A fault detection method based on two-layer distributed and integrated Bayesian reasoning

By adopting a fault detection method of two-layer distributed and integrated Bayesian inference in industrial systems, the problem of failure to fully consider data distribution characteristics and variable dynamic correlation in the prior art is solved, and more efficient fault detection performance is achieved .

CN117093932BActive Publication Date: 2024-11-29SHANGHAI SHIP & SHIPPING RES INST CO LTD +1

Patent Information

Application Number
CN202311082891.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-08-25
Publication Date
2024-11-29
Estimated Expiration
2043-08-25

AI Technical Summary

Technical Problem

The existing industrial system fault detection methods fail to fully consider the distribution characteristics of the data, the dynamic correlation of variables, local similarity, and the inability to emphasize the different contribution information of each local model, resulting in poor detection performance.

Method used

Using a fault detection method based on two-layer distributed and integrated Bayesian inference, the first-layer block division and the second-layer block division are performed on the samples by installing a state parameter monitoring sensor in an industrial system, and Gaussian blocks and non-Gaussian blocks are processed respectively. The control limits of each cluster subblock are calculated using PCA and ICA methods, and finally, the global monitoring statistics are calculated through Bayesian inference to achieve fault detection.

Benefits of technology

It effectively improves the performance of fault detection, can more comprehensively consider the distribution characteristics of data, the dynamic correlation and local similarity of variables, and emphasizes the different contribution information of each local model, thereby improving detection accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117093932B_ABST
    Figure CN117093932B_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of industrial system fault detection, and specifically to a fault detection method based on double-layer distribution and integrated Bayesian reasoning, including two stages of offline modeling and model testing. The present invention aims to solve the problem that the current industrial system fault detection method fails to fully consider the distribution characteristics of data, the dynamic correlation of variables, local similarity, and fails to emphasize the different contribution information of each local model, which ultimately leads to poor detection performance. A fault detection method based on double-layer distribution and integrated Bayesian reasoning is constructed. The double-layer distribution strategy is used to fully mine the local information such as the distribution characteristics, dynamics, and similarity of the monitoring data. The integrated Bayesian reasoning strategy highlights the contribution of different local monitoring models, and ultimately effectively improves the performance of fault detection. Therefore, the present invention is suitable for fault detection of dynamic processes with Gaussian and non-Gaussian concurrent characteristics.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of industrial system fault detection, and in particular to a fault detection method based on double-layer distributed and integrated Bayesian reasoning. Background Art

[0002] As the scale of modern industry continues to expand, the increasingly complex industrial systems are more likely to cause major safety accidents. In order to ensure the long-term stable operation of industrial systems, timely discovery and accurate detection of faults is a research difficulty in the industrial field.

[0003] The existing principal component analysis (PCA) method assumes that the data follows a Gaussian distribution and that samples at different times are independent. However, in actual industry, the above assumptions are not necessarily met. To this end, the dynamic PCA (DPCA) model and slow feature analysis are proposed to solve the problem of dynamic correlation, and independent component analysis (ICA) and dynamic independent component analysis (DICA) are proposed to deal with non-Gaussian processes. However, the above centralized methods cannot handle large-scale industrial data well. In recent studies, some scholars use distributed methods to handle large-scale data, such as distributed principal component analysis, distributed slow feature analysis, distributed cointegration analysis, and distributed canonical correlation analysis. However, the above distributed methods do not consider the distribution characteristics of the data, resulting in each block containing both Gaussian variables and non-Gaussian variables. The same method is still used to process variables with different distribution characteristics, which affects the detection results. Although some scholars have combined the advantages of PCA and ICA to solve the problem of concurrent Gaussian and non-Gaussian processes, they have failed to mine the local structural characteristics of the data from the perspective of variable similarity, and failed to emphasize the different contribution information of each local model, resulting in poor detection performance. Summary of the invention

[0004] The present invention solves the problem that the existing industrial system fault detection methods fail to fully consider the distribution characteristics of data, the dynamic correlation of variables, local similarities, and fail to emphasize the different contribution information of each local model, which ultimately leads to poor detection performance. A fault detection method based on double-layer distributed and integrated Bayesian reasoning is provided.

[0005] The technical solution claimed in the present invention is as follows:

[0006] A fault detection method based on double-layer distributed and integrated Bayesian reasoning is proposed. State parameter monitoring sensors are installed at important nodes of the industrial system to collect and process the state parameters to obtain a sample set X=[x1,x2,...,x m ]∈R n×m, the sample set X contains both normal samples and faulty samples, including an offline modeling stage and a model testing stage; the offline modeling stage includes the following steps:

[0007] S11: Select normal samples as training sets, calculate the column mean and column variance of the samples in the training sets, and perform standardization on the samples in the training sets;

[0008] S12: First-layer block division method: Use Jarque-Bera test to perform first-layer block division on the samples in the training set to obtain Gaussian blocks and non-Gaussian blocks of the training set; the Gaussian blocks contain Gaussian variables, and the non-Gaussian blocks contain non-Gaussian variables;

[0009] S13: Augment the training set Gaussian block and non-Gaussian block obtained in S12 to obtain augmented Gaussian sub-blocks and augmented non-Gaussian sub-blocks of the training set, respectively.

[0010] S14: Second-layer block division method: Use K-means clustering to perform second-layer block division on the variables in the augmented Gaussian sub-block and augmented non-Gaussian sub-block of the training set in S13, and obtain several Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the training set;

[0011] S15: Apply the PCA method to each Gaussian cluster sub-block of the training set in S14 to obtain the control limit of each Gaussian cluster sub-block of the training set; at the same time, apply the ICA method to each non-Gaussian cluster sub-block of the training set in S14 to obtain the control limit of each non-Gaussian cluster sub-block of the training set;

[0012] The model testing phase includes the following steps:

[0013] S21: Select fault samples as a test set, and perform normalization processing on the samples in the test set; the normalization processing is to subtract the column mean of the training set samples in S11 from each column value of the sample and then divide it by the column variance of the training set samples in S11;

[0014] S22: Perform first-layer block division on the samples in the normalized test set using the first-layer block division method in the offline modeling stage to obtain Gaussian blocks and non-Gaussian blocks of the test set; augment the Gaussian blocks and non-Gaussian blocks of the test set to obtain augmented Gaussian sub-blocks and augmented non-Gaussian sub-blocks of the test set, respectively; divide the variables in the augmented Gaussian sub-blocks and augmented non-Gaussian sub-blocks of the test set using the second-layer block division method in the offline modeling stage to obtain a number of Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the test set corresponding to the Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the training set in S14, respectively;

