimension for each of the n cells is obtained using n intermediate artificial neural networks. The final artificial neural network receives the m-dimensional feature values for each of the n cells from the n intermediate artificial neural networks according to the order of the respective cells on the time axis, thereby performing tasks (task) of classification, prediction, and the like with respect to the entire time-series data.

In order to reflect the time-varying characteristics of the successive cells, the m-dimensional characteristic value of the cell to be judged and the m-dimensional characteristic value of the cell temporally adjacent to the cell to be judged may be successively input to the input layer of the final artificial neural network. For the electrocardiographic signal, it is necessary to continuously input 3 beats of the preceding beat to be determined, the following beat to be determined, and the following beat to be determined, but the general time-series data may be input only to the preceding cell of the cell to be determined and 2 cells to be determined, or input only to the following cell of the cell to be determined and 2 cells to be determined. Of course, all of the preceding cell of the cell to be determined, and the subsequent cell of the cell to be determined may be 3.

The final artificial neural network receives m-dimensional eigenvalues for each cell, and therefore, the input layer of the final artificial neural network preferably has m nodes. Of course, the number of nodes of the input layer of the final artificial neural network may be m or more, so that other values or parameters than the m-dimensional characteristic values of the respective cells may be input.

Fig. 9a and 9b are diagrams illustrating node structures of an intermediate artificial neural network and a final artificial neural network used in a time-series data analysis method using artificial intelligence according to an embodiment of the present invention.

For convenience of explanation, in fig. 9a and 9b, the feature 831 of the first beat 821 obtained by dividing the electrocardiographic signal is input to the first intermediate artificial neural network 841, and the intermediate output data of the first intermediate artificial neural network 841 is input to the final artificial neural network 850.

When a feature extracted from each beat by Fast Fourier Transform (FFT), mel Cepstral Coefficient (MFCC), Filter Bank (filterbank), Wavelet (Wavelet), or the like is assumed to be a value of d dimension, the number of nodes of the input layer of each of the intermediate artificial neural networks 841, 842, 843 is preferably d. Of course, the number of nodes of the input layer of the intermediate artificial neural network may be d or more, so that other values or parameters than the d-dimensional features extracted from the respective beats may be input.

When d-dimensional feature 831 is extracted from first beat 821, d-dimensional feature 831 is input to d nodes located in input layer 841a of first intermediate artificial neural network 841. The input data is analyzed by the hidden layer 841b of the first intermediate artificial neural network 841, and the analysis result is output by m nodes located in the output layer 841c of the first intermediate artificial neural network 841.

The intermediate output data of the first intermediate artificial neural network 841, which is output by the m nodes located at the output layer 841c of the first intermediate artificial neural network 841, is input to the m nodes located at the input layer 850a of the final artificial neural network 850. The final artificial neural network 850 receives not only the intermediate output data of the first intermediate artificial neural network 841 but also the intermediate output data from the other intermediate artificial neural networks 842, 843 …, analyzes the time-varying characteristics of successive beats via the hidden layer 850b, and outputs the final output data. The final output data is output by c nodes located at the output layer 850c of the final artificial neural network 850.

Fig. 10 is a diagram for explaining a process of inputting intermediate output data output from a plurality of intermediate artificial neural networks to a final artificial neural network according to an embodiment of the present invention.

When the pulse to be determined in the cardiac electric signal is referred to as the second pulse 822, m-dimensional intermediate output data of the preceding pulse 821 of the pulse to be determined, the pulse 822 to be determined, and the subsequent pulse 832 of the pulse to be determined need to be continuously input.

That is, m-dimensional intermediate output data 841c, m-dimensional intermediate output data 842c, and m-dimensional intermediate output data 843c are sequentially input to the final artificial neural network. The m-dimensional intermediate output data 841c is obtained by inputting the d-dimensional feature 831 extracted from the preceding pulse 821 of the pulse to be determined into the first intermediate artificial neural network 841, the m-dimensional intermediate output data 842c is obtained by inputting the d-dimensional feature 832 extracted from the pulse 822 to be determined into the second intermediate artificial neural network 842, and the m-dimensional intermediate output data 843c is obtained by inputting the d-dimensional feature 833 extracted from the post-event pulse 823 of the pulse to be determined into the third intermediate artificial neural network 843.

The above contents are specifically organized as follows.

