Industrial process fault detection method and device based on enhanced data spatiotemporal neighborhood features

The TS-SSAE-CVA model, enhanced with spatiotemporal neighborhood features, solves the problem that traditional methods do not consider the dynamic correlation of sample data and neighborhood interaction, thus achieving more accurate fault detection and condition monitoring.

CN118092385BActive Publication Date: 2024-11-05UNIV OF SCI & TECH BEIJING
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410177636.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-02-08
Publication Date
2024-11-05
Estimated Expiration
2044-02-08

AI Technical Summary

Technical Problem

Traditional industrial process monitoring methods fail to effectively consider the dynamic correlations between sample data and the interactions between spatial and temporal neighborhoods, resulting in limited monitoring effectiveness.

Method used

A method for enhancing spatiotemporal neighborhood features is adopted. The spatiotemporal-stacked sparse autoencoder-normalized variable analysis (TS-SSAE-CVA) model is used to enhance features by acquiring hot strip rolling process data. Combined with k-nearest neighbor algorithm and Bayesian inference, monitoring statistics and control limits of state subspace and residual subspace are constructed.

Benefits of technology

It improves the accuracy and effectiveness of fault detection, better reflects the actual situation of the production process, extracts nonlinear dynamic features, and intuitively obtains information on process state changes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118092385B_ABST
    Figure CN118092385B_ABST
Patent Text Reader

Abstract

The present invention discloses an industrial process fault detection method and device with enhanced spatiotemporal neighborhood features of data, and relates to the technical field of industrial process monitoring. The method comprises: obtaining process data of hot strip rolling to be detected, and performing feature enhancement on the process data; inputting the feature-enhanced process data into a constructed spatiotemporal-stacked sparse autoencoder-canonical variable analysis TS-SSAE-CVA model; wherein the spatiotemporal-stacked sparse autoencoder-canonical variable analysis TS-SSAE-CVA model comprises a spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model and a canonical variable analysis CVA sub-model; according to the feature-enhanced process data, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model and the canonical variable analysis CVA sub-model, obtaining the hot strip rolling process fault detection result. The present invention can solve the problem that the traditional process monitoring method ignores the spatiotemporal neighborhood of data, does not consider the dynamic nature of data, and leads to insufficient fault detection efficiency.
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 process monitoring, and in particular to an industrial process fault detection method and device with enhanced data spatiotemporal neighborhood features. Background Art

[0002] The hot strip rolling process is a steel production process with complex mechanism, large scale, high efficiency and multiple processes. It is a typical complex process industry. The production line is mainly composed of heating furnace, rough rolling unit, hot output roller and flying shear, finishing rolling unit, laminar cooling and coiling unit, etc. Monitoring and fault diagnosis of the hot strip rolling process helps operators to understand the operation status of the production process in real time, and timely monitor the abnormal state that occurs during the operation, so as to locate the fault and take corresponding improvement measures, eliminate the abnormal state, improve product quality and production efficiency, and reduce safety hazards. With the rapid development of automatic control, as well as the application and optimization and upgrading of fieldbus systems, it is possible to obtain a large amount of data containing product information, processing information, and operation status information from the production process. Therefore, the process monitoring method based on data-driven has become a hot spot for hot strip rolling process monitoring.

[0003] Sensors are widely distributed in the hot strip rolling process. The high-dimensional data collected by different sensors can reflect the operating status of the process. Due to the complex nonlinear relationship between industrial process variables, there is also a time series correlation between samples at different times. Considering that the time neighborhood or spatial neighborhood of a sample at a specific moment can interact with it, the traditional fault diagnosis method based on multivariate statistical methods cannot achieve ideal monitoring results. Therefore, the research on the monitoring of the hot strip rolling process with enhanced spatial and temporal neighborhood characteristics of data has practical engineering significance.

[0004] Traditional process monitoring methods are divided into three categories: analytical model-based methods, data-driven methods, and knowledge-based methods. The production process of modern process industry is complex, involving complex physical and chemical processes, fuzzy and changeable boundary conditions, numerous variables, and an increasingly obvious trend of large-scale development. On the one hand, it is difficult to obtain an accurate mechanism model, and on the other hand, it is difficult to organize and establish a systematic expert knowledge base. Therefore, the mechanism model-based method and the knowledge-based method are only applicable to certain unit equipment with clear mechanisms in the process. In addition, modern industry has been equipped with a large number of sensors and other instrumentation equipment, which have stored a large amount of process historical data. With the rapid development of process data analysis technologies such as multivariate statistical analysis, statistical learning, cluster analysis, and neural networks, data-based methods have become the mainstream method and research hotspot for process industry process monitoring in the past 20 years. This method does not require the establishment of an accurate mechanism model, nor does it rely on expert knowledge. Instead, it guides process monitoring by extracting process information from the data. It has the advantages of flexible modeling, convenient implementation, simple and economical, and strong versatility.

[0005] Data-based process monitoring methods have received great attention. Modern industrial process data often exhibit dynamic, nonlinear and rich spatiotemporal characteristics. In the collected data, there is a certain autocorrelation between process variables and time, and there is mutual correlation between variables. In traditional methods, traditional autoencoders, such as stacked sparse autoencoders (SSAE), have excellent nonlinear feature extraction capabilities, but they ignore the dynamic correlation between sample data and lack the ability to extract local spatial features of data. At the same time, the current sample has spatial neighbors and temporal neighbors. Spatial neighbors refer to multiple samples composed of samples with the smallest distance from the current sample in the sample feature space, and temporal neighbors refer to multiple samples with the closest sampling time to the current time. Due to the spatial coupling between variables in process data and the correlation between samples, failure to consider spatiotemporal neighbors will limit the effect of process monitoring.

[0006] Therefore, the traditional process monitoring methods have the problems of ignoring the dynamic correlation between sample data, not considering the existence of spatial and temporal neighbors of samples at a given moment, and the interaction between temporal and spatial neighbors. Summary of the invention

[0007] The present invention aims at the problems that traditional process monitoring methods ignore the dynamic correlation between sample data, fail to consider the existence of spatial neighborhood and temporal neighborhood of moment samples, and the interaction between temporal neighborhood and spatial neighborhood.

[0008] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0009] In one aspect, the present invention provides an industrial process fault detection method with enhanced data spatiotemporal neighborhood features, the method being implemented by an electronic device, the method comprising:

[0010] S1. Obtain process data of hot rolling of the strip steel to be tested, and perform feature enhancement on the process data.

