阅读说明：本技术 一种基于注意力机制的机械通气人机异步检测识别方法 (Mechanical ventilation man-machine asynchronous detection and identification method based on attention mechanism ) 是由 熊富海 颜延 王磊 谯小豪 李慧慧 于 2021-09-23 设计创作，主要内容包括：本发明涉及医疗器械智能控制技术领域,特别涉及到一种基于注意力机制的机械通气人机异步检测识别方法；本发明先后分别构建核级注意力模块与时序注意力模块,且能从波形数据的卷积核特征层面以及信号的时序层面寻找到更加利于分别不同种人机异步类型的信号特征段,给予这些特征段更强的关注度,这样就可以更好地对人机异步做出区分。(The invention relates to the technical field of intelligent control of medical instruments, in particular to a mechanical ventilation man-machine asynchronous detection and identification method based on an attention mechanism; according to the invention, the kernel level attention module and the time sequence attention module are respectively constructed in sequence, and signal feature sections which are more beneficial to different man-machine asynchrony types can be found from the convolution kernel feature level of waveform data and the time sequence level of signals, so that the feature sections have stronger attention, and man-machine asynchrony can be better distinguished.)

1. A mechanical ventilation man-machine asynchronous detection and identification method based on an attention mechanism is characterized by comprising the following steps:

s1, constructing a one-dimensional convolution neural network model according to waveform data of the mechanical ventilation of the respirator;

step S2, respectively constructing a core level attention submodule and a time sequence attention submodule in sequence;

s3, sequentially connecting the core-level attention submodule and the time sequence attention submodule in series to form an organic integrated multi-attention comprehensive module;

s4, embedding and integrating a multi-attention integration module into the one-dimensional convolutional neural network model to form a multi-attention composite model;

and step S5, analyzing and judging the waveform data of the mechanical ventilation of the breathing machine through the multi-attention composite model.

2. The method for asynchronous detection and identification of mechanical ventilation man-machine based on attention mechanism as claimed in claim 1, wherein in step S2, a kernel-level attention submodule is first constructed, the kernel of which is composed of two models of fully connected layers sequentially connected in series, the features of the input convolution kernel are processed respectively by using one-dimensional global average pooling and one-dimensional global maximum pooling, so as to obtain global maximum pooling features and global average pooling features, and then the global maximum pooling features and the global average pooling features are processed in parallel through two fully connected layers of structures sequentially connected, and then the corresponding matrix position elements are added and then converted.

3. The method for asynchronous detection and identification of mechanical ventilation man-machine based on attention mechanism as claimed in claim 2, wherein in step S2, after a kernel level attention submodule is constructed, a time sequence attention submodule is constructed, the features input from convolution kernel are averaged pooled and maximally pooled, time sequence related averaged pooled feature matrix and maximal pooled feature matrix are extracted, the averaged pooled feature matrix and maximal pooled feature matrix are spliced to obtain a combined feature matrix, and then the combined feature matrix is converted through convolution network.

4. The method for asynchronous detection and identification of mechanical ventilation man-machine based on attention mechanism as claimed in claim 3, wherein in step S3, the core-level attention submodule and the time-series attention submodule are connected in series to form a multi-attention synthesis module, the input features are processed by the core-level attention submodule to form a core-level output feature matrix with core-level attention, and the core-level output features are processed by the time-series attention submodule to form time-series output features.

5. The method for human-machine asynchronous detection and identification of mechanical ventilation based on attention mechanism as claimed in claim 4, wherein in step S1, the waveform data includes flow rate, airway pressure and volume during mechanical ventilation of ventilator.

6. The method for human-machine asynchronous detection and identification of mechanical ventilation based on attention mechanism as claimed in claim 5, wherein in step S1, all data of flow, airway pressure and volume in the human-machine asynchronous event during the mechanical ventilation process of the ventilator for normal breathing, dual triggering and ineffective inspiration effort are collected, and the collected all data are preprocessed and labeled by segmentation.

7. The method according to claim 6, wherein the one-dimensional convolutional neural network model is sequentially composed of a convolutional layer, a maximal pooling layer, a convolutional layer, a fully-connected layer, and a fully-connected layer from front to back.

