阅读说明：本技术 用于设备监测的装置 (Device for monitoring equipment ) 是由 M·基奥亚 S·苏比亚 A·M·科特里瓦拉 I·阿米海 于 2020-03-18 设计创作，主要内容包括：本发明涉及一种用于设备监测的装置。该装置包括输入单元、处理单元和输出单元。输入单元配置成为处理单元提供用于设备的项的时间传感器数据的多个批。时间传感器数据的每个批包括作为时间的函数的多个时间传感器值。处理单元配置成处理时间传感器数据的多个批,以确定谱传感器数据的多个批。谱传感器数据的每个批包括作为频率的函数的多个谱传感器值。处理单元配置成实现至少一个统计过程算法,以处理对于谱传感器数据的多个批的多个谱传感器值,以便确定多个指标值。对于谱传感器数据的每个批,存在通过统计过程算法的每个所确定的指标值。每个统计过程算法具有关联阈值,以及处理单元配置成利用至少一个阈值和多个指标值来确定感兴趣的谱传感器数据的批,其具有大于对于关联统计过程算法的阈值的指标值。处理单元配置成基于对于感兴趣的谱传感器数据的批的多个谱传感器值来确定感兴趣的频率范围。(The invention relates to a device for monitoring equipment. The device comprises an input unit, a processing unit and an output unit. The input unit is configured to provide the processing unit with a plurality of batches of time sensor data for items of equipment. Each batch of temporal sensor data includes a plurality of temporal sensor values as a function of time. The processing unit is configured to process the multiple batches of temporal sensor data to determine multiple batches of spectral sensor data. Each batch of spectral sensor data includes a plurality of spectral sensor values as a function of frequency. The processing unit is configured to implement at least one statistical process algorithm to process a plurality of spectral sensor values for a plurality of batches of spectral sensor data to determine a plurality of indicator values. For each batch of spectral sensor data, there is an index value determined by each of the statistical process algorithms. Each statistical process algorithm has an associated threshold value, and the processing unit is configured to determine a batch of spectral sensor data of interest with at least one threshold value and a plurality of indicator values having an indicator value greater than the threshold value for the associated statistical process algorithm. The processing unit is configured to determine a frequency range of interest based on a plurality of spectral sensor values for a batch of spectral sensor data of interest.)

1. An apparatus for device monitoring, the apparatus comprising:

-an input unit;

-a processing unit; and

-an output unit;

wherein the input unit is configured to provide the processing unit with a plurality of batches of temporal sensor data for items of equipment, wherein each batch of temporal sensor data comprises a plurality of temporal sensor values as a function of time;

wherein the processing unit is configured to process the plurality of batches of temporal sensor data to determine a plurality of batches of spectral sensor data, wherein each batch of spectral sensor data comprises a plurality of spectral sensor values as a function of frequency;

wherein the processing unit is configured to implement at least one statistical process algorithm to process the plurality of spectral sensor values for the plurality of batches of spectral sensor data to determine a plurality of indicator values, wherein for each batch of spectral sensor data there is an indicator value determined by each of the statistical process algorithms;

Wherein each statistical process algorithm has an associated threshold value, and wherein the processing unit is configured to determine a batch of spectral sensor data of interest with at least one threshold value and the plurality of indicator values having an indicator value greater than the threshold value for the associated statistical process algorithm; and

wherein the processing unit is configured to determine a frequency range of interest based on the plurality of spectral sensor values for the batch of spectral sensor data of interest.

2. The apparatus of claim 1, wherein a time period between adjacent batches of temporal sensor data is greater than a time period between adjacent sensor data within a batch.

3. The apparatus of any of claims 1-2, wherein the at least one statistical process algorithm comprises hotelling.

4. The apparatus of claim 3, wherein for each batch of spectral sensor data there is an indicator value determined by the Hotelling statistics.

5. The apparatus of any of claims 3-4, wherein the batch of spectral sensor data of interest is determined when the indicator value determined for that batch by the Hotelling statistic is greater than the threshold value associated with the Hotelling statistic.