[0015] S23: Apply the PCA method to each Gaussian cluster sub-block of the test set in S22 to obtain the monitoring statistics of each Gaussian cluster sub-block of the test set; apply the ICA method to each non-Gaussian cluster sub-block of the test set in S22 to obtain the monitoring statistics of each non-Gaussian cluster sub-block of the test set;

[0016] S24: According to the control limits of the Gaussian cluster sub-blocks of the training set in S15 of the offline modeling stage and the monitoring statistics of the Gaussian cluster sub-blocks of the test set obtained in S23, the fault probability index and sample detection accuracy of each Gaussian cluster sub-block of the test set in S22 are calculated; and according to the control limits of the non-Gaussian cluster sub-blocks of the training set in S15 of the offline modeling stage and the monitoring statistics of the non-Gaussian cluster sub-blocks of the test set obtained in S23, the fault probability index and sample detection accuracy of each non-Gaussian cluster sub-block of the test set in S22 are calculated; secondly, the normalized weights of the Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the corresponding test set in S22 are calculated by the sample detection accuracy, and then Bayesian reasoning is performed to assign the normalized weights to the fault probability indicators of the Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the corresponding test set, and then the global monitoring statistics are obtained;

[0017] S25: Determine whether the global monitoring statistic in S24 exceeds a threshold value. If so, it indicates that a fault has been detected and the offline modeling phase is completed. If not, return to S11 to continue offline modeling.

[0018] Preferably, the standardization process refers to subtracting the mean of each column element of each sample in the training set from the mean of the column in which it is located, and then dividing it by the standard deviation of the column in which it is located, so that each column element obeys a normal distribution with a mean of 0 and a standard deviation of 1.

[0019] Preferably, the augmentation in S13 and S22 is performed by adding variable time lag values ​​in Gaussian blocks and non-Gaussian blocks.

[0020] S15 The calculation of the control limit of each Gaussian clustering sub-block of the training set comprises the following steps:

[0021] S1511: Use PCA to cluster the i-th Gaussian sub-block X of the training set i G Decompose it, the formula is as follows:

[0022]

[0023] Where T∈R n×k , They represent the score matrix, loading matrix and residual matrix respectively; k represents the number of principal components, which is obtained by calculating the cumulative percentage of variance (CPV); n represents the number of samples; m i represents the number of variables in the i-th Gaussian cluster sub-block of the training set; (·) TRepresents the transpose operation;

[0024] S1512: Calculate T by formula i 2 Statistics and SPE i Statistics, the formula is as follows:

[0025]

[0026]

[0027] in, represents the i-th Gaussian clustering sub-block of the training set; represents the load matrix; Λ k It consists of the first k largest eigenvalues ​​of the covariance matrix of the principal components; I represents the identity matrix; (·) T Represents the transpose operation;

[0028] S1513: Calculate S1512 by formula Statistics and SPE i The control limit of the Gaussian clustering sub-block of the training set corresponding to the statistic is as follows:

[0029]

[0030]

[0031] in, k represents the degree of freedom; n represents the number of samples; F k,n-k,α represents the F distribution with the first degree of freedom being k and the second degree of freedom being nk; α represents the significance level; C α is the standard deviation corresponding to a Gaussian distribution.

[0032] S15 The calculation of the control limit of each non-Gaussian clustering sub-block of the training set comprises the following steps:

[0033] S1521: cluster the jth non-Gaussian sub-block X of the training set by ICA j NG Decompose it, the formula is as follows:

[0034]

[0035] Among them, A∈R n×d , E represents the mixing matrix, independent component matrix and residual matrix respectively; n represents the number of samples; m j represents the number of variables in the jth non-Gaussian clustering sub-block of the training set; d represents the number of independent components;

[0036] S1522: Using FastICA method Decomposed into W and The product of:

[0037]

[0038] in, represents the reconstructed source signal matrix; W represents the separation matrix; is the jth non-Gaussian clustering sub-block of the training set; perform eigenvalue decomposition on the sample covariance matrix C, the formula is as follows:

[0039] C=UΔU T (8)

[0040] Among them, U represents the eigenvector, and Δ represents the diagonal matrix composed of eigenvalues;

[0041] Then perform the whitening transformation, namely:

[0042]

[0043] Among them, the matrix represents the jth non-Gaussian clustering sub-block of the training set;

[0044] Introduce the orthogonal matrix B to calculate the matrix Q:

[0045] Z=QAS=BS (10)

[0046] Among them, A∈R n×d , Respectively represent the mixing matrix and the independent component matrix; n represents the number of samples; m j represents the number of variables in the jth non-Gaussian clustering sub-block of the training set; d represents the number of independent components;

[0047] therefore, It can be estimated by the following formula:

[0048]

[0049] S1523: Determine the number of independent components d by the cumulative variance percentage (CPV), and finally obtain the jth non-Gaussian clustering sub-block of the training set and SPE j Monitoring statistics:

[0050]

[0051]

[0052] in,

[0053] S1524: Using kernel density estimation and SPE j Control limits for monitoring statistics and SPE limj ; The kernel density estimation is the process of fitting a data set using a suitable smoothed probability density function from a set of random samples, and is widely used in the estimation of probability density functions; control limits and SPE limj Calculated from the respective probability density function estimates; the control limits and SPE limj It is the control limit of the jth non-Gaussian cluster sub-block of the training set.

[0054] The PCA method is applied to each Gaussian clustering sub-block of the S22 test set as described in S23, as follows: 2 The probability of failure of the i-th Gaussian clustering sub-block in the test set during the online detection phase is calculated using the following formula under statistics:

[0055]

[0056] in, N and F represent normal and fault status, Represents the i-th Gaussian cluster subblock of the test set The corresponding T 2 Statistics, Represents the i-th Gaussian cluster subblock of the test set The corresponding control limits; represents the i-th Gaussian clustering sub-block of the test set; α represents the set confidence level.

[0057] The normalized weights described in S24 are and express, and They represent the normalized weights of the Gaussian cluster sub-blocks of the test set and the non-Gaussian cluster sub-blocks of the test set respectively. The calculation formulas for the normalized weights of the i-th Gaussian cluster sub-block of the test set and the j-th non-Gaussian sub-block of the test set are as follows:

[0058]

[0059] in, represents the sample detection accuracy of the i-th Gaussian clustering sub-block in the test set, It represents the sample detection accuracy of the j-th non-Gaussian clustering sub-block of the test set, and the sample detection accuracy is the ratio of samples whose states are correctly detected to the total samples.