8. The method for mechanical ventilation human-machine asynchronous detection and identification based on attention mechanism of claim 7, wherein in step S4, a multi-attention synthesis module is embedded and integrated into the one-dimensional convolutional neural network model to form a multi-attention composite model, the multi-attention composite model is sequentially composed of a convolutional layer, a maximum pooling layer, a first multi-attention layer, a convolutional layer, a maximum pooling layer, a second multi-attention layer, a convolutional layer, a fully-connected layer, and a fully-connected layer from front to back, and the first multi-attention layer and the second multi-attention layer include a kernel-level attention layer and a time-sequence attention layer.

9. The method for human-machine asynchronous detection and identification of mechanical ventilation based on attention mechanism as claimed in claim 8, wherein in step S5, waveform data during mechanical ventilation of the ventilator is analyzed and discriminated by the multi-attention composite model, and when the waveform data is detected as data in human-machine asynchronous events during mechanical ventilation of the ventilator with double-triggering or ineffective inhalation effort, the human-machine asynchronous events with double-triggering or ineffective inhalation effort can be discriminated.

10. The method for human-machine asynchronous detection and identification of mechanical ventilation based on attention mechanism as claimed in claim 6, wherein in step S1, the data segments of flow, airway pressure and volume during full-scale mechanical ventilation of ventilator are preprocessed and segmented, the segmented and labeled data segments are normalized and zero-filled, and then the uniformly processed full-scale data are divided into training set and testing set according to 7:3 ratio.

Technical Field

The invention relates to the technical field of intelligent control of medical instruments, in particular to a mechanical ventilation man-machine asynchronous detection and identification method based on an attention mechanism.

Background

The breathing machine is an important life device based on breathing function support, is widely applied to intensive care departments and general departments of hospitals, enters families, becomes auxiliary household devices for daily sleep and the like, and provides important auxiliary support for people with breathing dysfunction.

Generally speaking, one of the most important functions of a ventilator is its ventilation sensitivity, whether it can provide the same-frequency air supply/ventilation support when a patient or user needs to inhale, and whether it can switch the patient exhale in time, i.e. the ventilator needs to detect the inhalation demand of the patient in time, and detect the timing of switching the patient from the inhalation end to the exhalation, etc. These patient inspiratory and expiratory conditions may be characterized by the rate of airflow in the ventilation tubing that the ventilator is connected to the patient, tubing pressure, ventilatory capacity, and the like.

When the breathing requirement of a patient is asynchronous with the air supply switching action of a breathing machine, the airflow speed (Flow), airway pressure (Paw) and ventilation Volume (Volume) in the pipeline can show a certain mode, the modes have a certain rule and can be detected manually or by an algorithm, but the manual detection needs more time, related training needs to be carried out in advance, the labor cost is expensive, and the method is especially suitable for high-quality medical resources in China; it is necessary and important to develop a man-machine asynchronous detection algorithm based on artificial intelligence.

Disclosure of Invention

The invention mainly solves the technical problem of providing a mechanical ventilation man-machine asynchronous detection and identification method based on an attention mechanism, wherein a nuclear level attention module and a time sequence attention module are constructed, signal characteristic sections which are more beneficial to different man-machine asynchronous types can be found from a convolution kernel characteristic level of waveform data and a time sequence level of signals, and stronger attention is given to the characteristic sections, so that man-machine asynchronization can be better distinguished.

In order to solve the technical problems, the invention adopts a technical scheme that: the mechanical ventilation man-machine asynchronous detection and identification method based on the attention mechanism is provided, and comprises the following steps:

s1, constructing a one-dimensional convolution neural network model according to waveform data of the mechanical ventilation of the respirator;

step S2, respectively constructing a core level attention submodule and a time sequence attention submodule in sequence;

s3, sequentially connecting the core-level attention submodule and the time sequence attention submodule in series to form an organic integrated multi-attention comprehensive module;

s4, embedding and integrating a multi-attention integration module into the one-dimensional convolutional neural network model to form a multi-attention composite model;

and step S5, analyzing and judging the waveform data of the mechanical ventilation of the breathing machine through the multi-attention composite model.