To analyze the ecg signal to determine arrhythmia, deep learning may be used. The artificial neural network model includes a CNN model, an RNN model, and the like. The CNN model is less accurate because it is not suitable for analyzing time series data with time varying features. The CNN method is based on the positional information of the features, and therefore has a problem that the determination capability is greatly reduced for small distortion or noise. (Sabour, S., frost, N.Hinton, G.E.dynamic routing between capsules. InAdvances in Neural Information Processing Systems, pp.3859-3869,2017)

To overcome these problems, an RNN model may be used for analyzing time series data with time varying characteristics. The biggest problem with RNN models is their slow learning and analysis speed. Although RNNs can take advantage of more information and can reflect time-varying features with some improvement in accuracy, they need to remember more information than CNNs and therefore increase in complexity. Furthermore, since the RNN needs to use the association with consecutive cells as information, consecutive cells need to be sequentially processed. It is therefore necessary to process a long length of input, so that learning and analysis takes a longer time.

According to an embodiment of the present invention, in order to solve the problems of low accuracy of CNNs and inability of variable length input processing and the slow learning and analysis speed of RNNs, it is proposed to use a hierarchical RNN (hierarchical RNN with variable length boundaries for parallelization model) composed of a plurality of intermediate artificial neural networks and one final artificial neural network with variable length boundaries.

Existing RNN models treat n consecutive beats as one input. However, in the artificial intelligence model according to an embodiment of the present invention, it is divided into 2 steps of short-term rnn (short term rnn) processed by a plurality of intermediate artificial neural networks and long-term rnn (long term rnn) processed by one final artificial neural network for learning and analysis. First, consecutive n beats are used as input to an intermediate artificial neural network, respectively. As a result, an m-dimensional vector for one beat is obtained and then input to the final artificial neural network, and the electrocardiographic signal is analyzed to determine a disease such as arrhythmia.

The plurality of intermediate artificial neural networks process the plurality of beats, respectively, and the plurality of intermediate artificial neural networks are the same model. Therefore, when a plurality of beats are processed, parallel processing can be performed because there is no Dependency (Dependency) between them. Therefore, compared to the conventional RNN system in which all the electrocardiographic signals need to be input, it is possible to perform operations faster in proportion to the number of computing devices that perform parallel processing or the number of processors that perform parallel processing. Of course, a configuration in which a plurality of input beats are processed by an intermediate artificial neural network limited to less than the number of beats according to resources is also possible. In the structures, each beat is processed in parallel in a plurality of intermediate artificial neural networks, and the intermediate output data with the shortened length is processed in the final artificial neural network, so that the number of nodes of the hidden layer of each artificial neural network can be greatly reduced, and the available memory is increased. This increases the speed of diagnostics in a GPU environment with video memory (video memory) limitations. In addition, the data length which needs to be processed in the final artificial neural network is shortened, so that the calculation amount (complexity) is reduced, and the diagnosis speed is greatly improved.

With the time-series data analysis method using artificial intelligence according to an embodiment of the present invention, the length of input data when analyzing the electrocardiographic signals of fig. 3 is calculated as follows.

The cardiac signal of fig. 3a and 3b is divided into 9 beats (beats). When one beat is sampled at 250Hz, the processing length when extracting features from 9 beats is 250 × 9 ═ 2250.

When 9 features extracted from 9 beats are input to 9 intermediate artificial neural networks, the processing length is 1 × 9 ═ 9.

When m-dimensional intermediate output data outputted from the 9 intermediate artificial neural networks is inputted to the final artificial neural network, m-dimensional intermediate output data on the preceding beat, the judgment target beat, and the subsequent beat are successively inputted. As described above, the data is input to the final artificial neural network according to the following sequence.

(1) A feature value for the m-dimension of beat a, a feature value for the m-dimension of beat B, and a feature value for the m-dimension of beat C.

(2) A feature value for the m-dimension of beat B, a feature value for the m-dimension of beat C, and a feature value for the m-dimension of beat D.

(3) A feature value for the m-dimension of beat C, a feature value for the m-dimension of beat D, and a feature value for the m-dimension of beat E.

(4) A feature value for the m-dimension of the beat D, a feature value for the m-dimension of the beat E, and a feature value for the m-dimension of the beat F.

(5) A feature value for the m-dimension of the beat E, a feature value for the m-dimension of the beat F, and a feature value for the m-dimension of the beat G.

(6) A feature value for the m-dimension of the beat F, a feature value for the m-dimension of the beat G, and a feature value for the m-dimension of the beat H.