[0011] S2. Input the feature-enhanced process data into the constructed spatiotemporal-stacked sparse autoencoder-canonical variable analysis TS-SSAE-CVA model.

[0012] Among them, the spatiotemporal-stacked sparse autoencoder-canonical variate analysis TS-SSAE-CVA model includes a spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model and a canonical variate analysis CVA sub-model.

[0013] S3. Based on the process data after feature enhancement, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model and the canonical variable analysis CVA sub-model, the fault detection results of the hot strip rolling process are obtained.

[0014] Optionally, the construction process of the spatiotemporal-stacked sparse autoencoder-canonical variate analysis TS-SSAE-CVA model in S2 includes:

[0015] S21. Obtain training data for hot strip rolling.

[0016] S22. Perform feature enhancement on the training data to obtain training data after feature concatenation.

[0017] S23. Input the training data after feature splicing into the constructed spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model to obtain the dimensionality reduction data corresponding to the training data after feature splicing.

[0018] S24. Input the dimension-reduced data into the constructed canonical variable analysis (CVA) sub-model to obtain the control limits and global statistics of the TS-SSAE sub-model.

[0019] Optionally, the step of performing feature enhancement on the training data in S22 to obtain feature-concatenated training data includes:

[0020] S221. For each sample in the training data, use the k-nearest neighbor algorithm to perform feature enhancement based on the time neighborhood and feature enhancement based on the space neighborhood, respectively, to obtain training data with spatial neighborhood feature enhancement for each sample and training data with temporal neighborhood feature enhancement for each sample.

[0021] S222. Standardize the spatial neighborhood feature enhanced training data of each sample and the temporal neighborhood feature enhanced training data of each sample to obtain the standardized spatial neighborhood feature enhanced training data of each sample and the standardized temporal neighborhood feature enhanced training data of each sample.

[0022] S223, performing feature concatenation on each sample with the standardized spatial neighborhood feature enhanced training data of each sample and the standardized temporal neighborhood feature enhanced training data of each sample to obtain feature concatenated training data.

[0023] Optionally, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model includes a first TS-SSAE sub-model and a second TS-SSAE sub-model.

[0024] In S23, the training data after feature concatenation is input into the constructed spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model to obtain the dimensionality reduction data corresponding to the training data after feature concatenation, including:

[0025] The standardized spatial neighborhood feature enhanced training data is input into the first TS-SSAE sub-model to obtain the dimensionality reduction data corresponding to the standardized spatial neighborhood feature enhanced training data.

[0026] The standardized training data with enhanced temporal neighborhood features is input into the second TS-SSAE sub-model to obtain the dimensionality reduction data corresponding to the standardized training data with enhanced temporal neighborhood features.

[0027] Optionally, the construction process of the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model in S23 includes:

[0028] According to the training data after feature splicing and the back propagation algorithm, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model is trained to obtain the constructed TS-SSAE sub-model.

[0029] Optionally, in S24, the dimension reduction data is input into the constructed canonical variable analysis CVA sub-model to obtain the control limits and global statistics of the TS-SSAE sub-model, including:

[0030] S241. Input the dimension-reduced data into the constructed canonical variate analysis (CVA) sub-model to obtain the local statistics of the state subspace and the local statistics of the residual subspace of the TS-SSAE sub-model.

[0031] S242. According to the local statistics of the state subspace, the local statistics of the residual subspace and the kernel density estimation KDE algorithm, the control limits of the TS-SSAE submodel are obtained.

[0032] S243. According to the local statistics of the state subspace, the local statistics of the residual subspace and Bayesian reasoning, the global statistics of the TS-SSAE submodel are obtained.

[0033] Optionally, the CVA sub-model includes a first CVA sub-model and a second CVA sub-model.

[0034] In S241, the dimension reduction data is input into the constructed canonical variable analysis CVA sub-model to obtain the local statistics of the state subspace and the local statistics of the residual subspace of the TS-SSAE sub-model, including:

[0035] The reduced dimensionality data corresponding to the standardized spatial neighborhood feature enhanced training data are input into the first CVA sub-model to obtain the spatial neighborhood local state subspace and the spatial neighborhood local residual subspace of the reduced dimensionality data corresponding to the standardized spatial neighborhood feature enhanced training data, and the local statistics of the spatial neighborhood local state subspace and the local statistics of the spatial neighborhood local residual subspace are calculated based on the spatial neighborhood local state subspace and the spatial neighborhood local residual subspace of the reduced dimensionality data corresponding to the standardized spatial neighborhood feature enhanced training data.

[0036] The reduced dimensionality data corresponding to the standardized training data with enhanced temporal neighborhood features are input into the second CVA sub-model to obtain the temporal neighborhood local state subspace and the temporal neighborhood local residual subspace of the reduced dimensionality data corresponding to the standardized training data with enhanced temporal neighborhood features. The local statistics of the temporal neighborhood local state subspace and the local statistics of the temporal neighborhood local residual subspace are calculated based on the temporal neighborhood local state subspace and the temporal neighborhood local residual subspace of the reduced dimensionality data corresponding to the standardized training data with enhanced temporal neighborhood features.

[0037] Optionally, in S243, according to the local statistics of the state subspace, the local statistics of the residual subspace and Bayesian reasoning, the global statistics of the TS-SSAE submodel are obtained, including:

[0038] The failure probability of the local statistics of the state subspace and the failure probability of the local statistics of the residual subspace are calculated. The calculation formulas are shown in the following equations (1) and (2):

[0039]

[0040]

[0041] in, represents the failure probability of the local statistics of the state subspace, that is, T in the bth subblock 2 The probability of process abnormality under statistical conditions, F represents the abnormal conditions of the process, represents the T of the b-th sub-block 2 The control limit of the statistic, exp() represents an exponential function with the natural constant e as the base, T represents the b-th sub-block 2 Statistics, represents the T of the b-th sub-block 2 The prior probability of statistical failure, P(N) represents α, α is the confidence level, N represents the normal condition of the process, P(F) represents 1-α, P(F|Q b) represents the probability of process abnormality under the Q statistic condition of the b-th sub-block, P(Q b |F)=exp(-Q b,lim / Q b ), Q b,lim represents the control limit of the Q statistic of the b-th sub-block, Q b represents the Q statistic of the b-th sub-block, represents the prior probability of failure under the Q statistic of the b-th sub-block, P(Q b |N)=exp(-Q b / Q b,lim ).