As an improvement of the present invention, in step S2, a kernel-level attention submodule is first constructed, the kernel of which is formed by sequentially connecting two fully-connected layers in series, the features of the input convolution kernel are respectively processed by using one-dimensional global average pooling and one-dimensional global maximum pooling, so as to obtain global maximum pooling features and global average pooling features, and then the global maximum pooling features and the global average pooling features are subjected to parallel processing by two sequentially connected fully-connected layer structures, and then the addition of corresponding matrix position elements is performed, followed by conversion.

As a further improvement of the present invention, in step S2, after the kernel-level attention submodule is constructed, the time-sequence attention submodule is constructed, the average pooling and maximum pooling operations are performed on the features input from the convolution kernel, the average pooling and maximum pooling feature matrices related to the time sequence are extracted, the average pooling and maximum pooling feature matrices are spliced to obtain a combined feature matrix, and the combined feature matrix is converted by the convolution network.

As a further improvement of the present invention, in step S3, the core-level attention submodule and the time-series attention submodule are connected in series to form a multi-attention synthesis module, the input features are processed by the core-level attention submodule to form a core-level output feature matrix with core-level attention, and the core-level output features are processed by the time-series attention submodule to form time-series output features.

As a further improvement of the present invention, in step S1, the waveform data includes flow rate, airway pressure, and volume during mechanical ventilation of the ventilator.

As a further improvement of the present invention, in step S1, all data of flow, airway pressure, volume in the mechanical ventilation process of the ventilator for normal breathing, dual triggering and ineffective inspiration effort of the ventilator are collected, pre-processed and segment labeled.

As a further improvement of the invention, the one-dimensional convolutional neural network model is sequentially composed of a convolutional layer, a maximum pooling layer, a convolutional layer, a fully-connected layer and a fully-connected layer from front to back.

As a further improvement of the present invention, in step S4, a multi-attention integration module is embedded and integrated into the one-dimensional convolutional neural network model to form a multi-attention composite model, and the multi-attention composite model is sequentially composed of a convolutional layer, a maximum pooling layer, a first multi-attention layer, a convolutional layer, a maximum pooling layer, a second multi-attention layer, a convolutional layer, a fully-connected layer, and a fully-connected layer from front to back, where the first multi-attention layer and the second multi-attention layer include a nuclear-level attention layer and a time-sequence attention layer.

As a further improvement of the present invention, in step S5, waveform data during mechanical ventilation of the ventilator is analyzed and determined by the multi-attention composite model, and when the waveform data is detected as data in the asynchronous event during mechanical ventilation of the ventilator during a double-triggered or ineffective inspiratory effort, the asynchronous event of the double-triggered or ineffective inspiratory effort can be determined.

As a further improvement of the present invention, in step S1, the data segments of flow, airway pressure and volume during full volume ventilator mechanical ventilation are preprocessed and labeled separately, the labeled data segments are normalized and aligned with zero padding, and then the uniformly processed full volume data is divided into training set and testing set according to the ratio of 7: 3.

The invention has the beneficial effects that: compared with the prior art, the method and the device respectively construct the kernel level attention module and the time sequence attention module, can find signal characteristic sections which are more beneficial to different man-machine asynchronous types from the convolution kernel characteristic level of waveform data and the time sequence level of signals, and give stronger attention to the characteristic sections, so that man-machine asynchrony can be better distinguished.

Drawings

FIG. 1 is a block diagram of the steps of the present invention;

FIG. 2 is a diagram of a build kernel level attention submodule;

FIG. 3 is a diagram of a build timing attention submodule;

FIG. 4 is a block diagram of a combination module of a core level attention submodule and a timing attention submodule;

FIG. 5 is a block diagram of a multi-attention composite model.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

A ventilator is an important device that provides ventilatory support when a patient for some reason has insufficient respiratory capacity to meet the patient's own respiratory needs, providing valuable time for the treatment of the patient's primary illness.