(7) A feature value for the m-dimension of the beat G, a feature value for the m-dimension of the beat H, and a feature value for the m-dimension of the beat I.

That is, since the input of the length 3 is performed 7 times, when m-dimensional intermediate output data output from 9 intermediate artificial neural networks is input to the final artificial neural network, the processing length is 3 × 7 to 21.

The sum of all treatment lengths is 2250+9+21 ═ 2280. From this, it is found that the treatment length is reduced to about 1/3 compared with the treatment length 5250 of the conventional method shown in fig. 3a and 3 b.

As described above, according to the time-series data analysis method using artificial intelligence according to an embodiment of the present invention, the accuracy of analysis is higher than that of the conventional ecg analysis method using CNN, and the length of data to be processed is shorter than that of the conventional ecg analysis method using RNN, so that the analysis speed is faster and the real-time performance of ecg analysis can be ensured. According to the time series data analysis method using artificial intelligence, provided by the embodiment of the invention, electrocardiosignals are accurately analyzed in real time, and the checked person is informed in advance before heart diseases occur, so that measures such as medicines and hospital treatment can be performed in advance. Further, the early-stage correspondence of fatal heart diseases can be made swiftly, and thus, it can also be used for emergency rescue services.

Fig. 11 is a flowchart illustrating a time-series data analysis method using artificial intelligence according to an embodiment of the present invention.

As shown in fig. 11, a time-series data analysis method using artificial intelligence according to an embodiment of the present invention may include: a step S1110 of inputting, for each of a plurality of cells into which time-series data is divided on a time axis, a feature (feature) of each cell into an intermediate artificial neural network; a step S1120 of obtaining intermediate output data of m (m is a natural number of 2 or more) dimensions from the intermediate artificial neural network; a step S1130 of inputting intermediate output data of a plurality of units that are temporally adjacent to each other to the final artificial neural network, and obtaining final output data output from the final artificial neural network; and a step S1140 of generating a time-series data analysis result by using the final output data.

The respective steps have already been described in detail above, and therefore, the description thereof is omitted for the overlapping contents.

Fig. 12 is a diagram illustrating a case where an intermediate artificial neural network is embodied in a plurality of levels in an artificial neural network structure according to an embodiment of the present invention.

As shown in fig. 12, in the time-series data analysis method using artificial intelligence according to an embodiment of the present invention, the intermediate artificial neural network may be embodied in a plurality of levels.

The intermediate artificial neural networks of the respective levels output the characteristic values of the input data, and the characteristic values output by the intermediate artificial neural networks of the previous level are input to the intermediate artificial neural networks of the next level.

When the intermediate artificial neural network is composed of k levels, the feature of each of the plurality of cells is input to the level-1 intermediate artificial neural network, the feature value output by the respective level intermediate artificial neural network is input to the intermediate artificial neural network of the next level, the feature value output by the level-1 intermediate artificial neural network is input to the level-k intermediate artificial neural network, and the m-dimensional feature value output by the level-k intermediate artificial neural network is input to the final artificial neural network.

In fig. 12, the intermediate artificial neural network having a plurality of levels is designed in such a manner that a plurality of feature values output from a plurality of intermediate artificial neural networks of a previous level are input to one intermediate artificial neural network of a next level, which shows that data is processed by: the level 1 intermediate artificial neural network processes data of each unit (first unit), the level 2 intermediate artificial neural network processes data of a second unit grouping (grouping) a plurality of units which are immediately adjacent in time, and the level 3 intermediate artificial neural network processes data of a third unit grouping a plurality of second units which are immediately adjacent in time. At this time, the m-dimensional feature value output from the k-th median artificial neural network becomes an m-dimensional feature value of the k-th unit of data. At this time, the number of the intermediate artificial neural networks of each level may be different according to the grouping criteria.

Of course, the intermediate artificial neural network having a plurality of levels may be designed such that a characteristic value output by an intermediate artificial neural network of a previous level is represented by 1: the pattern 1 is input to an intermediate artificial neural network of the next level. At this time, the intermediate artificial neural networks of the respective levels may be embodied in the same number. For example, when the number of the level 1 intermediate artificial neural networks is n, the number of the level 2 intermediate artificial neural networks may also be embodied as n, the number of the level 3 intermediate artificial neural networks may also be embodied as n, …, and the number of the level k intermediate artificial neural networks may also be embodied as n.