6. The apparatus of any of claims 1-5, wherein the at least one statistical process algorithm comprises squared prediction error or Qstatistic.

7. The apparatus of claim 6, wherein for each batch of spectral sensor data there is an index value determined by the squared prediction error or Q-statistic.

8. The apparatus of any of claims 5-7, wherein the batch of spectral sensor data of interest is determined when the indicator value determined for that batch by the squared prediction error or Qstatistic is greater than the threshold value associated with the squared prediction error or Qstatistic.

9. The apparatus of claim 8 when dependent on claim 5 or any one of claims 6 to 7 when dependent on claim 5, wherein the batch of spectral sensor data of interest is determined when the indicator value determined for that batch by the hotelling statistic is greater than the threshold value associated with the hotelling statistic; or wherein the batch of spectral sensor data of interest is determined when the index value determined for that batch by the squared prediction error or Q statistic is greater than the threshold value associated with the squared prediction error or Q statistic.

10. The apparatus of any of claims 1-9, wherein the determination of the plurality of batches of spectral sensor data comprises utilizing a fourier transform algorithm on the temporal sensor value for each of the plurality of batches of temporal sensor data.

11. The apparatus of any one of claims 1-10, wherein the processing unit is configured to subdivide the spectral values for the batch of spectral data of interest into a plurality of frequency ranges, and wherein a frequency range of interest is determined as a frequency range exhibiting larger values than values associated with other frequency ranges.

12. A system for device monitoring, the system comprising:

-at least one sensor; and

-an apparatus for equipment monitoring according to any of claims 1-11;

wherein the at least one sensor is configured to acquire the plurality of batches of temporal sensor data.

13. A method for device monitoring, the method comprising:

a) providing a plurality of batches of temporal sensor data for items of equipment, wherein each batch of temporal sensor data comprises a plurality of temporal sensor values as a function of time;

b) processing the plurality of batches of temporal sensor data to determine a plurality of batches of spectral sensor data, wherein each batch of spectral sensor data includes a plurality of spectral sensor values as a function of frequency;

c) Implementing at least one statistical process algorithm to process a plurality of spectral sensor values for a plurality of batches of spectral sensor data in order to determine a plurality of indicator values, wherein for each batch of spectral sensor data there is an indicator value determined by each of the statistical process algorithms;

d) determining a batch of spectral sensor data of interest with a threshold value and a plurality of index values for each statistical process algorithm having an index value greater than the threshold value of the associated statistical process algorithm; and

e) determining a frequency range of interest based on the plurality of spectral sensor values for the batch of spectral sensor data of interest.

14. A computer program element for controlling an apparatus according to one of the claims 1 to 11 and/or a system according to claim 12, which, when being executed by a processor, is configured to carry out the method according to claim 13.

15. A computer readable medium having stored the computer program element of claim 14.

Technical Field

The present invention relates to an apparatus for device monitoring, a system for device monitoring, a method for device monitoring and to a computer program element and a computer readable medium.

Background

Currently in the process industry, inspection and maintenance of equipment with moving parts (e.g., rotating parts) contributes primarily to operating expenses. Condition monitoring of, for example, rotating equipment, typically relies on the collection and analysis of various types of sensors, such as acceleration, velocity and displacement sensors, microphones, acoustic arrays, torque meters, encoders, tachometers, image cameras, fiber optic sensors, thermal sensors and stress sensors. The signal processing based algorithm is then used to calculate health indicators for various failure modes for various equipment types (e.g., pumps, transmissions, and bearings). These health indicators are then tracked to detect or extrapolated to predict specific equipment failures.

Several solutions for monitoring of rotating equipment are known. These typical solutions rely on monitoring simple indicators (indexes), such as ISO 10186-1,7, which reflect the overall vibration level and compare to a predefined threshold, which is typically conservatively set.