During the process of ventilation of the respirator and a patient, the phenomenon that the ventilation of the respirator is not matched with and conflicts with the breathing requirement of the patient per se occurs at a high probability, and at the moment, a man-machine asynchrony phenomenon, namely an asynchrony phenomenon, occurs in the mechanical ventilation process between the respirator and the patient based on a certain channel (flow, airway pressure, volume and the like); when the patient is in the inspiratory phase when inspiration is required, the ventilator does not provide ventilatory/ventilation support, or is starved or inadequate; when a patient exhales, the breathing machine is not switched to an exhaling state in time, and man-machine asynchronism occurs in the two cases, wherein typical man-machine asynchronism types comprise ineffective inspiration efforts, double triggering and the like, and the ineffective inspiration efforts mean that inspiration efforts made by the patient when the patient has inspiration requirements are not captured by the breathing machine, so that the breathing machine can not provide ventilation support, and the danger of suffocation of the patient can be caused in the past. The double triggering is that the inspiration action of the patient is captured twice by the breathing machine in a short time, so that the two times of ventilation support of the breathing machine are triggered; it is characterized in that the air is inhaled twice continuously, and no time is left in the middle for exhalation; typically, patient effort is present throughout the dual-trigger breath, indicating that continued patient effort during the initial phases of the inspiratory and expiratory cycles triggers a second breath, i.e., dual-trigger.

The technical scheme that the asynchronous type of the breathing machine and a patient is predicted by adopting waveform signals of flow, pressure, volume and the like during mechanical ventilation of the breathing machine is less at present, and some technical schemes are often judged by adopting the characteristics of field expert selection design, so that a large amount of time of experts is consumed naturally, the cost is higher, and the potential practical deployment implementation is not facilitated; on the other hand, even if the characteristics manually extracted by experts are still to be verified in the aspects of expandability, generalization, individuation and the like, the schemes have certain limitations.

In the prior art, the attention mechanism is less applied to the man-machine asynchronous detection problem, in fact, the attention mechanism is more successfully and effectively applied to other fields of artificial intelligence, and the mechanism is also necessary to be applied to the mechanical ventilation man-machine asynchronous detection field, just as when a field expert of man-machine asynchronous recognition classifies (classifies) man-machine asynchronous types according to mechanical ventilation waveform data signals, the field expert focuses on waveform change conditions on certain sub-time sequence segments of the long-time sequence signal segments, and distinguishes different man-machine asynchronous types according to the sub-time sequence segments with obvious change differences, the attention mechanism can focus on distinguishing information of signal waveform segments with different man-machine asynchronous types, namely waveform time sequence segment information needing to be distinguished, so that important distinguishing information can be more intelligently noticed and learned, and better performance of distinguishing the man-machine asynchronous type is obtained.

Referring to fig. 1 to 5, a mechanical ventilation man-machine asynchronous detection and identification method based on attention mechanism of the present invention includes the following steps:

s1, constructing a one-dimensional convolution neural network model according to waveform data of the mechanical ventilation of the respirator;

step S2, respectively constructing a core level attention submodule and a time sequence attention submodule in sequence;

s3, sequentially connecting the core-level attention submodule and the time sequence attention submodule in series to form an organic integrated multi-attention comprehensive module;

s4, embedding and integrating a multi-attention integration module into the one-dimensional convolutional neural network model to form a multi-attention composite model;

and step S5, analyzing and judging the waveform data of the mechanical ventilation of the breathing machine through the multi-attention composite model.

In the invention, waveform data of Flow (Flow), airway pressure (Paw) and Volume (Volume) during mechanical ventilation of a respirator are used for judging two typical man-machine asynchrony (ineffective inspiration effort and double triggering) of interaction between a patient and the respirator, and according to the characteristic that different sections of corresponding waveform data pay different attention when experts manually judge the man-machine asynchrony, an attention mechanism is naturally introduced into a constructed one-dimensional convolutional neural network, so that the one-dimensional convolutional neural network integrating an attention module pays stronger attention to signal data sections of three groups of waveform data representing man-machine asynchrony characteristics, such as Flow, airway pressure and Volume, and the like, and the discrimination performance of the basic one-dimensional convolutional neural network can be better exerted to obtain better classification and identification performance.