[0042] According to the failure probability of the local statistics of the state subspace, the failure probability of the local statistics of the residual subspace and Bayesian reasoning, the global statistics of the TS-SSAE submodel are obtained. The calculation formula is shown in the following equations (3) and (4):

[0043]

[0044]

[0045] in, Indicates that in T 2 Global statistics and BIC of TS-SSAE submodel under statistical conditions Q represents the global statistics of the TS-SSAE sub-model under the Q statistic condition, B represents the number of samples, and b represents the b-th sample, that is, the b-th sub-block.

[0046] Optionally, according to the process data after feature enhancement, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model and the canonical variable analysis CVA sub-model in S3, the fault detection results of the hot strip rolling process are obtained, including:

[0047] According to the process data after feature enhancement, the spatiotemporal-sparse autoencoder TS-SSAE sub-model and the canonical variable analysis CVA sub-model, the control limits of the TS-SSAE sub-model and the global statistics of the TS-SSAE sub-model are obtained.

[0048] If the global statistics of the TS-SSAE sub-model is less than or equal to the control limit of the TS-SSAE sub-model, it is judged that the hot strip rolling process is running normally; if the global statistics of the TS-SSAE sub-model is greater than the control limit of the TS-SSAE sub-model, it is judged that a fault occurs in the hot strip rolling process.

[0049] On the other hand, the present invention provides an industrial process fault detection device with enhanced data spatiotemporal neighborhood features, which is applied to implement an industrial process fault detection method with enhanced data spatiotemporal neighborhood features, and the device comprises:

[0050] The acquisition module is used to acquire the process data of the hot rolling of the strip steel to be detected and to perform feature enhancement on the process data.

[0051] The input module is used to input the feature-enhanced process data into the constructed spatiotemporal-stacked sparse autoencoder-canonical variable analysis TS-SSAE-CVA model.

[0052] Among them, the spatiotemporal-stacked sparse autoencoder-canonical variate analysis TS-SSAE-CVA model includes a spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model and a canonical variate analysis CVA sub-model.

[0053] The output module is used to obtain the fault detection results of the hot strip rolling process based on the process data after feature enhancement, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model and the canonical variable analysis CVA sub-model.

[0054] Optionally, the input module is further configured to:

[0055] S21. Obtain training data for hot strip rolling.

[0056] S22. Perform feature enhancement on the training data to obtain training data after feature concatenation.

[0057] S23. Input the training data after feature splicing into the constructed spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model to obtain the dimensionality reduction data corresponding to the training data after feature splicing.

[0058] S24. Input the dimension-reduced data into the constructed canonical variable analysis (CVA) sub-model to obtain the control limits and global statistics of the TS-SSAE sub-model.

[0059] Optionally, the input module is further configured to:

[0060] S221. For each sample in the training data, use the k-nearest neighbor algorithm to perform feature enhancement based on the time neighborhood and feature enhancement based on the space neighborhood, respectively, to obtain training data with spatial neighborhood feature enhancement for each sample and training data with temporal neighborhood feature enhancement for each sample.

[0061] S222. Standardize the spatial neighborhood feature enhanced training data of each sample and the temporal neighborhood feature enhanced training data of each sample to obtain the standardized spatial neighborhood feature enhanced training data of each sample and the standardized temporal neighborhood feature enhanced training data of each sample.

[0062] S223, performing feature concatenation on each sample with the standardized spatial neighborhood feature enhanced training data of each sample and the standardized temporal neighborhood feature enhanced training data of each sample to obtain feature concatenated training data.

[0063] Optionally, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model includes a first TS-SSAE sub-model and a second TS-SSAE sub-model.

[0064] Input module, further used to:

[0065] The standardized spatial neighborhood feature enhanced training data is input into the first TS-SSAE sub-model to obtain the dimensionality reduction data corresponding to the standardized spatial neighborhood feature enhanced training data.

[0066] The standardized training data with enhanced temporal neighborhood features is input into the second TS-SSAE sub-model to obtain the dimensionality reduction data corresponding to the standardized training data with enhanced temporal neighborhood features.

[0067] Optionally, the input module is further configured to:

[0068] According to the training data after feature splicing and the back propagation algorithm, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model is trained to obtain the constructed TS-SSAE sub-model.

[0069] Optionally, the input module is further configured to:

[0070] S241. Input the dimension-reduced data into the constructed canonical variate analysis (CVA) sub-model to obtain the local statistics of the state subspace and the local statistics of the residual subspace of the TS-SSAE sub-model.

[0071] S242. According to the local statistics of the state subspace, the local statistics of the residual subspace and the kernel density estimation KDE algorithm, the control limits of the TS-SSAE submodel are obtained.

[0072] S243. According to the local statistics of the state subspace, the local statistics of the residual subspace and Bayesian reasoning, the global statistics of the TS-SSAE submodel are obtained.

[0073] Optionally, the CVA sub-model includes a first CVA sub-model and a second CVA sub-model.

[0074] Input module, further used to:

[0075] The reduced dimensionality data corresponding to the standardized spatial neighborhood feature enhanced training data are input into the first CVA sub-model to obtain the spatial neighborhood local state subspace and the spatial neighborhood local residual subspace of the reduced dimensionality data corresponding to the standardized spatial neighborhood feature enhanced training data, and the local statistics of the spatial neighborhood local state subspace and the local statistics of the spatial neighborhood local residual subspace are calculated based on the spatial neighborhood local state subspace and the spatial neighborhood local residual subspace of the reduced dimensionality data corresponding to the standardized spatial neighborhood feature enhanced training data.

[0076] The reduced dimensionality data corresponding to the standardized training data with enhanced temporal neighborhood features are input into the second CVA sub-model to obtain the temporal neighborhood local state subspace and the temporal neighborhood local residual subspace of the reduced dimensionality data corresponding to the standardized training data with enhanced temporal neighborhood features. The local statistics of the temporal neighborhood local state subspace and the local statistics of the temporal neighborhood local residual subspace are calculated based on the temporal neighborhood local state subspace and the temporal neighborhood local residual subspace of the reduced dimensionality data corresponding to the standardized training data with enhanced temporal neighborhood features.