Fig. 13 is a diagram illustrating a time-series data analysis method when an intermediate artificial neural network is embodied in a plurality of levels in an artificial intelligence structure according to an embodiment of the present invention.

In the time-series data analysis method using artificial intelligence according to an embodiment of the present invention, the time-series data analysis method when the intermediate artificial neural network is composed of a plurality of levels may include: a step S1310 of inputting, for each of a plurality of cells into which time-series data is divided on a time axis, a feature (feature) of each cell to the level 1 intermediate artificial neural network; a step S1320 of inputting intermediate output data output from the intermediate artificial neural network at the level i (i is a natural number equal to or greater than 1) to the intermediate artificial neural network at the level i + 1; a step S1330 of obtaining intermediate output data from the level k intermediate artificial neural network; and a step S1340 of sequentially inputting the intermediate output data of the level k to the final artificial neural network and obtaining final output data from the final artificial neural network.

When the intermediate artificial neural network having a plurality of levels is designed such that a plurality of feature values output by a plurality of intermediate artificial neural networks of a previous level are input to one intermediate artificial neural network of a next level, the step of sequentially inputting the level k intermediate output data to the final artificial neural network may include: and a step of sequentially inputting level k intermediate output data regarding data of a k-th unit that is temporally adjacent to the final artificial neural network.

In one aspect, when the intermediate artificial neural network having a plurality of levels is designed such that a characteristic value output by an intermediate artificial neural network of a previous level is represented by 1: the step of sequentially inputting the level k intermediate output data on the temporally adjacent cells to the final artificial neural network when the mode 1 is input to one intermediate artificial neural network of the next level is the same as that when the level of the intermediate artificial neural network is 1 as described in fig. 9 and 10, and thus the description thereof will be omitted for redundant contents.

The above description has been made of a learning method using time-series data of artificial intelligence according to an embodiment of the present invention.

Hereinafter, a learning method of time-series data using artificial intelligence according to an embodiment of the present invention will be described with reference to fig. 14 to 17. Specifically, a learning method of time-series data using artificial intelligence according to an embodiment of the present invention is described by describing a process of learning electrocardiographic signals using a plurality of artificial neural networks according to an embodiment of the present invention.

As is well known, the learning process of an artificial neural network using learning data is substantially similar in many parts to the input data analysis process using an artificial neural network. The difference is only that the step of comparing the output value with the actual value to calculate errors and changing the weight value of the artificial neural network by back-propagating the errors is added in the process of learning the artificial neural network.

Hereinafter, a method of learning time-series data according to an embodiment of the present invention will be described focusing on a difference from a method of analyzing time-series data according to an embodiment of the present invention.

Fig. 14 is a flowchart illustrating an electrocardiogram learning method using artificial intelligence according to an embodiment of the present invention.

As shown in fig. 14, the electrocardiogram learning method using artificial intelligence according to an embodiment of the present invention may include: an electrocardiographic signal collection step S1410; an electrocardiographic signal segmentation step S1420; a beat combination step S1430; a feature (feature) extraction step S1440; and an artificial neural network learning step S1450.

The electrocardiographic signal collection step S1410 is a step of collecting the electrocardiographic signals received by the server 200 from the electrocardiograph apparatus 100 or the electrocardiographic signals directly input by the user. Generally, the judgment result such as arrhythmia is not labeled on all electrocardiographic signals measured for a long time, but labeled on a beat unit basis. The known data include MIT-BIH Arrhythmia Database (Arrhytmia Database).

The electrocardiographic signal dividing step S1421 is a step of dividing the electrocardiographic signal into units of beats.

The beat combination step S1430 is a step of combining a preceding beats and b subsequent beats with the labeled beat.

Fig. 15 is a diagram illustrating learning data used in an electrocardiogram learning method using artificial intelligence according to an embodiment of the present invention.

As shown in fig. 15, according to the electrocardiogram learning method using artificial intelligence according to an embodiment of the present invention, a preceding beat and b subsequent beats are combined with a labeled beat and used as learning data.

According to the method for analyzing time-series data using artificial intelligence according to an embodiment of the present invention, time-varying features, which are generated by connecting a plurality of beats adjacent in time, such as the occurrence interval with a preceding beat or a succeeding beat and the height difference with a preceding beat or a succeeding beat, are analyzed together without analyzing a single beat. Therefore, the tagged beats should be learned not only by the learning but also by the learning in combination with the pre-beat and the post-beat of the tag.