The method mainly comprises a plurality of processes, namely, a one-dimensional convolution neural network model with a good structure is constructed for waveform data such as flow, airway pressure, capacity and the like, and the one-dimensional convolution neural network model is a basic use case; secondly, two attention submodules are constructed, the attention of the two attention submodules to a specific section of waveform data and different sections of convolution kernel characteristic values are deepened from two angles of a time sequence (time domain) and a kernel level characteristic domain respectively, the two attention sub-modules form an organic integrated multi-attention comprehensive module by a mode of firstly carrying out core-level attention and then carrying out time sequence attention in series (the invention can also adopt other modes to replace the mode of firstly carrying out core-level attention and then carrying out time sequence attention in series, for example, the mode of firstly carrying out time sequence attention and then carrying out core-level attention in series to combine the multi-attention comprehensive module or the mode of combining the core-level attention and the time sequence attention in parallel (parallel processing connection) to combine the multi-attention comprehensive module), the multi-attention comprehensive module can be flexibly embedded into a position behind a convolution structure of the one-dimensional convolution neural network according to requirements; thirdly, a multi-attention comprehensive module is well embedded and integrated into a one-dimensional convolution neural network module, so that the one-dimensional convolution neural network model based on the multi-attention mechanism can be trained like a conventional neural network model, and the whole model has good performance through certain parameter modulation; by applying the composite model, the man-machine asynchronous condition expressed by the relevant signal data of the breathing machine in the mechanical ventilation process of the patient in some scenes can be analyzed and judged, and reference and assistance are provided for doctors.

In the invention, two typical human-machine asynchronous types of double triggering and ineffective inspiration efforts are judged according to waveform data during mechanical ventilation of a respirator, a main body of the method is realized by adopting a composite model of a one-dimensional convolution neural network combined with a composite attention mechanism, the composite model mainly receives input of three-section signals of flow, airway pressure and capacity during mechanical ventilation of the respirator, the three-section signals flow through a designed multi-attention module after passing through two convolution layers of the convolution network, so that more effective characteristics in a plurality of groups of waveform data are extracted, and whether the two typical asynchronous types and what the specific asynchronous type are in the mechanical ventilation process of the respirator for a patient can be judged more easily according to the later stage of the more effective characteristic model.

During mechanical ventilation of the respirator, three groups of waveform data points of flow, airway pressure and volume are generated at a sampling rate of 50 Hz; when the strength of the inspiration of the patient reaches a certain threshold value, the breathing machine is triggered to give the patient a certain amount (time and volume) of ventilation support; when the patient transitions to the expiratory state, the ventilator is again triggered to remove ventilatory support.

In step S1, the waveform data includes flow, airway pressure, and volume during mechanical ventilation of the ventilator; collecting all data of flow, airway pressure and volume in a mechanical ventilation process of a breathing machine, namely normal breathing, double triggering and ineffective inspiration effort of the breathing machine, preprocessing all the collected data, and carrying out segmentation labeling; specifically, simulation verification is carried out based on real data, the data are derived from a known invasive respirator and a precision simulation lung, an experiment is carried out by utilizing a stable case of ARDS (acute respiratory distress syndrome) really and reliably simulated by the simulation lung, and two typical human-computer asynchronous events in the mechanical ventilation process of the respirator, namely a double-trigger and ineffective inspiration effort, are simulated based on the case.

The collected segments of respiratory events are segmented and labeled by a skilled technician trained in the art, so that all labeled data comprising three categories of Dual Trigger (DT), ineffective inspiratory effort (IEE) and normal breathing (NORM) segmented by breathing cycle are obtained. The initial full data with the labels can be used for the pre-processing of subsequent data and is finally used by an algorithm model; before training, firstly preprocessing the Flow (Flow), airway pressure (Paw) and Volume (Volume) data segments of a full-Volume ventilator during mechanical ventilation, standardizing and aligning the segmented and labeled data segments, then dividing the uniformly processed full-Volume data into a training set and a testing set according to the proportion of 7:3, thus completing the Flow of early-stage data processing, completing the processing and preparation work of the data, and then hierarchically constructing a multi-attention module and a composite model based on the module enhancement.