[0077] Optionally, the input module is further configured to:

[0078] The failure probability of the local statistics of the state subspace and the failure probability of the local statistics of the residual subspace are calculated. The calculation formulas are shown in the following equations (1) and (2):

[0079]

[0080]

[0081] in, represents the failure probability of the local statistics of the state subspace, that is, T in the bth subblock 2 The probability of process abnormality under statistical conditions, F represents the abnormal conditions of the process, represents the T of the b-th sub-block 2 The control limit of the statistic, exp() represents an exponential function with the natural constant e as the base, represents the T of the b-th sub-block 2 Statistics, represents the T of the b-th sub-block 2 The prior probability of statistical failure, P(N) represents α, α is the confidence level, N represents the normal condition of the process, P(F) represents 1-α, P(F|Q b) represents the probability of process abnormality under the Q statistic condition of the b-th sub-block, P(Q b |F)=exp(-Q b,lim / Q b ), Q b,lim represents the control limit of the Q statistic of the b-th sub-block, Q b represents the Q statistic of the b-th sub-block, represents the prior probability of failure under the Q statistic of the b-th sub-block, P(Q b |N)=exp(-Q b / Q b,lim ).

[0082] According to the failure probability of the local statistics of the state subspace, the failure probability of the local statistics of the residual subspace and Bayesian reasoning, the global statistics of the TS-SSAE submodel are obtained. The calculation formula is shown in the following equations (3) and (4):

[0083]

[0084]

[0085] in, Indicates that in T 2 Global statistics and BIC of TS-SSAE submodel under statistical conditions Q represents the global statistics of the TS-SSAE sub-model under the Q statistic condition, B represents the number of samples, and b represents the b-th sample, that is, the b-th sub-block.

[0086] Optionally, the output module is further configured to:

[0087] According to the process data after feature enhancement, the spatiotemporal-sparse autoencoder TS-SSAE sub-model and the canonical variable analysis CVA sub-model, the control limits of the TS-SSAE sub-model and the global statistics of the TS-SSAE sub-model are obtained.

[0088] If the global statistics of the TS-SSAE sub-model is less than or equal to the control limit of the TS-SSAE sub-model, it is judged that the hot strip rolling process is running normally; if the global statistics of the TS-SSAE sub-model is greater than the control limit of the TS-SSAE sub-model, it is judged that a fault occurs in the hot strip rolling process.

[0089] On the one hand, an electronic device is provided, which includes a processor and a memory, wherein at least one instruction is stored in the memory, and the at least one instruction is loaded and executed by the processor to implement the industrial process fault detection method with enhanced spatiotemporal neighborhood features of data.

[0090] On the one hand, a computer-readable storage medium is provided, in which at least one instruction is stored, and the at least one instruction is loaded and executed by a processor to implement the above-mentioned industrial process fault detection method with enhanced spatiotemporal neighborhood features of data.

[0091] Compared with the prior art, the above technical solution has at least the following beneficial effects:

[0092] The above scheme utilizes the spatial neighborhood and temporal neighborhood of the sample at the current moment to expand the neighbor features of the current sample. The addition of temporal neighborhood information will be helpful in dealing with dynamic problems. The introduction of spatial similarity allows the reconstructed sample to consider the correlation between time and space at the same time, which can more fully reflect the actual situation of the production process and reflect the fault information contained therein.

[0093] By utilizing the feature expression capability of SSAE, the learned high-dimensional sparse features are helpful for fault detection and are more suitable for practical application processes.

[0094] CVA is used to further extract nonlinear dynamic characteristics and construct monitoring statistics and control limits in the state subspace and residual subspace to achieve ideal monitoring effects.

[0095] By using Bayesian decision making to fuse local models, the state change information of the hot rolling process can be obtained more intuitively. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0097] Figure 1 It is a schematic flow chart of an industrial process fault detection method with enhanced data spatiotemporal neighborhood features provided by an embodiment of the present invention;

[0098] Figure 2 It is a schematic diagram of the process flow of hot strip rolling provided by an embodiment of the present invention;

[0099] Figure 3 is a schematic diagram of a finishing rolling process provided by an embodiment of the present invention;

[0100] Figure 4 is a schematic diagram of a model of a sparse stacked autoencoder provided in an embodiment of the present invention;

[0101] Figure 5 Schematic diagram of a spatiotemporal feature enhancement-sparse stacking autoencoder model provided by an embodiment of the present invention;

[0102] Figure 6 It is a block diagram of an industrial process fault detection device with enhanced data spatiotemporal neighborhood features provided by an embodiment of the present invention;

[0103] Figure 7 It is a structural schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0104] In order to make the purpose, technical solution and advantages of the embodiment of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all of the embodiments. Based on the described embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0105] like Figure 1 As shown, an embodiment of the present invention provides an industrial process fault detection method with enhanced data spatiotemporal neighborhood features, which can be implemented by an electronic device. Figure 1 The flowchart of the industrial process fault detection method with enhanced data spatiotemporal neighborhood features is shown in FIG. The processing flow of the method may include the following steps:

[0106] S1. Obtain process data of hot rolling of the strip steel to be tested, and perform feature enhancement on the process data.

[0107] In a feasible implementation, the hot strip rolling process is a steel production process with complex mechanism, large scale, high efficiency and multiple steps, and is a typical complex process industry. Figure 2 As shown in FIG. 1 , the production line is mainly composed of a heating furnace, a roughing mill, a hot output roller and a flying shear, a finishing mill, a laminar cooling and a coiling unit, etc. The present invention is described using the process data of the finishing mill as test data. Figure 3 As shown, the finishing mill group is usually composed of seven frames, and the frames are connected together through a control loop. Each frame has two support rolls on the left and right, two working rolls in the middle, and is equipped with a hydraulic system to provide the required rolling force and bending roll force. The thickness of the steel and the smooth forward movement of the strip can be controlled. Therefore, the finishing process data collected on site will be used as test data.

[0108] Furthermore, for each sample, feature enhancement is performed mainly based on the temporal neighborhood and the spatial neighborhood. After the test data is enhanced with the corresponding temporal and spatial features, the data with the corresponding feature enhancement is used for standardization, and the standardized data is the training data for each sub-model.

[0109] For example, the samples in the finishing process data are enhanced in spatiotemporal features with spatial neighborhood as the main body and then standardized to obtain preprocessed data 1; the samples in the finishing process data are enhanced in spatiotemporal features with time neighborhood as the main body and then standardized to obtain preprocessed data 2.