Indicators associated with signal processing methods related to specific faults of specific components (e.g. bearings, gearbox, pump) have also been implemented. However, the industrial use of such high-grade indices still suffers from the necessity of: a specific detection threshold is defined for each rotating device that varies with load, rotational speed, and other operating parameters. The implementation of these methods is cumbersome and time consuming due to the lack of ability to systematically select such thresholds.

There is a need to address this problem.

Disclosure of Invention

It would therefore be advantageous to have an improved capability for monitoring a device having moving parts in order to detect whether the device is degrading or is about to degrade or is damaged.

The object of the invention is solved with the subject matter of the independent claims, wherein further embodiments are incorporated in the dependent claims.

In a first aspect, there is provided an apparatus for equipment monitoring, the apparatus comprising:

-an input unit;

-a processing unit; and

-an output unit.

The input unit is configured to provide the processing unit with a plurality of batches (batch) of time sensor data for items of the equipment. Each batch of temporal sensor data includes a plurality of temporal sensor values as a function of time. The processing unit is configured to process the multiple batches of temporal sensor data to determine multiple batches of spectral sensor data. Each batch of spectral sensor data includes a plurality of spectral sensor values as a function of frequency. The processing unit is configured to implement at least one statistical process algorithm to process a plurality of spectral sensor values for a plurality of batches of spectral sensor data to determine a plurality of indicator values. For each batch of spectral sensor data, there is an index value determined by each of the statistical process algorithms. Each statistical process algorithm has an associated threshold value, and the processing unit is configured to determine a batch of spectral sensor data of interest with at least one threshold value and a plurality of indicator values having an indicator value greater than the threshold value for the associated statistical process algorithm. The processing unit is configured to determine a frequency range of interest based on a plurality of spectral sensor values for a batch of spectral sensor data of interest.

In this way, an unsupervised technique for monitoring industrial equipment (e.g., rotating equipment) is implemented. An abnormal increase in the overall level of the sensed value (e.g., vibration) is detected and the spectral components most correlated to the detected abnormal increase in the sensed value are isolated. Thus, the analyst can have a first indication of the type of fault affecting the device based on knowledge of the spectral components contributing to the increase in the overall sensed value level (e.g., vibration of the rotating device).

In other words, the above diagnostic system is provided with a higher level of knowledge of simple changes in sensed values (e.g., vibrations), wherein the new technique enables abnormal sensed value levels to be detected and the associated spectrum (frequency components) that is the primary cause of that abnormality to be identified to be determined.

In an example, a time period between adjacent batches of temporal sensor data is greater than a time period between adjacent sensor data within a batch.

In an example, the at least one statistical process algorithm includes hotelling.

In an example, for each batch of spectral sensor data, there is an indicator value determined by hotelling statistics.

In an example, a batch of spectral sensor data of interest is determined when an indicator value determined for that batch by hotelling statistics is greater than a threshold value associated with hotelling statistics.

In an example, the at least one statistical process algorithm includes a squared prediction error or Q statistic.

In an example, for each batch of spectral sensor data, there is an index value determined by the squared prediction error or Q statistic.

In an example, a batch of spectral sensor data of interest is determined when an index value determined for that batch by the squared prediction error or Q statistic is greater than a threshold value associated with the squared prediction error or Q statistic.

In an example, when the indicator value determined for that batch by hotelling statistics is greater than a threshold value associated with hotelling statistics, determining a batch of spectral sensor data of interest; or wherein a batch of spectral sensor data of interest is determined when the index value determined for that batch by the squared prediction error or Q statistic is greater than a threshold value associated with the squared prediction error or Q statistic.

In other words, both indicators (determined using SPE or Q statistics or T2 hotelling statistics) are monitored separately and if one of them exceeds its limit, an analysis of the corresponding batch is performed to determine the problematic spectral frequencies.

In other words, two individual statistical control charts (calculated or index values for a batch) are used to detect anomalous observations by comparing them to threshold limits. Hotelling's T2 statistics are used in the principal component space, while squared prediction error (SPE or Q) statistics are used in the residual space.