In step S2, a kernel-level attention submodule is first constructed, the kernel of which is composed of two models in which fully-connected layers are sequentially connected in series, the features of the input convolution kernel are processed respectively using one-dimensional global average pooling and one-dimensional global maximum pooling to obtain global maximum pooling features and global average pooling features, and then the global maximum pooling features and the global average pooling features are converted after being added to corresponding matrix position elements after passing through the two fully-connected layers of the sequentially-connected structures in parallel; specifically, the core of the kernel-level attention mechanism is composed of two fully-connected layer sequentially-connected models, for the feature of the feature map of the input convolution kernel, firstly, one-dimensional global average pooling and one-dimensional global maximum pooling are respectively used for processing, the global average pooling has feedback on each signal point on the feature map, the global maximum pooling only has feedback of gradient on the place with the maximum response in the feature map when gradient back propagation calculation is carried out, the global maximum pooling can supplement the global average pooling to a certain extent, then deformation operation is carried out on the data after the two pooling, and 1 middle dimension with the value of 1 is added, so that the pooled data is the same as the input data in the dimension number; after the deformation operation, the method is equivalent to compressing the feature map feature data of the input convolution kernel in a time sequence dimension, and the data are compressed and spread to the dimension of the convolution kernel, so that two sets of kernel-level preliminary features of maximum pooling and average pooling related to the dimension of the convolution kernel are extracted.

Note that the first of the two fully-connected layers of the kernel-level attention submodel reduces the last dimension of the input feature matrix by a certain ratio (for example, the ratio is 8, so that when the last dimension of the input feature matrix is 16, the last dimension of the input feature matrix passes through the first fully-connected layer and becomes 2), and the second fully-connected layer increases the last dimension of the feature matrix input from the first fully-connected layer, so that the last dimension (kernel-level dimension) of the feature matrix input from the first fully-connected layer initially is the same as the last dimension of the feature matrix output from the second fully-connected layer. When the feature map of the input convolution kernel is transformed by two pooling methods to obtain the global maximum pooling feature and the global average pooling feature (namely the two sets of kernel-level preliminary features mentioned above), the two sets of kernel-level preliminary features transformed by the pooling operation are added to corresponding matrix position elements after passing through the two sequentially-connected full-connection layer structures of the kernel in parallel, and the addition can fully enhance the attention to important kernel-level features in the feature map, promote classification and classificationIdentifying the performance of the model, converting the added result by using a hard _ sigmoid function, and multiplying the attention matrix obtained after conversion by the feature map matrix of the initial input convolution kernel according to elements (if necessary, executing a broadcasting mechanism) to obtain the final characteristics processed by the whole kernel-level attention submodule, wherein the process can be briefly described by a formula as follows:

wherein, W₁And W₀The weight of the multi-layer perceptron MLP is adopted, the corner mark k represents a kernel-level attention module, the kernel-level attention submodule can be roughly described by using FIG. 2, the process of constructing the kernel-level attention submodule can be explained in a more concise and concise manner by combining with FIG. 2, the constructed kernel-level attention submodule is A, the initial airway pressure, flow and capacity data are converted by a one-dimensional convolution structure to form the input features of the kernel-level attention submodule, and the shape of the input feature matrix can be set as [ None,50,32]]The input feature matrix is a 3-dimensional matrix, the 1 st dimension value is None, which represents the data volume input into the attention submodule according to batches, and None represents that no limitation is made; the 2 nd dimension value 50 represents the time-sequential dimension of the input feature matrix; the 3 rd dimension value 32 represents the kernel-level dimension of the input feature matrix; a1: parallel input feature matrixes are subjected to one-dimensional maximum pooling and one-dimensional average pooling to form two sets of kernel-level preliminary features, and both pooled shapes are [ None,32]Then converted into shape of [ None,1,32 ]]Two sets of nuclear-level preliminary features; a2: the two sets of nuclear-level preliminary features are respectively input into an MLP structure to be processed and converted to form two sets of nuclear-level processing features, and the shape of the two sets of nuclear-level processing features is unchanged and is [ None,1,32](ii) a A3: adding corresponding elements to the two groups of MLP processed nuclear-level processing characteristics, and forming a converted nuclear-level attention matrix after the added characteristics are acted by a hard _ sigmoid function, wherein shape is kept as [ None,1,32 ]](ii) a A4: the transformed kernel-level attention matrix and input feature matrixPerforming matrix multiplication operation of corresponding elements to obtain a final kernel-level attention feature matrix processed for the input features, wherein shape is [ None,50,32]]。