The number of preceding beats a and the number of following beats b to be combined with the labeled beats can be arbitrarily determined. The number a of the preceding beats and the number b of the subsequent beats determined at this time should be input using the same number even when the analysis is performed.

The feature extraction step S1440 is a step of extracting features of each beat. The extracted features may be obtained by using a variety of methods such as Fast Fourier Transform (FFT), Mel Frequency Cepstral Coefficient (MFCC), Filter Bank (Filter Bank), Wavelet (Wavelet), and the like, which are widely used in signal processing. The feature extraction algorithm used in learning is preferably the same as the feature extraction algorithm used in analysis.

The artificial neural network learning step S1450 is a step of learning the artificial neural network using a technique such as a well-known back propagation and gradient descent method. The artificial neural network learning step S1450 is explained in detail below with reference to fig. 16.

Fig. 16 is a diagram illustrating a time-series data learning method using artificial intelligence according to an embodiment of the present invention.

A process of learning time-series data (e.g., cardiac electrical signals) using artificial intelligence according to an embodiment of the present invention is explained with reference to fig. 16.

When a database labeled with beat units is used, such as the MIT-BIH arrhythmia database, the learning data 1610 should be an electrocardiographic signal obtained by joining one or more preceding beats and one or more following beats to the labeled beat in the beat joining step S1430. If the database used is a database in which the analysis results of all the electrocardiographic signals are labeled as labels in the electrocardiographic signals, not on a beat-by-beat basis, the electrocardiographic signals with labels can be used as the learning data 1610 without the need of combining beats with step S1430.

When a database in which tags are labeled in units of cells is used, the learning data should be time-series data in which one or more cells are bound to a labeled cell. If the database used is a database in which the analysis results of the entire time-series data are labeled as tags in the respective time-series data, not in units of cells, the labeled time-series data can be directly used as the learning data.

When the learned data 1610 is input, the learned beat is divided into a plurality of beats 1621, 1622, 1623 … according to a predetermined criterion. The same method is preferably used for the electrocardiographic signal segmentation method used for learning and the electrocardiographic signal segmentation method used for analysis.

This is illustrated generically as time series data, which can be described as: when learning data is input, the learning data is divided into a plurality of cells according to a predetermined criterion. The same method is preferably used for dividing the time-series data used for learning and for dividing the time-series data used for analysis.

Then, an operation of extracting features 1631, 1632, 1633 … of the respective beats is performed for the plurality of beats 1621, 1622, 1623 …, respectively. The same algorithm is preferably used for the feature extraction algorithm of the electrocardiographic signal used for learning and the feature extraction algorithm of the electrocardiographic signal used for analysis.

This is illustrated generically as time series data, which can be described as: when time series data is divided into a plurality of units (units), features of the respective units are extracted for the plurality of units. The feature extraction algorithm for the time-series data used for learning and the feature extraction algorithm for the time-series data used for analysis preferably use the same algorithm.

Then, an operation of inputting the features 1631, 1632, 1633 … of the respective beats to the intermediate artificial neural networks 1641, 1642, 1643 … is performed. When the learning data 1610 is combined by n beats, a total of n features are extracted, and in this case, the intermediate artificial neural network is also preferably n.

This is illustrated generically as time series data, which can be described as: when the time-series data is divided into n units (units), and the features of the respective units are extracted, the features of the respective units are input to n intermediate artificial neural networks.

Then, the m-dimensional feature value for each of the n beats output by the n intermediate artificial neural networks is input to the final artificial neural network 1650 according to the order of the individual beats on the time axis, thereby obtaining final output data on the learning data.

Describing it as generalized time-series data, when the time-series data is divided into n units (units), m-dimensional feature values for each of the n units are obtained using n intermediate artificial neural networks, and finally the artificial neural networks receive the m-dimensional feature values for each of the n units from the n intermediate neural networks in accordance with the order of the respective units on the time axis, thereby obtaining classification values, prediction values, and the like for the learning data.

When the final output data on the learning data is obtained from the final artificial neural network 1650, the final output data on the learning data is compared with the label labeled on the learning data. At this time, the final output data of the learning data becomes an output value, and the flag of the learning data becomes an actual value. The error between the output value and the actual value is calculated by using a loss function (loss function) or the like, and then the error is propagated in the reverse direction.