In an example, the determining of the plurality of batches of spectral sensor data includes utilizing a fourier transform algorithm for the temporal sensor value of each of the plurality of batches of temporal sensor data.

In an example, the processing unit is configured to subdivide spectral values for a batch of spectral data of interest into a plurality of frequency ranges. The frequency range of interest is determined to be a frequency range that exhibits a value that is greater than values associated with other frequency ranges.

In a second aspect, there is provided a system for equipment monitoring, the system comprising: at least one sensor configured to acquire a plurality of batches of temporal sensor data; and an apparatus for equipment monitoring according to the first aspect.

In a third aspect, there is provided a method for device monitoring, the method comprising:

a) providing a plurality of batches of temporal sensor data for items of equipment, wherein each batch of temporal sensor data comprises a plurality of temporal sensor values as a function of time;

b) processing the plurality of batches of temporal sensor data to determine a plurality of batches of spectral sensor data, wherein each batch of spectral sensor data includes a plurality of spectral sensor values as a function of frequency;

c) Implementing at least one statistical process algorithm to process a plurality of spectral sensor values for a plurality of batches of spectral sensor data in order to determine a plurality of indicator values, wherein for each batch of spectral sensor data there is an indicator value determined by each of the statistical process algorithms;

d) determining a batch of spectral sensor data of interest with a threshold value for each statistical process algorithm and a plurality of index values having index values greater than the threshold value for the associated statistical process algorithm; and

e) the frequency range of interest is determined based on a plurality of spectral sensor values for a batch of spectral sensor data of interest.

According to another aspect, there is provided a computer program element for controlling an apparatus or system as previously described, which, when being executed by a processing unit, is adapted to carry out the method steps as previously described.

According to another aspect, there is also provided a computer readable medium having stored the computer element as previously described.

The above aspects and examples will become apparent from and elucidated with reference to the embodiments described hereinafter.

Drawings

Exemplary embodiments will now be described with reference to the following drawings:

FIG. 1 shows an example of a principal component space spanned by two principal components and the remaining space of a data set for three process variables;

FIG. 2 shows a schematic representation of examples of determined indicators for different batches of statistical limits for one of those statistics, either calculated using T2 Hotelling statistics or calculated using PRE or Q statistics; and

FIG. 3 shows an example of a spectral frequency energy spectrum for a batch that exceeds a statistical limit.

Detailed Description

The presently provided apparatus, systems, and methods for equipment monitoring are now described in detail with reference to fig. 1-3. Examples of an apparatus for device monitoring include an input unit, a processing unit, and an output unit. The input unit is configured to provide the processing unit with a plurality of batches of time sensor data for items of equipment. Each batch of temporal sensor data includes a plurality of temporal sensor values as a function of time. The processing unit is configured to process the multiple batches of temporal sensor data to determine multiple batches of spectral sensor data. Each batch of spectral sensor data includes a plurality of spectral sensor values as a function of frequency. The processing unit is configured to implement at least one statistical process algorithm to process a plurality of spectral sensor values for a plurality of batches of spectral sensor data to determine a plurality of indicator values. For each batch of spectral sensor data, there is an index value determined by each of the statistical process algorithms. Each statistical process algorithm has an associated threshold value, and the processing unit is configured to determine a batch of spectral sensor data of interest with at least one threshold value and a plurality of indicator values having an indicator value greater than the threshold value for the associated statistical process algorithm. The processing unit is configured to determine a frequency range of interest based on a plurality of spectral sensor values for a batch of spectral sensor data of interest.

Thus, the apparatus is capable of operating in real time, as well as being at a site of a device, and monitoring that device. Alternatively, the device can analyze previously acquired data to determine abnormal behavior and those frequencies at which the frequency is problematic.

According to an example, a time period between adjacent batches of temporal sensor data is greater than a time period between adjacent sensor data within a batch.