[0060] Global monitoring statistics BIC-C2 The calculation formula is as follows:

[0061]

[0062] Where h represents the number of Gaussian clustering sub-blocks in the test set; and They represent the normalized weights of the i-th Gaussian cluster sub-block of the test set and the j-th non-Gaussian cluster sub-block of the test set respectively; N and F represent the normal and fault states respectively; Represents the i-th Gaussian cluster sub-block of the test set; represents the jth non-Gaussian cluster sub-block of the test set; For T 2 Statistics, under the condition that a fault occurs, Probability of failure; Indicates that in T 2 Under statistics, for Probability of failure; Indicates that for I 2 Statistics, under the condition that a fault occurs, The probability of failure; Indicates that 2 Under statistics, Probability of failure;

[0063] The calculation formula of the global monitoring statistic BIC-SPE is as follows:

[0064]

[0065] in N and F represent normal and fault status respectively; Represents the i-th Gaussian cluster sub-block of the test set; represents the jth non-Gaussian cluster sub-block of the test set; and They represent respectively under the SPE statistic, and The normalized weight of It means that under the SPE statistic, when it is known that a fault has occurred, The probability of failure; It means that under the SPE statistic, when it is known that a fault has occurred, The probability of failure; Indicates that under the SPE statistic, for Probability of failure; Indicates that under the SPE statistic, The probability of failure.

[0066] Preferably, the fault detection method based on double-layer distributed and integrated Bayesian reasoning further includes an online monitoring stage; the online monitoring stage includes the following steps:

[0067] S31: Select online samples as monitoring sets, and perform normalization processing on the samples in the monitoring sets; the normalization processing is to subtract the column mean of the training set samples in S11 from each column value of the sample and then divide it by the column variance of the training set samples in S11;

[0068] S32: Perform first-layer block division on the samples in the normalized monitoring set using the first-layer block division method in the offline modeling stage to obtain Gaussian blocks and non-Gaussian blocks of the monitoring set; augment the Gaussian blocks and non-Gaussian blocks of the monitoring set to obtain augmented Gaussian sub-blocks and augmented non-Gaussian sub-blocks of the monitoring set, respectively; divide the variables in the augmented Gaussian sub-blocks and augmented non-Gaussian sub-blocks of the monitoring set using the second-layer block division method in the offline modeling stage to obtain a number of Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the monitoring set corresponding to the Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the monitoring set in S14, respectively;

[0069] S33: applying the PCA method to each Gaussian cluster sub-block of the monitoring set in S32 to obtain the monitoring statistics of each Gaussian cluster sub-block of the monitoring set; applying the ICA method to each non-Gaussian cluster sub-block of the monitoring set in S32 to obtain the monitoring statistics of each non-Gaussian cluster sub-block of the monitoring set;

[0070] S34: According to the control limits of the Gaussian cluster sub-blocks of the training set in S15 of the offline modeling stage and the monitoring statistics of the Gaussian cluster sub-blocks of the monitoring set obtained in S33, the fault probability index and sample detection accuracy of each Gaussian cluster sub-block of the monitoring set in S32 are calculated; and according to the control limits of the non-Gaussian cluster sub-blocks of the training set in S15 of the offline modeling stage and the monitoring statistics of the non-Gaussian cluster sub-blocks of the monitoring set obtained in S33, the fault probability index and sample detection accuracy of each non-Gaussian cluster sub-block of the monitoring set in S32 are calculated; secondly, the normalized weights of the Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the corresponding monitoring set in S32 are calculated by the sample detection accuracy, and then Bayesian reasoning is performed to assign the normalized weights to the fault probability indicators of the Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the corresponding monitoring set, thereby obtaining the global monitoring statistics;

[0071] S35: Determine whether the obtained global monitoring statistic exceeds the threshold. If so, it means that a fault is detected; if not, it means that the system is working normally, and return to S31 to continue online monitoring.

[0072] Beneficial effects:

[0073] The present invention provides a fault detection method based on double-layer distribution and integrated Bayesian reasoning, which is used for fault detection of industrial systems. A double-layer distribution strategy is adopted, normal samples are selected as training sets, and each sample in the training set is divided into a training set Gaussian block and a non-Gaussian block by using a Jarque-Bera test. Considering the autocorrelation or cross-correlation problem of each monitoring variable in the training set Gaussian block and the non-Gaussian block, the training set Gaussian block and the non-Gaussian block are augmented to solve the dynamic problem, and then a K-means clustering method is used to perform a second-layer block division on the variables in the augmented Gaussian sub-block and the augmented non-Gaussian sub-block of the training set to obtain the training set Gaussian cluster sub-block and the non-Gaussian cluster sub-block, so that the variables in each cluster sub-block have the same distribution characteristics (Gaussian or non-Gaussian) and local similarity, thereby realizing full mining of local information such as distribution characteristics, dynamics, similarity and the like of monitoring data, and solving the problem that the existing industrial system fault detection method fails to fully consider the distribution characteristics of data, the dynamic correlation of variables, and the local similarity. Due to the uncertainty and inconsistency of the monitoring results of the above-mentioned clustering sub-blocks, the present invention proposes to integrate Bayesian reasoning to highlight the contribution of different local monitoring models, that is, to distribute the sample detection accuracy of the above-mentioned clustering sub-blocks as weights to the fault probability indicators of each clustering sub-block, and obtain a comprehensive global monitoring statistic, which ultimately effectively improves the performance of fault detection and solves the problem that the existing industrial system fault detection method cannot emphasize the different contribution information of each local model, which ultimately leads to poor detection performance.

[0074] The present invention first selects normal samples as training sets, and performs first-layer block division and second-layer block division on the samples in the training sets to obtain several Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the training sets, and calculates the control limits of each of the above cluster sub-blocks. However, the control limits of each cluster sub-block obtained from the samples in the training sets cannot evaluate the fault detection performance of the method proposed by the present invention. Therefore, faulty samples known to be faulty are selected to evaluate the fault detection performance of the method of the present invention. For this purpose, faulty samples are selected to form a test set, and the method of S22-S25 of the present invention is used to predict whether the sample is a faulty sample. By comparing the predicted results with the actual results, the fault detection rate of the entire test set is obtained, and then the fault detection performance of the proposed method can be evaluated according to the fault detection rate. For unknown samples to be detected in practice, they can be directly detected online using the steps described in S31-S35 to determine whether the unknown sample is a faulty sample. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 It is a flow chart of the scheme of the present invention. DETAILED DESCRIPTION

[0076] The present invention is further described in detail below in conjunction with the accompanying drawings. It should be understood that the present invention is not limited to the following technical solutions.