[0110] S2. Input the feature-enhanced process data into the constructed spatiotemporal-stacked sparse autoencoder-canonical variable analysis TS-SSAE-CVA model.

[0111] Among them, the spatiotemporal-stacked sparse autoencoder-canonical variate analysis TS-SSAE-CVA model includes a spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model and a canonical variate analysis CVA (Canonical Variate Analysis) sub-model.

[0112] Optionally, the construction process of the spatiotemporal-stacked sparse autoencoder-canonical variable analysis TS-SSAE-CVA model in S2 may include the following steps S21-S24:

[0113] S21. Obtain training data for hot strip rolling.

[0114] The training data is the normal operation data at continuous moments.

[0115] In a feasible implementation, the present invention studies the hot strip rolling process with spatiotemporal characteristics, wherein the finishing rolling process plays an important role in the entire hot strip rolling process, and can ensure high-precision, high-efficiency, continuous and stable production of the final product. Therefore, the finishing rolling process data collected on-site is used as training data to construct a monitoring model for the hot strip rolling process.

[0116] S22. Perform feature enhancement on the training data to obtain training data after feature concatenation.

[0117] In a feasible implementation, feature enhancement is performed on each sample in the standardized training data, mainly based on the temporal neighborhood and mainly based on the spatial neighborhood, to obtain training data obtained by concatenating the newly added features with the original features.

[0118] Optionally, the above step S22 may include the following steps S221-S223:

[0119] S221. For each sample in the training data, use the k-nearest neighbor algorithm to perform feature enhancement based on the time neighborhood and feature enhancement based on the space neighborhood, respectively, to obtain feature-enhanced training data; wherein the feature-enhanced training data includes training data with spatial neighborhood feature enhancement and training data with temporal neighborhood feature enhancement.

[0120] S222, standardizing the feature-enhanced training data to obtain standardized training data; wherein the standardized training data includes standardized spatial neighborhood feature-enhanced training data and standardized temporal neighborhood feature-enhanced training data.

[0121] S223, performing feature concatenation on the training data and the standardized training data to obtain feature concatenated training data.

[0122] In a feasible implementation, the KNN (k-Nearest Neighbor) algorithm is used to construct the spatiotemporal features of the training data based on the time neighborhood and the space domain, respectively, to form training data with enhanced spatiotemporal features.

[0123] Specifically, the training data of each sample in the training data is enhanced with the time neighborhood as the main body and the spatial neighborhood as the main body, respectively, to obtain the training data after the newly added features are spliced, which may include the following steps A221-A224:

[0124] A221. Considering the spatial and temporal neighborhoods of each sample in the continuous moment process, for the TS-SSAE (Tempora Spatial-Sparse Stacked Autoencoder) method, the spatial neighborhood can be selected under the constraints of the time window, and the spatial similarity of the current sample is added to the temporal neighborhood. Then the neighborhood reconstruction information and the sample at the current moment are combined as the input of SSAE to extract important information of the current sample and the neighborhood.

[0125] SSAE Figure 4 As shown, the objective function is shown in the following formula (1), and the sparsity restriction makes the extracted intermediate layer features contain the most important information about the current sample and the reconstruction neighborhood:

[0126]

[0127] In the formula, the left part is the reconstruction error of the autoencoder, the objective function is the right part of the equation, n represents the number of samples, x (i) represents the i-th actual sample, represents the i-th prediction sample, β is the weight controlling the sparsity penalty, and S is the number of hidden layer neurons. It is shown by the following formula (2). represents the average activation of hidden unit j, as shown in the following formula (3). ρ is a sparsity parameter with a value of zero, and its value determines the sparsity of the neuron:

[0128]

[0129]

[0130] Minimizing the right side of equation (2) will give and ρ are as equal as possible, so that the average activation of the hidden unit is smaller, in order to achieve the purpose of sparse hidden layers.

[0131] Furthermore, the TS-SSAE model is Figure 5 As shown in the figure, the proposed method can be divided into two parts. First, the original process matrix is ​​defined as X = (x1, x2, ..., x n )∈R m×n , where n is the number of samples and m is the number of features. Considering the dynamic characteristics of the process, the current sample can only use historical samples. The number of spatial neighbors is k, and the time window length L = 2k. The TS-SSAE method consists of TS-SSAE1 and TS-SSAE2.

[0132] In TS-SSAE1-CVA, for the current sample x t , using KNN from the time window L = (x t-1 , x t-2 , …, x t-L ) selects k spatial neighbors, which can be expressed as Represents the current sample x t The i-th spatial neighbor of i represents the time deviation from the current moment. There is a dynamic relationship between the samples in the appropriate time window and the current moment, and these neighbors are related to x t The Euclidean distance between them is the smallest, so we can assume that they are t Are highly relevant.

[0133] A222. Calculate the time weight. Calculate the time weight for the spatial domain. For the spatial neighborhood X=(x1, x2, ..., x n )∈R m×n , taking into account the serial correlation on the time scale. First, calculate each neighbor and the current sample x t The temporal distance between them can then be used to construct the temporal weight. The temporal distance is shown in the following equation (4), which considers the temporal deviation of all k spatial neighbors and converts it into the relevant contribution of the nth neighbor to the current sample. In addition, the Gaussian kernel function is introduced to strengthen the temporal constraints at different times. Finally, the temporal weight is shown in the following equation (5), and the weight of each neighbor is expressed as w t,i , the sum of the weights is set to 1:

[0134]

[0135]

[0136] Based on the time weight w t,i , the spatial neighbors can be reconstructed as shown in the following formula (6):

[0137]

[0138] The reconstructed sample obtained by formula (6) It means reconstructing neighbors with high similarity through time series correlation to expand the current sample as neighborhood features.

[0139] A223. Construct the TS-SSAE model. The TS-SSAE-1 model is based on spatial neighbors and uses serial correlation as a time weight to reconstruct neighbors to expand the current sample. Therefore, it also considers the topological structure of time and space. The input at the current time can be expressed as Will be used as input for SSAE.