In an example, the time period between adjacent batches of temporal sensor data is one of: 1 hour; 2 hours; 3 hours; 4 hours; 5 hours; 6 hours; 7 hours; 8 hours; 12 hours; for 24 hours. The time period between batches can be different from the time periods described above.

In an example, the time period between adjacent sensor data within a batch is one of: 0.0001 s; 0.0005 s; 0.001 s; 0.002 s; 0.003 s; 0.004 s; 0.005 s; 0.01 s; 0.02 s; 0.05 s.

In an example, each batch of sensor data is acquired over a time period of one of: 10 s; 20 s; 30 s; 60 s; 120 s.

Thus, for example, the sensor is able to acquire data at a rate of 1 kHz in 60 seconds. Then wait 6 hours and again acquire data at a rate of 1 kHz in 60 seconds. However, different rates of sensor data acquisition over different time periods can be utilized, and with different durations between batches of sensor data.

According to an example, the at least one statistical process algorithm comprises hotelling statistics.

According to an example, for each batch of spectral sensor data, there is an index value determined by hotelling statistics.

According to an example, a batch of spectral sensor data of interest is determined when the index value determined for that batch by the hotelling statistics is greater than a threshold value associated with the hotelling statistics.

According to an example, the at least one statistical process algorithm comprises a squared prediction error or Q-statistic.

According to an example, for each batch of spectral sensor data, there is an index value determined by the squared prediction error or Q-statistic.

According to an example, a batch of spectral sensor data of interest is determined when an index value determined for that batch by the squared prediction error or Q-statistic is greater than a threshold value associated with the squared prediction error or Q-statistic.

According to an example, a batch of spectral sensor data of interest is determined when the index value determined for that batch by the hotelling statistics is greater than a threshold value associated with the hotelling statistics. Alternatively, a batch of spectral sensor data of interest is determined when the index value determined for that batch by the squared prediction error or Q statistic is greater than a threshold value associated with the squared prediction error or Q statistic.

According to an example, the determining of the plurality of batches of spectral sensor data includes utilizing a fourier transform algorithm for the temporal sensor value of each of the plurality of batches of temporal sensor data.

In an example, the fourier transform algorithm is a fast fourier transform algorithm.

According to an example, the processing unit is configured to subdivide the spectral values of the batch of spectral data of interest into a plurality of frequency ranges, and wherein the frequency range of interest is determined as a frequency range exhibiting a larger value than values associated with other frequency ranges.

In an example, the frequency range of interest is determined as a frequency range exhibiting a spectral power value that is greater than spectral power values associated with other frequency ranges.

In an example, the sensor data is rotational sensor data.

It is therefore appreciated that the above-described apparatus, when coupled to at least one sensor (which acquires multiple batches of temporal sensor data), provides a system for equipment monitoring.

Moreover, an example relates to a method for device monitoring, the method comprising:

a) providing a plurality of batches of temporal sensor data for items of equipment, wherein each batch of temporal sensor data comprises a plurality of temporal sensor values as a function of time;

b) Processing the plurality of batches of temporal sensor data to determine a plurality of batches of spectral sensor data, wherein each batch of spectral sensor data includes a plurality of spectral sensor values as a function of frequency;

c) implementing at least one statistical process algorithm to process a plurality of spectral sensor values for a plurality of batches of spectral sensor data in order to determine a plurality of indicator values, wherein for each batch of spectral sensor data there is an indicator value determined by each of the statistical process algorithms;

d) determining a batch of spectral sensor data of interest with a threshold value for each statistical process algorithm and a plurality of index values having index values greater than the threshold value for the associated statistical process algorithm; and

e) the frequency range of interest is determined based on a plurality of spectral sensor values for a batch of spectral sensor data of interest.

In an example, a time period between adjacent batches of temporal sensor data is greater than a time period between adjacent sensor data within a batch.