[0077] A fault detection method based on double-layer distributed and integrated Bayesian reasoning is proposed. State parameter monitoring sensors are installed at important nodes of the industrial system to collect and process the state parameters to obtain a sample set X=[x1,x2,...,x m ]∈R n×m , the sample set X contains both normal samples and fault samples, and each sample contains m variables, such as Figure 1 As shown, it includes an offline modeling stage and a model testing stage; the offline modeling stage includes the following steps:

[0078] S11: Select normal samples as training set X train , calculate the training set X train The column mean and column variance of the sample (training set X train The mean and variance of the variables with the same index in all samples in the training set X train The samples in the training set X are standardized. The standardized processing refers to train Each column element of the training set X is subtracted from the mean of the column in which it is located (column mean), and then divided by the standard deviation of the column (column variance), so that each column element obeys a normal distribution with a mean of 0 and a standard deviation of 1; train is a matrix. The number of samples in the training set is not fixed and can be any number. Therefore, the number of rows in the training set is not fixed, but the number of columns is fixed to m.

[0079] S12: First-level block division method: Use Jarque-Bera (JB) test to divide the samples in the training set into first-level blocks and obtain X train =[X Gaussian ,X Non-Gaussian ];X Gaussian represents the Gaussian block of the training set, which is a matrix whose columns consist of the values ​​of the Gaussian variables; X Non-Gaussian represents the non-Gaussian block of the training set, which is also a matrix, and the columns of the matrix are composed of the values ​​of non-Gaussian variables; assuming that the number of variables in a sample is 50 (the index of the variable in the sample is 1, 2, 3..., 50), the first layer of block division is completed by JB test and is divided into Gaussian blocks and non-Gaussian blocks. In addition to the variables, the Gaussian blocks and non-Gaussian blocks also contain the index of the variables. In actual industrial production, state parameter monitoring sensors are installed at important nodes of the industrial system. After the state parameters are collected and processed, the type of the variable corresponding to each index in each sample is determined (that is, in one of the samples, the variable corresponding to index 1 is a Gaussian variable, then in other samples, the corresponding index 1 should also be a Gaussian variable. In one sample, the corresponding index 2 is a non-Gaussian variable, then in other samples, the corresponding index 2 is also a non-Gaussian variable, and so on);

[0080] S13: Considering the autocorrelation or cross-correlation problems of variables in the training set Gaussian block and the training set non-Gaussian block; in the training set Gaussian block of S12, the augmentation is performed by adding the time lag value of the variable, that is, the training set Gaussian block is augmented by adding the observation results of the past moment to obtain the augmented Gaussian sub-block of the training set; similarly, in the training set non-Gaussian block of S12, the augmentation is performed by adding the time lag value of the variable, that is, the non-Gaussian block is augmented by adding the observation results of the past moment to obtain the augmented non-Gaussian sub-block of the training set.

[0081] S14: Second-layer block division method: Use K-means clustering to perform second-layer block division on the variables in the augmented Gaussian sub-block and augmented non-Gaussian sub-block of the training set in S13, and obtain h training set Gaussian clustering sub-blocks. and b training set non-Gaussian clustering sub-blocks The training set Gaussian cluster sub-block and non-Gaussian cluster sub-block both contain several variables and indexes corresponding to the variables. Assuming that a sample contains 18 variables, 12 of the variables are divided into a Gaussian block in the training set through the JB test, and the remaining 6 variables are divided into a non-Gaussian block in the training set. The training set Gaussian block and the training set non-Gaussian block are divided into the second layer of blocks through K-means clustering. Assuming that the number of clusters of the training set Gaussian block is 2, the training set Gaussian block is divided into two training set Gaussian cluster sub-blocks. Assuming that each training set Gaussian cluster sub-block contains 6 variables (it can also be the first training set Gaussian cluster The first sub-block contains 4 variables, and the second training set Gaussian cluster sub-block contains 8 variables), and the training set Gaussian cluster sub-block indexes are 1, 3, 5, 7, 8, 10 and 2, 4, 6, 9, 11, 12 respectively; assuming that the number of clusters of the obtained training set non-Gaussian block is 2, the training set non-Gaussian block is divided into two training set non-Gaussian cluster sub-blocks. Assuming that each training set non-Gaussian cluster sub-block contains 3 variables (it can also be that the first training set non-Gaussian cluster sub-block contains 2 variables and the second training set non-Gaussian cluster sub-block contains 4 variables), the training set non-Gaussian cluster sub-block indexes are 13, 15, 18 and 14, 16, 17 respectively.

[0082] S15: Gaussian clustering sub-blocks of the h training sets in S14 Using the PCA (Principal Component Analysis) method, we get the Gaussian clustering sub-blocks of the h training sets. Control limits and Control limits; where The i-th Gaussian clustering subblock in the training set corresponds to S152 Control limits of statistics; SPE limi The i-th Gaussian clustering subblock in the training set corresponds to SPEi The control limits of statistics mainly include the following processes:

[0083] S1511: cluster the i-th Gaussian sub-block of the training set by PCA Decompose it, the formula is as follows:

[0084]

[0085] Where T∈R n×k , They represent the score matrix, loading matrix and residual matrix respectively; k represents the number of principal components, which is obtained by calculating the cumulative percentage of variance (CPV); n represents the number of samples; m i represents the number of variables in the i-th Gaussian cluster sub-block of the training set; (·) T Represents the transpose operation;

[0086] S1512: Calculate the T of the i-th Gaussian clustering sub-block in the training set by the formula i 2 Statistics and SPE i Statistics, the formula is as follows:

[0087]

[0088]

[0089] in, represents the i-th Gaussian clustering sub-block of the training set; represents the load matrix; Λ k It consists of the first k largest eigenvalues ​​of the covariance matrix of the principal components; I represents the identity matrix; (·) T Represents the transpose operation;

[0090] S1513: Calculate the control limit of the i-th Gaussian clustering sub-block of the training set by using the formula, the formula is as follows:

[0091]

[0092]

[0093] in, k represents the degree of freedom; n represents the number of samples; F k,n-k,α represents the F distribution with the first degree of freedom being k and the second degree of freedom being nk; α represents the significance level; C α is the standard deviation corresponding to a Gaussian distribution.