[0140] A224, TS-SSAE model Figure 5 As shown, it can be constructed by the following method. In TS-SSAE-1, the spatial neighborhood is the main body, but x t The temporal neighborhood of x t With more obvious serial correlation, the addition of temporal neighbor information will be beneficial in dealing with dynamic problems. In TS-SSAE-2, temporal neighbors are used to reconstruct neighbor information. First, for the current sample x t , select m temporal neighbors in the time window L; they can be expressed as For the current sample x t Time Neighborhood The present invention considers the similarity on the spatial scale. The spatial similarity is shown in the following formula (7), and then the spatial weight is calculated as shown in the following formula (8). The introduction of spatial similarity allows the reconstruction sample to consider the correlation between time and space at the same time. The defined reconstruction expression is shown in the following formula (9):

[0141]

[0142]

[0143]

[0144] Then, similar to the TS-SSAE1 method, construct Used as input for SSAE to build the TS-SSAE2 model.

[0145] S23. Input the training data after feature splicing into the constructed spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model to obtain the dimensionality reduction data corresponding to the training data after feature splicing.

[0146] Among them, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model includes a first TS-SSAE sub-model and a second TS-SSAE sub-model.

[0147] Optionally, the construction process of the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model in S23 includes:

[0148] According to the training data after feature splicing and the back propagation algorithm, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model is trained to obtain the constructed TS-SSAE sub-model.

[0149] In a feasible implementation, the training data is spliced ​​with time neighborhood as the main body and space neighborhood as the main body features, KNN is used to calculate the spatial distance and time distance between samples, BP (Error Back Propagation) is used to train the SSAE model respectively, and the loss function is MAE (Mean Absolute Error) to obtain a trained SSAE model.

[0150] Optionally, the above step S23 may include the following steps S231-S232:

[0151] S231. Input the standardized spatial neighborhood feature enhanced training data into the first TS-SSAE sub-model to obtain the dimensionality reduction data corresponding to the standardized spatial neighborhood feature enhanced training data.

[0152] S232. Input the standardized training data with enhanced temporal neighborhood features into the second TS-SSAE sub-model to obtain the dimensionality reduction data corresponding to the standardized training data with enhanced temporal neighborhood features.

[0153] In a feasible implementation, the test data collected on-site is input into the trained SSAE sub-model after spatiotemporal feature enhancement. The trained SSAE sub-model can obtain the hidden layer features after vector dimension compression through forward propagation.

[0154] For example, by inputting the data 1 of the finishing rolling process after being enhanced with spatial neighborhood as the main feature into the trained SSAE sub-model 1, TS dimension reduction data 1 can be obtained; by inputting the data 2 of the finishing rolling process after being enhanced with temporal neighborhood as the main feature into the trained SSAE sub-model 2, TS dimension reduction data 2 can be obtained.

[0155] S24. Input the dimension-reduced data into the constructed canonical variable analysis (CVA) sub-model to obtain the control limits and global statistics of the TS-SSAE sub-model.

[0156] Optionally, the above step S24 may include the following steps S241-S243:

[0157] S241. Input the dimension-reduced data into the constructed canonical variate analysis (CVA) sub-model to obtain the local statistics of the state subspace and the local statistics of the residual subspace of the TS-SSAE sub-model.

[0158] In a feasible implementation, the reduced-dimensional training data are respectively input into the CVA sub-model, and the parameters are tuned to obtain the reduced-dimensional state space and residual space of each set of feature-enhanced training data; based on the reduced-dimensional state space and residual space of each set of feature-enhanced training data, the local statistics and control limits of the corresponding model are obtained, and then the global statistics and control limits are obtained.

[0159] The CVA sub-model includes a first CVA sub-model and a second CVA sub-model.

[0160] Optionally, the above step S241 may include the following steps S2411-S2412:

[0161] S2411. Utilizing the nonlinear characteristics of the training data after spatiotemporal feature splicing, the reduced-dimensional data corresponding to the standardized spatial neighborhood feature-enhanced training data is input into the first CVA sub-model to obtain the spatial neighborhood local state subspace and the spatial neighborhood local residual subspace of the reduced-dimensional data corresponding to the standardized spatial neighborhood feature-enhanced training data, and the local statistics of the spatial neighborhood local state subspace and the local statistics of the spatial neighborhood local residual subspace are calculated.

[0162] S2412. Input the reduced-dimensional data corresponding to the standardized time neighborhood feature enhanced training data into the second CVA sub-model, obtain the time neighborhood local state subspace and the time neighborhood local residual subspace of the reduced-dimensional data corresponding to the standardized time neighborhood feature enhanced training data, and calculate the local statistics of the time neighborhood local state subspace and the local statistics of the time neighborhood local residual subspace.

[0163] In a feasible implementation, the reduced-dimensional training data is input into the CVA sub-model, and the parameters are tuned to obtain the state subspace and residual subspace of the training data after preprocessing by each TS-SSAE model, which may include:

[0164] CVA constructs past vectors and future vectors, and then can establish the maximum linear combination of the correlation between the past and the future.

[0165] Canonical variable analysis is a dimensionality reduction method that collects process variables at time k in a given normal operating system in a fault diagnosis method that does not involve quality variables. Among them, m x Represents the number of process variables. The past vector p is constructed as shown in the following formula (10): k and the future vector f k , and then the maximum linear combination of past and future correlations of process variables can be established:

[0166]

[0167]

[0168] Among them, l and h represent the lag of historical and future data vectors, that is, the length of historical and future observations.

[0169] For N sampling data, M = Nl-h+1 generates k∈[l+1,l+M], the past and future data vectors p at the moment k 、f k , to form the past and future Hankel matrices. Construct Y p and Y f The data matrix representing the past and the future is shown in the following formula (11):

[0170] Y p =[p p+1 , p p+2 , ..., p p+M ]∈R mp×M

[0171]

[0172] Then use equations (12) to (17) to estimate the sample covariance and cross covariance, construct the Hankel matrix H, and obtain the optimal projection matrix J and the canonical correlation coefficient matrix D for the historical and future data matrices. p , Y f Projecting into the canonical variable space, thus calculating the canonical matrix C with maximum correlation p , C f , determine the state vector z k and determine the residual vector e k , for Y p , Y f After standardization, Y p With Y f The sample covariance and cross covariance of can be estimated according to formula (12):

[0173]

[0174] The Hankel matrix H is constructed as follows:

[0175]

[0176] Where U and V are the left and right singular matrices of the Hankel matrix H; the singular value σ i (i=1, 2, ..., r) are arranged from large to small to form a diagonal matrix, which is defined as ∑.