In an example, the time period between adjacent batches of temporal sensor data is one of: 1 hour; 2 hours; 3 hours; 4 hours; 5 hours; 6 hours; 7 hours; 8 hours; 12 hours; for 24 hours.

In an example, the time period between adjacent sensor data within a batch is one of: 0.0001 s; 0.0005 s; 0.001 s; 0.002 s; 0.003 s; 0.004 s; 0.005 s; 0.01 s; 0.02 s; 0.05 s.

In an example, each batch of sensor data is acquired over a time period of one of: 10 s; 20 s; 30 s; 60 s; 120 s.

In an example, in step c), the at least one statistical process algorithm comprises hotelling.

In an example, for each batch of spectral sensor data, there is an indicator value determined by hotelling statistics.

In an example, in step d), a batch of spectral sensor data of interest is determined when the indicator value determined for that batch by hotelling statistics is greater than a threshold value associated with hotelling statistics.

In an example, in step c), the at least one statistical process algorithm comprises a squared prediction error or Q-statistic.

In an example, for each batch of spectral sensor data, there is an index value determined by the squared prediction error or Q statistic.

In an example, in step d), a batch of spectral sensor data of interest is determined when the index value determined for that batch by the squared prediction error or Q statistic is greater than a threshold value associated with the squared prediction error or Q statistic.

In an example, a batch of spectral sensor data of interest is determined when an indicator value determined for that batch by hotelling statistics is greater than a threshold value associated with hotelling statistics. Alternatively, a batch of spectral sensor data of interest is determined when the index value determined for that batch by the squared prediction error or Q statistic is greater than a threshold value associated with the squared prediction error or Q statistic.

In an example, step b) includes utilizing a fourier transform algorithm for the time sensor value for each of the plurality of batches of time sensor data.

In an example, the fourier transform algorithm is a fast fourier transform algorithm.

In an example, the sensor data is rotational sensor data.

The following detailed description relates to monitoring of rotating equipment. In the new approach, multivariate statistical process control is used in the new approach. MSPCs have been used to monitor industrial processes and support process operators or process engineers in troubleshooting process anomalies-see Kresta, J. V., Macgregor, J. F., and Marlin, T.E. (1991). "multivariable static monitoring of process operating performance" (The Canadian Journal of Chemical Engineering, 69(1), 35-47.)

However, this approach in a variation has been used for monitoring of devices (e.g., rotating devices). While the standard MSPC approach uses time series collected from process sensors, it has been found that rotating equipment can be better characterized by spectral features. It has been found that faults or malfunctions in rotating equipment have features that are localized in the frequency domain but not necessarily in the time domain due to the periodicity caused by the rotation.

In summary, in the new technique, a T2 hotelling metric and a Q or Squared Prediction Error (SPE) metric are calculated for each batch of collected sensor values and compared to appropriate statistical limits. In a particular example, the "vibration" measurements are performed in a "batch" manner, i.e., measurements are recorded periodically over a given period of time. Acceleration/velocity sensor data measurements are recorded, for example, every 6 hours. In the following, reference is made to a time index (time index) k, where k =1, 2, 3, ·, n denotes a batch of data acquired at a relative time of 6 hours, 12 hours, 18 hours,..,. n × 6 hours at a rate of 1 KHz over a time period of, for example, one minute. Thus, in this particular example, each batch or record of data acquired every 6 hours has 1000 × 60=60000 data samples or values.

A T2 hotelling and Q contribution plot is determined from one or more identified batches of the plurality of batches acquired every 6 hours. The T2 and Q contribution plots are then used to evaluate the frequency bin (frequency bin) that is most correlated with the T2 and Q index deviations from their corresponding statistical limits. This indication allows the user to monitor the change in the spectrum of the vibration signal and helps the user understand the type of fault the rotating device is experiencing.

Thus, two individual statistical control charts are used to detect anomalous observations by comparing them to threshold limits. Hotelling's T2 statistics are used in the principal component space and squared prediction error (SPE or Q) statistics are used in the residual space.