[0094] At the same time, for the b training set non-Gaussian clustering sub-blocks in S14 Using the ICA (Independent Component Analysis) method, we get the b training set non-Gaussian clustering sub-blocks Control limits and Control limits; where The jth non-Gaussian clustering sub-block in the training set corresponds to S1523 Control limits for monitoring statistics; SPE limj The jth non-Gaussian clustering sub-block in the training set corresponds to SPE in S1523 j Monitor the control limits of statistics; mainly includes the following processes:

[0095] S1521: Cluster the jth non-Gaussian sub-block of the training set by ICA Decompose it, the formula is as follows:

[0096]

[0097] Among them, A∈R n×d , E represents the mixing matrix, independent component matrix and residual matrix respectively; n represents the number of samples; m j represents the number of variables in the jth non-Gaussian clustering sub-block of the training set; d represents the number of independent components;

[0098] S1522: Using FastICA method Decomposed into W and The product of:

[0099]

[0100] in, represents the reconstructed source signal matrix; W represents the separation matrix; is the jth non-Gaussian clustering sub-block of the training set;

[0101] Perform eigenvalue decomposition on the sample covariance matrix C, the formula is as follows:

[0102] C=UΔU T (8)

[0103] Among them, U represents the eigenvector, and Δ represents the diagonal matrix composed of eigenvalues;

[0104] Then perform the whitening transformation, namely:

[0105]

[0106] Among them, the matrix represents the jth non-Gaussian clustering sub-block of the training set;

[0107] Introduce the orthogonal matrix B to calculate the matrix Q:

[0108] Z=QAS=BS (10)

[0109] Among them, A∈R n×d , Respectively represent the mixing matrix and the independent component matrix; n represents the number of samples; m j represents the number of variables in the jth non-Gaussian clustering sub-block of the training set; d represents the number of independent components;

[0110] therefore, It can be estimated by the following formula:

[0111]

[0112] S1523: Determine the number of independent components d by the cumulative variance percentage (CPV), and finally obtain the jth non-Gaussian clustering sub-block of the training set and SPE j Monitoring statistics:

[0113]

[0114]

[0115] in,

[0116] S1524: Using kernel density estimation and SPE j Control limits for monitoring statistics and SPE limj ; The kernel density estimation is the process of fitting a data set using a suitable smoothed probability density function from a set of random samples, and is widely used in the estimation of probability density functions; control limits and SPE limj Calculated from the respective probability density function estimates.

[0117] The model testing phase includes the following steps:

[0118] S21: Select fault samples as a test set, and perform normalization processing on the samples in the test set; the normalization processing is to subtract the column mean of the training set samples in S11 from each column value of the sample and then divide it by the column variance of the training set samples in S11;

[0119] S22: For the sample x in the test set, the first-layer block division method in the offline modeling stage is used to perform the first-layer block division on the normalized sample in the test set to obtain the test set Gaussian block and non-Gaussian block; the test set Gaussian block and non-Gaussian block are augmented to obtain the test set augmented Gaussian sub-block and augmented non-Gaussian sub-block respectively; the variables in the augmented Gaussian sub-block and augmented non-Gaussian sub-block of the test set are divided by the second-layer block division method in the offline modeling stage to obtain h test set Gaussian cluster sub-blocks corresponding to the training set Gaussian cluster sub-block and non-Gaussian cluster sub-block in S14 respectively. and b test set non-Gaussian clustered sub-blocks The test set Gaussian cluster sub-block and non-Gaussian cluster sub-block contain several variables and indexes corresponding to the variables, based on the assumption of S14: the training set Gaussian block is divided into two training set Gaussian cluster sub-blocks, each training set Gaussian cluster sub-block contains 6 variables, and the training set Gaussian cluster sub-block indexes are 1, 3, 5, 7, 8, 10 and 2, 4, 6, 9, 11, 12 respectively; the training set non-Gaussian block is divided into two training set non-Gaussian cluster sub-blocks, each training set non-Gaussian cluster sub-block contains 3 variables, and the training set non-Gaussian cluster sub-block indexes are They are 13, 15, 18 and 14, 16, 17 respectively; correspondingly: the test set Gaussian block is divided into two test set Gaussian cluster sub-blocks, each test set Gaussian cluster sub-block contains 6 variables, and the test set Gaussian cluster sub-block indexes are 1, 3, 5, 7, 8, 10 and 2, 4, 6, 9, 11, 12 respectively; the test set non-Gaussian block is divided into two test set non-Gaussian cluster sub-blocks, each test set non-Gaussian cluster sub-block contains 3 variables, and the test set non-Gaussian cluster sub-block indexes are 13, 15, 18 and 14, 16, 17 respectively.

[0120] S23: Gaussian clustering of the h test set sub-blocks in S22 Using the PCA model, we get the Gaussian clustering sub-blocks of the h test sets. and Monitoring statistics; among them, is the T of the i-th Gaussian cluster subblock of the test set 2 Monitoring statistics; SPE i is the SPE monitoring statistic of the i-th Gaussian cluster sub-block of the test set;

[0121] At the same time, for the non-Gaussian clustering sub-blocks of the test set Using the ICA model, we get the b non-Gaussian clustering sub-blocks of the test set. and Monitoring statistics, where is the I of the jth non-Gaussian clustered sub-block of the test set 2 Monitoring Statistics and SPE j is the SPE monitoring statistic of the jth non-Gaussian clustered sub-block of the test set.

[0122] S24: Using T of the i-th Gaussian clustering sub-block of the test set 2 Monitoring Statistics and the control limits of the ith Gaussian clustering subblock of the corresponding S1513 training set To calculate the class conditional probability and Then, the probability of failure of the i-th Gaussian clustering sub-block in the test set during the model testing phase is calculated by the following formula:

[0123]

[0124] The failure probability index of the i-th Gaussian cluster sub-block of the test set is obtained as:

[0125] in, N and F indicate normal and fault status; Represents the i-th Gaussian cluster sub-block of the test set; For T 2 Statistics, under the condition that a fault occurs, Probability of failure; For T 2 Statistics, under the condition that no fault occurs, The probability that no failure occurs; For T 2 Statistics, The probability of failure; α represents the confidence level.