In step S2, after the kernel-level attention submodule is constructed, the time-sequence attention submodule is constructed, average pooling and maximum pooling operations are performed on the input features from the convolution kernel, an average pooling feature matrix and a maximum pooling feature matrix related to the time sequence are extracted, the average pooling feature matrix and the maximum pooling feature matrix are spliced to obtain a combined feature matrix, and the combined feature matrix is converted through a convolution network; specifically, the time series attention module construction process is as follows: for feature map features from convolution kernel in input, performing average pooling and maximum pooling operations on convolution kernel dimension of the feature matrix, which is equivalent to compressing the convolution kernel dimension, thereby extracting time-sequence-related average pooling and maximum pooling feature matrices, then splicing the two feature matrices on convolution kernel dimension to obtain a combined feature matrix, inputting the combined feature matrix into a one-dimensional convolution neural network of one layer, wherein the convolution kernel number of the convolution neural network is 1, an activation function adopts hard _ sigmoid without using bias and the like, after the convolution neural network is converted, the convolution kernel dimension of the combined feature matrix is compressed to 1, and finally, the time sequence attention matrix converted by the convolution network is multiplied by an initial input feature matrix from outside the time sequence attention submodule according to elements (if necessary, a broadcasting mechanism is also executed) to obtain a final feature map feature matrix which is subjected to the whole process The characteristics of the time sequence attention submodule after processing can be briefly described by the following formula:

wherein F represents the characteristics of the input,denotes the hard _ sigmoid function, f⁵Representing a one-dimensional convolution operation with a convolution kernel size of 5, the subscript t representing the time sequenceThe attention submodule, which can be schematically described with reference to fig. 3, can be described in a more concise manner with reference to fig. 3 to describe the process of constructing the time-series attention submodule, where the constructed time-series attention submodule is denoted as B, data processed by a one-dimensional convolution structure is generally used as a "transformed feature" of the time-series attention submodule, and a shape of the "transformed feature" can be denoted as [ None,50,32]. The build process may be as follows as shown in this figure; b1: for the input transformed feature matrix, maximum pooling and average pooling compression in kernel-level dimension are respectively performed to form two primary sub-feature matrices with pooling time sequences, wherein shape of each primary sub-feature matrix is [ None,50,1]. B2: and then splicing the two groups of pooled time sequence primary sub-feature matrixes to form a time sequence primary feature matrix, wherein shape of the time sequence primary feature matrix is [ None,50, 2]](ii) a B3: the time sequence primary characteristic matrix is input into a network of 1 convolution layer for processing conversion, the number of convolution kernels of the convolution network convolution layer is 1, the size of the convolution kernels is 5, an activation function also adopts a hard _ sigmoid function, a time sequence processing characteristic matrix is formed after conversion, and shape of the time sequence processing characteristic matrix is changed back to [ None,50,1](ii) a B4: matrix multiplication operation of corresponding elements is carried out on the converted time sequence processing characteristic matrix and the initial converted characteristic, and the final time sequence attention characteristic matrix processed on the converted characteristic is obtained as a result, and shape of the time sequence processing characteristic matrix is [ None,50,32]]。