Monitoring in principal component space

When the scores follow a gaussian multivariate distribution, hotelling statistics are used. The score has a zero mean, and the estimated sample covariance matrixIs a diagonal matrix. For observationThe Hotelling statistics for t-scores are given by

Where P is a matrix whose columns are the load vectors, NPC is the number of principal components retained in the model,is a diagonal matrix whose elements are the eigenvalues of the sample covariance matrix (singular values of the data matrix X) that remain in the model in descending order of magnitude. T is the score in the principal component space.

Holterin (a)The statistics are the sum of the scaling scores. It combines the information from all scores into a single index. When calculatingWhen the system is counted, only the load corresponding to the larger singular value is included. The smaller singular values corresponding to noise are inverted (inverted) in the calculation of this statistic. The exclusion of these smaller singular values allows for a better representation of the process behavior and robust anomaly detection within the model.Statistics are the distance between the projection of the observed principal component space and the origin of the principal component space.Threshold value ofCalculated using the following equation

WhereinIs the confidence level and N is the number of observations in the data matrix.With a Fisher distribution of NPC, N-NPC degrees of freedomAnd (4) key points. From the above equation, the control limit of the score can be derived

The above formula describing the confidence for the scoreElliptical zones (see fig. 1 for an example with m =3, NPC = 2). Variable pairs at given observationsIs given by

Monitoring in the remaining space

The monitoring in the remaining space uses Q statistics defined below,

whereinIs the row vector of the error matrix E. The Q statistic is the euclidean norm of the observed deviations from their projection onto the principal component space. The control limit of Q can be approximated as

WhereinIs andthe value of the normal distribution to which the percentile corresponds,

and is。

Control limitA threshold value representing random variations in the process. The individual contribution of a variable to Q at a given observation is given by the element of E corresponding to the observation.

Referring to FIG. 1, this shows a data set that includes three variables. It can be seen that the three-dimensional data points (open circles) can be reduced to the two dimensions spanned by the principal components with a first principal component (along the major axis of the ellipse) and a second principal component (the minor axis of the ellipse) having higher explained variance. Consider a new observation (labeled as a solid circle of PC 1) that lies in the two-dimensional space spanned by the principal components, but lies in the scoreOutside the control limits of (c). New data points marked by other solid circles were found to be acceptable when projected into the principal component space, although the data points are different from other points in the model. This variation is captured by Q statistics.

Therefore, Multivariate Statistical Process Control (MSPC) has been used to monitor industrial equipment in new ways. Principal Component Analysis (PCA) can be utilized. PCA allows a dataset to be represented over a lower dimensional space. In addition, it separates the observation space into two subspaces. One subspace captures the process trend, while the other subspace captures the effect of random noise or new anomaly changes that are not part of the model.

Setting data matrixThere are n observations collected from m sensors. PCA is a data reduction technique that extracts the maximum variance in the data in orthogonal directions called principal components. These principal components are linear combinations of variables that contain useful information (variability of the process data). PCA decomposes the data matrix into orthogonal vectors called the payload vector (p) and the score vector (t). Decomposition is performed using Singular Value Decomposition (SVD) of the data matrix X:

whereinComprising real non-negative singular values in order of decreasing amplitude ( ). Right eigenvectorIs the load vector. To avoid modeling the noise present in the data, only the r larger eigenvalues and their corresponding load vectors are retained. The amount of variance captured in the direction of principal component i is. The data matrix X, when projected onto the space formed by the individual load vectors (p), is

WhereinCalled the scoring matrix, and the columns of t are orthogonal. From the first column of the scoring matrixThe variance ratio captured (corresponding to the first principal component of X) is represented by the second columnThe captured variance is large. Score againThe projection into the original m-dimensional space is given,

whereinIs the modeling information.Andthe difference between them is called residual matrix E and the space spanned by the residual matrix is called residual space. This residual space corresponds to the variance in the smaller eigenvalues that are not included in the model.