[0177] Calculate the optimal projection matrix for historical and future data matrices And the normalized correlation coefficient matrix D. Using J and L, Y p , Y f Projecting into the canonical variable space, thus calculating the canonical matrix C with maximum correlation p , C f :

[0178] C p =JY p ∈R mp×M

[0179]

[0180]

[0181] The first n eigenvalues ​​of the correlation coefficient matrix D are the largest, and the best n value is selected as the dimension after dimensionality reduction according to the AIC criterion, where n < mp. The state vector z is determined using the first n main singular values k , and the remaining singular values ​​determine the residual vector e k . Use equation (15) to describe the state subspace and residual space:

[0182] z k =J n p k ∈R n

[0183] e k =F n p k ∈R mp (15)

[0184] In the formula, I is a unit vector of order n.

[0185] S242. According to the local statistics of the state subspace, the local statistics of the residual subspace and the kernel density estimation KDE algorithm, the control limits of the TS-SSAE submodel are obtained.

[0186] In one possible implementation, T 2 The monitoring statistics of the state subspace and residual subspace are constructed with the Q statistic. The statistics at the k sampling time are calculated as shown in formula (16):

[0187]

[0188]

[0189] The kernel density estimation method can be used to obtain T 2 The probability density function of Q is used to obtain the corresponding control limit. The kernel selected by the present invention is the Gaussian kernel function, as shown in formula (17):

[0190]

[0191] S243. According to the local statistics of the state subspace, the local statistics of the residual subspace and Bayesian reasoning, the global statistics of the TS-SSAE submodel are obtained.

[0192] Optionally, the above step S243 may include the following steps S2431-S2432:

[0193] S2431, define the failure probability of the local statistics of the state subspace generated by each sub-block and the failure probability of the local statistics of the residual subspace are calculated by the following equations (18) and (19):

[0194]

[0195]

[0196] in, represents the failure probability of the local statistics of the state subspace, that is, T in the bth subblock 2 The probability of process abnormality under the statistical conditions, N and F represent the normal and abnormal conditions of the process, respectively. Represents T of the b-th sub-block 2 The prior probabilities of normal and faulty conditions under the conditions of the statistics and Q statistics, P(N) and P(F), can be expressed as α and 1-α respectively, where α is the confidence level. P(Q b |N)、P(Q b |F) can be calculated by the following formulas (20), (21), (22) and (23):

[0197]

[0198]

[0199] P(Q b |N)=exp(-Q b / Q b,lim ) (twenty two)

[0200] P(Q b |F)=exp(-Q b,lim / Q b ) (twenty three)

[0201] in, Respectively represent the T of the b-th sub-block 2 Statistics and their control limits, exp() represents an exponential function with the natural constant e as the base, Q b , Q b,lim represent the Q statistic and its control limit of the b-th sub-block respectively.

[0202] S2432. According to the failure probability of the local statistics of the state subspace, the failure probability of the local statistics of the residual subspace and Bayesian reasoning, the global statistics of the TS-SSAE submodel are obtained. The calculation formula is shown in the following formula (24) (25):

[0203]

[0204]

[0205] Based on the Bayesian fusion strategy, two monitoring statistics can be obtained: and BIC Q , Indicates that in T 2 Global statistics and BIC of TS-SSAE submodel under statistical conditions Q represents the global statistics of the TS-SSAE sub-model under the Q statistic condition, B represents the number of samples, and b represents the b-th sample, that is, the b-th sub-block. Statistics and BIC Q The control limit of is the confidence level α. BIC Q Control limit exceeded confidence level or BIC Q >α, a fault is detected and further investigation is necessary. Otherwise, the monitoring process is considered normal.

[0206] For example, the SSAE hidden layer output data 1 with spatial neighbors as the main body and spatiotemporal features enhanced is input into the corresponding CVA model 1 to obtain the local state subspace and residual subspace of the reduced dimension data 1, and then the monitoring statistics of the state subspace 1 and residual subspace 1 of the hidden layer output data 1 are calculated; the SSAE hidden layer output data 2 with temporal neighbors as the main body and spatiotemporal features enhanced is input into the corresponding CVA model 2 to obtain the local state subspace 2 and residual subspace 2 of the reduced dimension data 2, and then the monitoring statistics of the state subspace and residual subspace of the hidden layer output data 2 are calculated; the global statistics and control limits are obtained based on Bayesian reasoning and the local state subspace 1 and local state subspace 2, the residual subspace statistics 1 and the residual subspace statistics 2.

[0207] S3. Based on the process data after feature enhancement, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model and the canonical variable analysis CVA sub-model, the fault detection results of the hot strip rolling process are obtained.

[0208] Optionally, the above step S3 may be:

[0209] According to the process data after feature enhancement, the spatiotemporal-sparse autoencoder TS-SSAE sub-model and the canonical variable analysis CVA sub-model, the control limits of the TS-SSAE sub-model and the global statistics of the TS-SSAE sub-model are obtained.

[0210] If the global statistic is less than or equal to the control limit, the hot strip rolling process is running normally; if the global statistic is greater than the control limit, the hot strip rolling process fails.

[0211] Aiming at the spatiotemporal characteristics and dynamics of steel rolling process data, the present invention utilizes the time neighbors and space neighbors at the current moment as the main body, enhances the spatiotemporal characteristics of industrial process data as the input of model training, utilizes the sparse stacked autoencoder to make the learned high-dimensional sparse features helpful for fault detection, trains the TS-SSAE-CVA model, obtains the statistics and control limits of the sub-model, and then obtains the global statistics; if the global statistics is less than the control limit, the hot strip rolling process runs normally, and if the global statistics is greater than the control limit, a fault occurs in the hot strip rolling process.

[0212] In an embodiment of the present invention, the spatial neighborhood and temporal neighborhood of the sample at the current moment can be used to expand the neighborhood features of the current sample. The addition of temporal neighborhood information will be beneficial for processing dynamic problems. The introduction of spatial similarity allows the reconstructed sample to consider the correlation between time and space at the same time, which can more fully reflect the actual situation of the production process and reflect the fault information contained therein.

[0213] By utilizing the feature expression capability of SSAE, the learned high-dimensional sparse features are helpful for fault detection and are more suitable for practical application processes.

[0214] CVA is used to further extract nonlinear dynamic characteristics and construct monitoring statistics and control limits in the state subspace and residual subspace to achieve ideal monitoring effects.

[0215] By using Bayesian decision making to fuse local models, the state change information of the hot rolling process can be obtained more intuitively.