[0126] Similarly, using the I of the jth non-Gaussian clustering sub-block of the test set 2 Monitoring Statistics and the control limits of the jth non-Gaussian clustered sub-block of the training set in S1513 To calculate the class conditional probabilities: and Then, the probability of failure of the jth non-Gaussian clustering sub-block in the test set is calculated by the following formula:

[0127]

[0128] The failure probability index of the jth non-Gaussian clustering sub-block of the test set is obtained as:

[0129] in, N and F indicate normal and fault status; Indicates that for I 2 Statistics, under the condition that a fault occurs, Probability of failure; Indicates that for I 2 Statistics, under the condition that no fault occurs, The probability that no failure occurs; Indicates that for I 2 Statistics, for Probability of failure; represents the jth non-Gaussian cluster sub-block of the test set; α represents the confidence level;

[0130] Then calculate the sample detection accuracy of each Gaussian cluster sub-block and non-Gaussian cluster sub-block of the corresponding test set in S22 respectively, and calculate the normalized weights of the Gaussian cluster sub-block and non-Gaussian cluster sub-block of the corresponding test set in S22 by the detection accuracy and The specific formula is as follows:

[0131]

[0132] in, represents the sample detection accuracy of the i-th Gaussian clustering sub-block in the test set, represents the sample detection accuracy of the jth non-Gaussian clustering sub-block of the test set; the sample detection accuracy is the ratio of samples whose states are correctly detected to the total samples;

[0133] The normalized weights are assigned to the failure probability indicators of each cluster sub-block in the test set, and the sum is obtained to obtain the integrated global monitoring statistic BIC-C 2 :

[0134]

[0135] Similarly, the following formula can be used to obtain the global monitoring statistic BIC-SPE, and the thresholds of both statistics are 1-α:

[0136]

[0137] in N and F represent normal and fault status respectively; Represents the i-th Gaussian cluster sub-block of the test set; represents the jth non-Gaussian cluster sub-block of the test set; and They represent respectively under the SPE statistic, and The normalized weight of It means that under the SPE statistic, when it is known that a fault has occurred, Probability of failure; It means that under the SPE statistic, when it is known that a fault has occurred, Probability of failure; Indicates that under the SPE statistic, for Probability of failure; Indicates that under the SPE statistic, The probability of failure.

[0138] S25: Determine whether the global monitoring statistic in S24 exceeds the threshold. If so, it means that a fault is detected and the offline modeling stage is qualified; if not, it means that the fault detection is wrong, and return to S11 to continue offline modeling.

[0139] After the offline modeling is qualified, the online samples are tested using the steps in the model testing phase, including the following steps:

[0140] S31: Select online samples as monitoring sets, and perform normalization processing on the samples in the monitoring sets; the normalization processing is to subtract the column mean of the training set samples in S11 from each column value of the sample and then divide it by the column variance of the training set samples in S11;

[0141] S32: Perform first-layer block division on the samples in the normalized monitoring set using the first-layer block division method in the offline modeling stage to obtain Gaussian blocks and non-Gaussian blocks of the monitoring set; augment the Gaussian blocks and non-Gaussian blocks of the monitoring set to obtain augmented Gaussian sub-blocks and augmented non-Gaussian sub-blocks of the monitoring set, respectively; divide the variables in the augmented Gaussian sub-blocks and augmented non-Gaussian sub-blocks of the monitoring set using the second-layer block division method in the offline modeling stage to obtain a number of Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the monitoring set corresponding to the Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the monitoring set in S14, respectively;

[0142] S33: applying the PCA method to each Gaussian cluster sub-block of the monitoring set in S32 to obtain the monitoring statistics of each Gaussian cluster sub-block of the monitoring set; applying the ICA method to each non-Gaussian cluster sub-block of the monitoring set in S32 to obtain the monitoring statistics of each non-Gaussian cluster sub-block of the monitoring set;

[0143] S34: According to the control limits of the Gaussian clustering sub-blocks of the training set in S15 of the offline modeling stage and the monitoring statistics of the Gaussian clustering sub-blocks of the monitoring set obtained in S33 (the Gaussian clustering sub-blocks of the training set and the Gaussian clustering sub-blocks of the monitoring set correspond to each other, and both are taken as the i-th), calculate the fault probability index and sample detection accuracy of each Gaussian clustering sub-block of the monitoring set in S32 (the control limits of the i-th Gaussian clustering sub-block of the training set and the monitoring statistics of the i-th Gaussian clustering sub-block of the monitoring set are used to calculate the fault probability index and sample detection accuracy of the i-th Gaussian clustering sub-block of the monitoring set); and according to the control limits of the non-Gaussian clustering sub-blocks of the training set in S15 of the offline modeling stage and the monitoring statistics of the non-Gaussian clustering sub-blocks of the monitoring set obtained in S33 (the control limits of the i-th Gaussian clustering sub-block of the training set and the monitoring statistics of the i-th Gaussian clustering sub-block of the monitoring set are used to calculate the fault probability index and sample detection accuracy of the i-th Gaussian clustering sub-block of the monitoring set); The non-Gaussian clustering sub-blocks in the training set and the non-Gaussian clustering sub-blocks in the monitoring set correspond to each other, and both are taken as the jth), and the fault probability index and sample detection accuracy of each non-Gaussian clustering sub-block in the monitoring set in S32 are calculated (the control limit of the jth non-Gaussian clustering sub-block in the training set and the monitoring statistic of the jth non-Gaussian clustering sub-block in the monitoring set are used to calculate the fault probability index and sample detection accuracy of the jth non-Gaussian clustering sub-block in the monitoring set); secondly, the normalized weights of the corresponding Gaussian clustering sub-blocks and non-Gaussian clustering sub-blocks in the monitoring set in S32 are calculated through the sample detection accuracy, and then Bayesian reasoning is performed to assign the normalized weights to the fault probability indexes of the corresponding Gaussian clustering sub-blocks and non-Gaussian clustering sub-blocks in the monitoring set, and then the global monitoring statistics are obtained;

[0144] S35: Determine whether the obtained global monitoring statistic exceeds the threshold. If so, it means that a fault is detected; if not, it means that the system is working normally, and return to S31 to continue online monitoring.

[0145] The calculation formulas in the above S32-S34 are as described in the above S22-S24.