To test new observation vectorsThe vectors are projected onto the model. The projection of the test vector into the space formed by the load vectors gives the score of the test data.

Thus, returning to the acquisition and processing of sensor data, the deviceA vector representing the collected sensor values, e.g., accelerometer measurements collected starting at time k within time interval T.

Is provided withCalculated within a time interval (k, k + T) using, for example, a Fast Fourier Transform (FFT) algorithmSpectrum of (a). Each batch calculation of collected sensor values. k =1, 2, 3, 4,. n, where n is the number of batches at m frequency channels. As described above, each batch can be taken every 6 hours, and each batch can have, for example, 6000 data samples taken every 0.001 seconds. However, each batch can be acquired with a different time period between them, such that there are, for example, 2 hours between two batches, and then the next batch is acquired, for example, after 6 hours. However, sensor values are read every Ts seconds (e.g., every 0.001s) within the batch.

Thus, a batch is a series of sensor values collected during the time interval T. For each batch, a fourier transform is calculated to estimate the corresponding power spectrum. The actual implementation of the power spectrum relies on a numerical algorithm that computes the fourier transform: fast fourier transform algorithm (FFT). The FFT computes discrete spectra, i.e., values of the spectrum at a finite number of frequencies. The frequencies or frequency bands in which the power spectrum is evaluated are referred to as "frequency channels".

Superimposing the calculated spectra for all batches yields the following matrix

Principal Component (PCA) decomposition of X yields the following approximation:

wherein

And E is the residual (approximate) error, t is the vector of scores, and p is the vector of loads.

The T2 Hotelling index and the Q (SPE) index are then calculated for each batch index (batch index) and compared to the appropriate statistical limit. This is shown in fig. 2. In fig. 2, the T2 hotelling and PRE or Q indices are calculated for each batch, where k is the batch (dataset index) as described above. As shown in FIG. 2, the batch 12 exceeds the limit, where only one set of indicator values calculated for the T2 Hotelling statistic is shown and only one threshold is shown for the sake of brevity only, but in practice there are two plots, one for the T2 Hotelling statistic and one for the PRE or Q statistical analysis. Thus, as shown, batch k =12 is above the statistical limit T2_limOr Q_lim. This batch and above the statistical limit T2_limOr Q_limAre associated with abnormal vibration levels. Therefore, these data sets (batches) are then investigated by analyzing the contribution plots to determine which frequency(s) best explain the high value of T2 or Q. This contribution plot is shown in fig. 3 for batch 12 and is determined from the FFT-determined spectral content for that batch with an abnormal vibration level. In this way, the T2 and Q contribution plots are used to evaluate the frequency bins most correlated with the T2 and Q index deviations for the corresponding statistical limits. In FIG. 3, frequency bins [ l00Hz-150 Hz ] for batch 12 ]Significantly higher than the other spectral components, indicating that the detected change in the Q or T2 indicator passes through [ l00Hz-150 Hz]The presence of components in the frequency range. This indication allows the user to monitor the change in the spectrum of the vibration signal and helps him to understand the type of fault that the rotating device is experiencing.

In another exemplary embodiment, a computer program or a computer program element is provided, which is characterized by being configured to perform the method steps of the method according to one of the preceding embodiments on a suitable system.

The computer program element may thus be stored on a computer unit, which may also be an integral part of the embodiment. This calculation unit may be configured to perform or cause to be performed the steps of the above-described method. Further, it may be configured to operate the components of the above-described apparatus and/or system. The computing unit can be configured to operate automatically and/or execute commands of a user. The computer program may be loaded into working memories of a data processor. Thus, the data processor may be equipped to perform a method according to one of the preceding embodiments.

According to a further exemplary embodiment of the present invention, a computer-readable medium (for example a CD-ROM) is proposed, wherein the computer-readable medium has stored thereon a computer program element, which is described by the previous section.

While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive. The invention is not limited to the disclosed embodiments. Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from a study of the drawings, the disclosure, and the appended claims.