The weights of the final artificial neural network 1650 and the weights of the intermediate artificial neural networks 1641, 1642, 1643 … are adjusted by backpropagation. The weight adjustment method of the back propagation and artificial neural network is described in detail in https:// en. wikipedia. org/wiki/Backpropagation, etc., and thus detailed description thereof is omitted.

At this time, it is to be noted the setting of the learning rates of the final artificial neural network 1650 and the intermediate artificial neural networks 1641, 1642, 1643 …. 'learning rate' refers to a step size (step size) for finding a minimum value in a gradient descent algorithm for adjusting the weight of the artificial neural network.

But cannot learn the intermediate artificial neural network in parallel (parallel). Since there is only one correct answer for multiple input beats. Therefore, back propagation with respect to multiple input beats needs to be performed for one correct answer. When the number of beats is n, and the number of the intermediate artificial neural networks is also n, the n intermediate artificial neural networks need to share (share) a weight value with each other, and the learning rate of the intermediate artificial neural networks when the backward propagation needs to be set to be 1/n times of the learning rate of the final artificial neural network.

In one aspect, the learning method of the artificial neural network is the same when the intermediate artificial neural network is embodied in a plurality of levels. After the learning data is analyzed using the same method as the method of analyzing the artificial neural network when the intermediate artificial neural network is embodied in a plurality of levels (refer to fig. 12 and 13), when the final output data on the learning data is obtained from the final artificial neural network, the final output data (output value) on the learning data is compared with the label (actual value) marked on the learning data. And calculating the error between the output value and the actual value by utilizing a loss function (lossfunction) and the like, and then performing back propagation, thereby adjusting the weight of the final artificial neural network and the weights of the intermediate artificial neural networks of a plurality of levels.

Fig. 17 is a diagram illustrating a time-series data learning method using artificial intelligence according to an embodiment of the present invention.

As shown in fig. 17, a time-series data learning method using artificial intelligence according to an embodiment of the present invention may include: a step S1710 of inputting, for each of a plurality of units in which the time-series data is divided on the time axis, a feature (feature) of each unit to the intermediate artificial neural network; a step S1720 of obtaining intermediate output data of m (m is a natural number of 2 or more) dimensions from the intermediate artificial neural network; a step S1730 of inputting intermediate output data of a plurality of units temporally adjacent to each other to the final artificial neural network, and obtaining final output data output from the final artificial neural network; a step S1740 of comparing the final output data with the tag marked in time series and calculating an error; and a step S1740 of reversely propagating the error and adjusting the weights of the intermediate artificial neural network and the final artificial neural network.

The respective steps have already been described in detail above, and therefore, the description thereof is omitted for the overlapping contents.

Fig. 18 is a hardware configuration diagram of an exemplary computing device that can embody a time-series data analysis method or a time-series data learning method using artificial intelligence according to an embodiment of the present invention.

As shown in fig. 18, an exemplary computing device that may embody a time-series data analysis method or a time-series data learning method using artificial intelligence according to an embodiment of the present invention may include: one or more processors 1810; a memory (storage) 1820; a memory (memory)1830 that loads (loads) a computer program for execution by processor 1810; a network interface 1840; and a bus (bus). Fig. 18 shows only structural elements related to the embodiment of the present invention. Therefore, it will be apparent to those skilled in the art that the present invention may include other general structural elements in addition to the structural elements shown in fig. 18.

The processor 1810 controls the overall operation of each configuration of a computing device, which may embody a time-series data analysis method or a time-series data learning method using artificial intelligence according to an embodiment of the present invention. The processor 1810 may include: a Central Processing Unit (CPU), a microprocessor Unit (MPU), a Micro Controller Unit (MCU), a Graphics Processing Unit (GPU), or any other type of processor known in the art to which the present invention pertains. Also, processor 1810 may perform computations for more than one application or program (to perform a method according to an embodiment of the invention). A computing device that can embody a time series data analysis method or a time series data learning method using artificial intelligence according to an embodiment of the present invention may be provided with one or more processors.

The memory 1820 may store more than one computer program non-temporarily. The memory 1820 may include: such as a non-volatile Memory, such as a Read Only Memory (ROM), an Erasable Programmable Read Only Memory (EPROM), an Electrically Erasable Programmable Read Only Memory (EEROM), and a flash Memory, a hard disk, a removable disk, or any other form of computer-readable recording medium known in the art to which the present invention pertains. The memory 1820 stores therein a computer program for executing a time-series data analysis method or a time-series data learning method using artificial intelligence according to an embodiment of the present invention.