Claims

1. A fault detection method based on double-layer distributed and integrated Bayesian reasoning, in which state parameter monitoring sensors are installed at important nodes of the industrial system, the state parameters are collected and processed, and a sample set X=[x1,x2,...,x m ]∈R n×m , the sample set X contains both normal samples and faulty samples, characterized in that, It includes an offline modeling stage and a model testing stage; the offline modeling stage includes the following steps: S11: Select normal samples as training sets, calculate the column mean and column variance of the samples in the training sets, and perform standardization on the samples in the training sets; S12: First-layer block division method: Use Jarque-Bera test to perform first-layer block division on the samples in the training set to obtain Gaussian blocks and non-Gaussian blocks of the training set; the Gaussian blocks contain Gaussian variables, and the non-Gaussian blocks contain non-Gaussian variables; S13: Augment the training set Gaussian block and the non-Gaussian block obtained in S12 to obtain an augmented Gaussian sub-block and an augmented non-Gaussian sub-block of the training set, respectively; S14: Second-layer block division method: Use K-means clustering to perform second-layer block division on the variables in the augmented Gaussian sub-block and augmented non-Gaussian sub-block of the training set in S13, and obtain several Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the training set; S15: Apply the PCA method to each Gaussian cluster sub-block of the training set in S14 to obtain the control limit of each Gaussian cluster sub-block of the training set; at the same time, apply the ICA method to each non-Gaussian cluster sub-block of the training set in S14 to obtain the control limit of each non-Gaussian cluster sub-block of the training set; The model testing phase includes the following steps: S21: Select fault samples as a test set, and perform normalization processing on the samples in the test set; the normalization processing is to subtract the column mean of the training set samples in S11 from each column value of the sample and then divide it by the column variance of the training set samples in S11; S22: Perform first-layer block division on the samples in the normalized test set using the first-layer block division method in the offline modeling stage to obtain Gaussian blocks and non-Gaussian blocks of the test set; augment the Gaussian blocks and non-Gaussian blocks of the test set to obtain augmented Gaussian sub-blocks and augmented non-Gaussian sub-blocks of the test set, respectively; divide the variables in the augmented Gaussian sub-blocks and augmented non-Gaussian sub-blocks of the test set using the second-layer block division method in the offline modeling stage to obtain a number of Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the test set corresponding to the Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the training set in S14, respectively; S23: Apply the PCA method to each Gaussian cluster sub-block of the test set in S22 to obtain the monitoring statistics of each Gaussian cluster sub-block of the test set; apply the ICA method to each non-Gaussian cluster sub-block of the test set in S22 to obtain the monitoring statistics of each non-Gaussian cluster sub-block of the test set; S24: According to the control limits of the Gaussian cluster sub-blocks of the training set in S15 of the offline modeling stage and the monitoring statistics of the Gaussian cluster sub-blocks of the test set obtained in S23, the fault probability index and sample detection accuracy of each Gaussian cluster sub-block of the test set in S22 are calculated; and according to the control limits of the non-Gaussian cluster sub-blocks of the training set in S15 of the offline modeling stage and the monitoring statistics of the non-Gaussian cluster sub-blocks of the test set obtained in S23, the fault probability index and sample detection accuracy of each non-Gaussian cluster sub-block of the test set in S22 are calculated; secondly, the normalized weights of the Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the test set in S22 are calculated by the sample detection accuracy, and then Bayesian reasoning is performed to assign the normalized weights to the fault probability indicators of the Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the corresponding test set, thereby obtaining the global monitoring statistics; The normalized weights are and express, and They represent the normalized weights of the Gaussian cluster sub-blocks of the test set and the non-Gaussian cluster sub-blocks of the test set respectively. The calculation formulas for the normalized weights of the i-th Gaussian cluster sub-block of the test set and the j-th non-Gaussian sub-block of the test set are as follows: in, represents the sample detection accuracy of the i-th Gaussian clustering sub-block in the test set, represents the sample detection accuracy of the jth non-Gaussian clustering sub-block of the test set, where the sample detection accuracy is the ratio of samples whose states are correctly detected to the total samples; Global monitoring statistics BIC-C 2 The calculation formula is as follows: Where h represents the number of Gaussian clustering sub-blocks in the test set; and They represent the normalized weights of the i-th Gaussian cluster sub-block of the test set and the j-th non-Gaussian cluster sub-block of the test set respectively; N and F represent the normal and fault states respectively; Represents the i-th Gaussian cluster sub-block of the test set; represents the jth non-Gaussian cluster sub-block of the test set; For T 2 Statistics, under the condition that a fault occurs, The probability of failure; Indicates that in T 2 Under statistics, for The probability of failure; Indicates that for I 2 Statistics, under the condition that a fault occurs, The probability of failure; Indicates that 2 Under statistics, The probability of failure; The calculation formula of the global monitoring statistic BIC-SPE is as follows: in N and F represent normal and fault status respectively; Represents the i-th Gaussian cluster sub-block of the test set; represents the jth non-Gaussian cluster sub-block of the test set; and They represent respectively under the SPE statistic, and The normalized weight of It means that under the SPE statistic, when it is known that a fault has occurred, The probability of failure; It means that under the SPE statistic, when it is known that a fault has occurred, The probability of failure; Indicates that under the SPE statistic, for The probability of failure; Indicates that under the SPE statistic, The probability of failure; S25: Determine whether the global monitoring statistic in S24 exceeds a threshold value. If so, it indicates that a fault has been detected and the offline modeling phase is completed. If not, return to S11 to continue offline modeling.

2. The fault detection method based on double-layer distributed and integrated Bayesian reasoning according to claim 1 is characterized in that: The standardization process refers to subtracting the mean of each column element of each sample in the training set from the mean of the column in which it is located, and then dividing it by the standard deviation of the column, so that each column element obeys a normal distribution with a mean of 0 and a standard deviation of 1.

3. The fault detection method based on double-layer distributed and integrated Bayesian reasoning according to claim 1 is characterized in that: The augmentation described in S13 and S22 is performed by adding variable time lag values ​​in the Gaussian block and the non-Gaussian block.