In step S3, the core-level attention submodule and the time-series attention submodule are connected in series to form a multi-attention synthesis module, the input features are processed by the core-level attention submodule to form a core-level output feature matrix with core-level attention, and the core-level output feature flow is processed by the time-series attention submodule to form a time-series output feature; specifically, the core-level attention submodule and the time sequence attention submodule are connected in series to form a multi-attention comprehensive module, a scheme of firstly connecting the core-level attention and then connecting the time sequence attention in series is adopted, the combination module can enhance the attention of feature fragments of a plurality of levels and angles such as a core feature level, a signal time sequence level and the like on an input feature matrix, and the performance of a composite model using the attention comprehensive module is improved; the multi-attention module also has considerable flexibility and can be applied after convolution structures at different positions of a model containing one-dimensional convolution structures. The multi-attention module can be schematically described with reference to fig. 4, and the process of constructing the multi-attention integration module can be described in a more concise manner with reference to fig. 4, where a constructed multi-stage attention integration module is denoted as C, where the features output from the convolutional layer are used as the input feature matrix of the attention module, and shape is denoted as [ None,50,32], and the construction process can be as follows as shown in the figure; c1: processing the input features by a core-level attention submodule to form a core-level output feature matrix with core-level attention, wherein shape of the core-level output feature matrix is [ None,50,32 ]; c2: the core-level output feature is processed by the time sequence attention submodule to form a time sequence output feature, the shape of the time sequence output feature is [ None,50,32], the time sequence output feature processed by the time sequence attention submodule is the final output feature of the whole multi-level attention comprehensive module, namely, a redefined feature, the redefined feature is a comprehensive feature with the support of core level and time sequence attention, and the segment of the input data representing the feature data has larger weight attention.

In step S4, a multi-attention integration module is embedded and integrated into the one-dimensional convolutional neural network model to form a multi-attention composite model, the multi-attention composite model is sequentially composed of a convolutional layer, a maximum pooling layer, a first multi-attention layer, a convolutional layer, a maximum pooling layer, a second multi-attention layer, a convolutional layer, a fully-connected layer and a fully-connected layer from front to back, and the first multi-attention layer and the second multi-attention layer both include a kernel-level attention layer and a time-sequence attention layer; specifically, the one-dimensional convolutional neural network model is sequentially composed of a convolutional layer, a maximum pooling layer, a convolutional layer, a full-link layer and a full-link layer; the multi-attention module is respectively arranged behind the two maximum pooling layers, so that a multi-attention mechanism-based enhanced one-dimensional convolutional neural network composite model is formed, and the composite model has classification and identification performances which are obviously superior to those of a basic convolutional model and has practical value; the composite model can be schematically described with reference to fig. 5.

In step S5, the waveform data during mechanical ventilation of the ventilator is analyzed and determined by the multi-attention composite model, and when the waveform data is detected as data in the asynchronous event during mechanical ventilation of the ventilator during a double-triggering or ineffective inspiratory effort, the asynchronous event can be determined as a double-triggering or ineffective inspiratory effort.

The nuclear level attention mechanism and the time sequence attention mechanism can extract the characteristics needing attention in different aspects in three groups of waveform data of flow, airway pressure and capacity from a plurality of levels, and the attention of the characteristics is enhanced, so that the classification performance is improved; the comprehensive attention module formed by combining the two attention sub-modules has flexible deployment degree, can be placed behind different convolution layers of various one-dimensional convolution neural networks such as use cases to strengthen the attention degree on characteristics from different angles and positions, and the composite model integrating the attention of different levels into the neural network with the one-dimensional convolution structure and having a multi-attention mechanism can better capture data segments representing human-machine asynchronous types in three-guided waveform data of flow, airway pressure and capacity, and can better identify the typing of the human-machine asynchronous types.

The invention has been verified through experimental simulation, the result is excellent, the proposed composite classification model based on the multi-attention mechanism has good identification performance on detection of two asynchronous types of double triggering and ineffective inspiration effort of typical cases, and the average identification accuracy is about 0.9621.

For different random numbers, the corresponding accuracy rate cases of a basic case 1DCNN (one-dimensional convolutional neural network) model and a 1DCNN composite model based on multi-attention enhancement are shown in table 1; it can be seen from table 1 that the proposed multi-attention mechanism-based enhanced 1DCNN composite model has better comprehensive performance.

Table 1 table for comparing the accuracy of the multi-attention machine system enhanced composite model with the accuracy of the basic case model based on different random numbers:

the above description is only an embodiment of the present invention, and not intended to limit the scope of the present invention, and all modifications of equivalent structures and equivalent processes performed by the present specification and drawings, or directly or indirectly applied to other related technical fields, are included in the scope of the present invention.