The memory 1830 stores various data, commands, and/or information. The memory 1830 may load one or more computer programs from the storage 1820 to perform a time series data analysis method or a time series data learning method using artificial intelligence according to an embodiment of the present invention.

The bus may provide a communication function between structural elements of the computing device, wherein the computing device may embody a time-series data analysis method or a time-series data learning method using artificial intelligence according to an embodiment of the present invention. The Bus may be embodied as various types of buses such as an Address Bus (Address Bus), a Data Bus (Data Bus), and a Control Bus (Control Bus).

The network interface 1840 supports wired and wireless communication of a computing device, which may embody a time-series data analysis method or a time-series data learning method using artificial intelligence according to an embodiment of the present invention. The network interface 1840 may support various communication methods other than internet communication. To this end, the network interface 1840 may include a communication module well known in the art to which the present invention pertains.

A computer program that executes a time-series data analysis method or a time-series data learning method using artificial intelligence according to an embodiment of the present invention is loaded into the memory 1830, so that the processor 1810 may include instructions (instruction) for executing the time-series data analysis method or the time-series data learning method using artificial intelligence according to an embodiment of the present invention.

Hereinafter, a time-series data analysis apparatus using artificial intelligence according to an embodiment of the present invention will be described.

The time-series data analysis device according to an embodiment of the present invention includes: a network interface 1840 that receives the time-series data; one or more processors 1810; a memory 1830 loaded with a computer program for execution by the processor 1841; and a memory 1820 storing a computer program. Wherein the computer program may comprise: a command for inputting a feature (feature) of each cell to an intermediate artificial neural network for each of a plurality of cells into which time-series data are divided on a time axis, and obtaining intermediate output data of m (m is a natural number of 2 or more) dimensions from the intermediate artificial neural network; a command for inputting intermediate output data of a plurality of units which are temporally adjacent to each other to the final artificial neural network to obtain final output data output from the final artificial neural network; and a command for generating an analysis result of the time-series data by using the final output data.

The respective commands of the computer program stored in the memory 1820 of the time-series data analysis apparatus according to an embodiment of the present invention are used to perform the aforementioned time-series data analysis method using artificial intelligence, and thus a description of the repetitive contents is omitted.

Hereinafter, a time-series data learning apparatus using artificial intelligence according to an embodiment of the present invention will be described.

The time-series data learning device according to an embodiment of the present invention includes: a network interface 1840 that receives the time-series data; one or more processors 1810; a memory 1830 loaded with a computer program for execution by the processor 1841; and a memory 1820 storing a computer program. Wherein the computer program may comprise: a command for inputting a feature (feature) of each cell to an intermediate artificial neural network for each of a plurality of cells into which time-series data are divided on a time axis, and obtaining intermediate output data of m (m is a natural number of 2 or more) dimensions from the intermediate artificial neural network; a command for inputting intermediate output data of a plurality of units which are temporally adjacent to each other to the final artificial neural network to obtain final output data output from the final artificial neural network; and a command to compare the final output data with tags tagged in the time series data.

The respective commands of the computer program stored in the memory 1820 of the time-series data learning apparatus according to an embodiment of the present invention are used to execute the aforementioned time-series data learning method using artificial intelligence, and thus a description of the repetitive contents is omitted.

In one aspect, the concepts of the present invention, as described with reference to the figures, can be embodied in computer readable code on a computer readable medium. The computer-readable recording medium may be a removable storage medium (CD, DVD, blu-ray disc, USB storage device, removable hard disk) or a fixed storage medium (ROM, RAM, computer fixed hard disk). The computer program recorded in the computer-readable recording medium can be transmitted to another computing apparatus via a network such as the internet and installed in the other computing apparatus, and can be used in the other computing apparatus.

Acts may be shown in the drawings in a particular order, but it should not be understood that the acts must be performed in the particular order shown, or in sequence, or that all acts shown are performed, to achieve desirable results. In certain situations, multitasking and parallel processing may be advantageous. Further, the separation of various structures in the embodiments described above should not be understood as requiring such separation, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

Although the embodiments of the present invention have been described with reference to the drawings, it will be understood by those skilled in the art that the present invention may be embodied in other specific forms without changing the technical spirit or essential characteristics thereof. It is therefore to be understood that the above described embodiments are illustrative in all respects, not restrictive.