4. The fault detection method based on double-layer distributed and integrated Bayesian reasoning according to any one of claims 1 to 3, characterized in that: S15 The calculation of the control limit of each Gaussian clustering sub-block of the training set comprises the following steps: S1511: Use PCA to cluster the i-th Gaussian sub-block of the training set Decompose it, the formula is as follows: Where T∈R n×k , They represent the score matrix, loading matrix and residual matrix respectively; k represents the number of principal components, which is obtained by calculating the cumulative variance percentage CPV; n represents the number of samples; m i represents the number of variables in the i-th Gaussian cluster sub-block of the training set; (·) T Represents the transpose operation; S1512: Calculate T using the formula i 2 Statistics and SPE i Statistics, the formula is as follows: in, Represents the i-th Gaussian clustering sub-block of the training set; represents the load matrix; Λ k It consists of the first k largest eigenvalues ​​of the covariance matrix of the principal components; I represents the identity matrix; (·) T Represents the transpose operation; S1513: Calculate T in S1512 by formula i 2 Statistics and SPE i The control limit of the Gaussian clustering sub-block of the training set corresponding to the statistic is as follows: in, , N = 1, 2, 3; k represents the degree of freedom; n represents the number of samples; F k,n-k,α represents the F distribution with the first degree of freedom being k and the second degree of freedom being nk; α represents the significance level; C α is the standard deviation corresponding to a Gaussian distribution.

5. The fault detection method based on double-layer distributed and integrated Bayesian reasoning according to claim 4 is characterized in that: The calculation of the control limits of each non-Gaussian clustering sub-block of the training set in S15 includes the following steps: S1521: Cluster the jth non-Gaussian sub-block of the training set by ICA Decompose it, the formula is as follows: Among them, A∈R n×d , E represents the mixing matrix, independent component matrix and residual matrix respectively; n represents the number of samples; m j represents the number of variables in the jth non-Gaussian clustering sub-block of the training set; d represents the number of independent components; S1522: Using FastICA method Decomposed into W and The product of: in, represents the reconstructed source signal matrix; W represents the separation matrix; is the jth non-Gaussian clustering sub-block of the training set; perform eigenvalue decomposition on the sample covariance matrix C, the formula is as follows: C=UΔU T (8) Among them, U represents the eigenvector, and Δ represents the diagonal matrix composed of eigenvalues; Then perform the whitening transformation, namely: Among them, the matrix represents the jth non-Gaussian clustering sub-block of the training set; Introduce the orthogonal matrix B to calculate the matrix Q: Z=QAS=BS (10) Among them, A∈R n×d , Respectively represent the mixing matrix and the independent component matrix; n represents the number of samples; m j represents the number of variables in the jth non-Gaussian clustering sub-block of the training set; d represents the number of independent components; therefore, It is estimated by the following formula: S1523: Determine the number of independent components d by the cumulative variance percentage CPV, and finally obtain the jth non-Gaussian clustering sub-block of the training set and SPE j Monitoring statistics: in, S1524: Using kernel density estimation and SPE j Control limits for monitoring statistics and SPE limj ; The kernel density estimation is the process of fitting a data set using a suitable smoothed probability density function from a set of random samples, and is widely used in the estimation of probability density functions; control limits and SPE limj Calculated from the respective probability density function estimates; the control limits and SPE limj It is the control limit of the jth non-Gaussian cluster sub-block of the training set.

6. The fault detection method based on double-layer distributed and integrated Bayesian reasoning according to claim 5 is characterized in that: The PCA method is applied to each Gaussian clustering sub-block of the S22 test set as described in S23, as follows: 2 The probability of failure of the i-th Gaussian clustering sub-block in the test set during the online detection phase is calculated using the following formula under statistics: in, N and F represent normal and fault status, T i 2 Represents the i-th Gaussian cluster subblock of the test set The corresponding T 2 Statistics, Represents the i-th Gaussian cluster subblock of the test set The corresponding control limits; represents the i-th Gaussian clustering sub-block of the test set; α represents the set confidence level.

7. The fault detection method based on double-layer distributed and integrated Bayesian reasoning according to claim 1 is characterized in that: It also includes an online monitoring stage; the online monitoring stage includes the following steps: S31: Select online samples as monitoring sets, and perform normalization processing on the samples in the monitoring sets; the normalization processing is to subtract the column mean of the training set samples in S11 from each column value of the sample and then divide it by the column variance of the training set samples in S11; S32: Perform first-layer block division on the samples in the normalized monitoring set using the first-layer block division method in the offline modeling stage to obtain Gaussian blocks and non-Gaussian blocks of the monitoring set; augment the Gaussian blocks and non-Gaussian blocks of the monitoring set to obtain augmented Gaussian sub-blocks and augmented non-Gaussian sub-blocks of the monitoring set, respectively; divide the variables in the augmented Gaussian sub-blocks and augmented non-Gaussian sub-blocks of the monitoring set using the second-layer block division method in the offline modeling stage to obtain a number of Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the monitoring set corresponding to the Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the monitoring set in S14, respectively; S33: applying the PCA method to each Gaussian cluster sub-block of the monitoring set in S32 to obtain the monitoring statistics of each Gaussian cluster sub-block of the monitoring set; applying the ICA method to each non-Gaussian cluster sub-block of the monitoring set in S32 to obtain the monitoring statistics of each non-Gaussian cluster sub-block of the monitoring set; S34: According to the control limits of the Gaussian cluster sub-blocks of the training set in S15 of the offline modeling stage and the monitoring statistics of the Gaussian cluster sub-blocks of the monitoring set obtained in S33, the fault probability index and sample detection accuracy of each Gaussian cluster sub-block of the monitoring set in S32 are calculated; and according to the control limits of the non-Gaussian cluster sub-blocks of the training set in S15 of the offline modeling stage and the monitoring statistics of the non-Gaussian cluster sub-blocks of the monitoring set obtained in S33, the fault probability index and sample detection accuracy of each non-Gaussian cluster sub-block of the monitoring set in S32 are calculated; secondly, the normalized weights of the Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the corresponding monitoring set in S32 are calculated by the sample detection accuracy, and then Bayesian reasoning is performed to assign the normalized weights to the fault probability indicators of the Gaussian cluster sub-blocks and non-Gaussian cluster sub-blocks of the corresponding monitoring set, thereby obtaining the global monitoring statistics; S35: Determine whether the obtained global monitoring statistic exceeds the threshold. If so, it means that a fault is detected; if not, it means that the system is working normally, and return to S31 to continue online monitoring.

Citation Information

Patent Citations

  • Batch process fault monitoring method based on multi-stage ICA-SVDD

    CN107272655A

  • Multi-condition fault monitoring method based on distributed PCA

    CN108762228A

Cited By

  • Method and system for monitoring abnormity of multiple performance indexes in strip steel hot rolling process

    CN122346780A