[0216] like Figure 6 As shown, an embodiment of the present invention provides an industrial process fault detection device 600 with enhanced data spatiotemporal neighborhood features. The device 600 is applied to implement an industrial process fault detection method with enhanced data spatiotemporal neighborhood features. The device 600 includes:

[0217] The acquisition module 610 is used to acquire the process data of the hot rolling of the steel strip to be detected and perform feature enhancement on the process data.

[0218] The input module 620 is used to input the feature-enhanced process data into the constructed spatiotemporal-stacked sparse autoencoder-canonical variable analysis TS-SSAE-CVA model.

[0219] Among them, the spatiotemporal-stacked sparse autoencoder-canonical variate analysis TS-SSAE-CVA model includes a spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model and a canonical variate analysis CVA sub-model.

[0220] The output module 630 is used to obtain the fault detection results of the hot strip rolling process according to the process data after feature enhancement, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model and the canonical variable analysis CVA sub-model.

[0221] Optionally, the input module 620 is further configured to:

[0222] S21. Obtain training data for hot strip rolling.

[0223] S22. Perform feature enhancement on the training data to obtain training data after feature concatenation.

[0224] S23. Input the training data after feature splicing into the constructed spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model to obtain the dimensionality reduction data corresponding to the training data after feature splicing.

[0225] S24. Input the dimension-reduced data into the constructed canonical variable analysis (CVA) sub-model to obtain the control limits and global statistics of the TS-SSAE sub-model.

[0226] Optionally, the input module 620 is further configured to:

[0227] S221. For each sample in the training data, use the k-nearest neighbor algorithm to perform feature enhancement based on the time neighborhood and feature enhancement based on the space neighborhood, respectively, to obtain training data with spatial neighborhood feature enhancement for each sample and training data with temporal neighborhood feature enhancement for each sample.

[0228] S222. Standardize the spatial neighborhood feature enhanced training data of each sample and the temporal neighborhood feature enhanced training data of each sample to obtain the standardized spatial neighborhood feature enhanced training data of each sample and the standardized temporal neighborhood feature enhanced training data of each sample.

[0229] S223, performing feature concatenation on each sample with the standardized spatial neighborhood feature enhanced training data of each sample and the standardized temporal neighborhood feature enhanced training data of each sample to obtain feature concatenated training data.

[0230] Optionally, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model includes a first TS-SSAE sub-model and a second TS-SSAE sub-model.

[0231] The input module 620 is further configured to:

[0232] The standardized spatial neighborhood feature enhanced training data is input into the first TS-SSAE sub-model to obtain the dimensionality reduction data corresponding to the standardized spatial neighborhood feature enhanced training data.

[0233] The standardized training data with enhanced temporal neighborhood features is input into the second TS-SSAE sub-model to obtain the dimensionality reduction data corresponding to the standardized training data with enhanced temporal neighborhood features.

[0234] Optionally, the input module 620 is further configured to:

[0235] According to the training data after feature splicing and the back propagation algorithm, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model is trained to obtain the constructed TS-SSAE sub-model.

[0236] Optionally, the input module 620 is further configured to:

[0237] S241. Input the dimension-reduced data into the constructed canonical variate analysis (CVA) sub-model to obtain the local statistics of the state subspace and the local statistics of the residual subspace of the TS-SSAE sub-model.

[0238] S242. According to the local statistics of the state subspace, the local statistics of the residual subspace and the kernel density estimation KDE algorithm, the control limits of the TS-SSAE submodel are obtained.

[0239] S243. According to the local statistics of the state subspace, the local statistics of the residual subspace and Bayesian reasoning, the global statistics of the TS-SSAE submodel are obtained.

[0240] Optionally, the CVA sub-model includes a first CVA sub-model and a second CVA sub-model.

[0241] The input module 620 is further configured to:

[0242] The reduced dimensionality data corresponding to the standardized spatial neighborhood feature enhanced training data are input into the first CVA sub-model to obtain the spatial neighborhood local state subspace and the spatial neighborhood local residual subspace of the reduced dimensionality data corresponding to the standardized spatial neighborhood feature enhanced training data, and the local statistics of the spatial neighborhood local state subspace and the local statistics of the spatial neighborhood local residual subspace are calculated based on the spatial neighborhood local state subspace and the spatial neighborhood local residual subspace of the reduced dimensionality data corresponding to the standardized spatial neighborhood feature enhanced training data.

[0243] The reduced dimensionality data corresponding to the standardized training data with enhanced temporal neighborhood features are input into the second CVA sub-model to obtain the temporal neighborhood local state subspace and the temporal neighborhood local residual subspace of the reduced dimensionality data corresponding to the standardized training data with enhanced temporal neighborhood features. The local statistics of the temporal neighborhood local state subspace and the local statistics of the temporal neighborhood local residual subspace are calculated based on the temporal neighborhood local state subspace and the temporal neighborhood local residual subspace of the reduced dimensionality data corresponding to the standardized training data with enhanced temporal neighborhood features.

[0244] Optionally, the input module 620 is further configured to:

[0245] The failure probability of the local statistics of the state subspace and the failure probability of the local statistics of the residual subspace are calculated. The calculation formulas are shown in the following equations (1) and (2):

[0246]

[0247]

[0248] in, represents the failure probability of the local statistics of the state subspace, that is, T in the bth subblock 2 The probability of process abnormality under statistical conditions, F represents the abnormal conditions of the process, represents the T of the b-th sub-block2 The control limit of the statistic, exp() represents an exponential function with the natural constant e as the base, represents the T of the b-th sub-block 2 Statistics, represents the T of the b-th sub-block 2 The prior probability of statistical failure, P(N) represents α, α is the confidence level, N represents the normal condition of the process, P(F) represents 1-α, P(F|Q b ) represents the probability of process abnormality under the Q statistic condition of the b-th sub-block, P(Q b |F)=exp(-Q b,lim / Q b ), Q b,lim represents the control limit of the Q statistic of the b-th sub-block, Q b represents the Q statistic of the b-th sub-block, represents the prior probability of failure under the Q statistic of the b-th sub-block, P(Q b |N)=exp(-Q b / Q b,lim ).

[0249] According to the failure probability of the local statistics of the state subspace, the failure probability of the local statistics of the residual subspace and Bayesian reasoning, the global statistics of the TS-SSAE submodel are obtained. The calculation formula is shown in the following equations (3) and (4):

[0250]

[0251]

[0252] in, Indicates that in T 2 Global statistics and BIC of TS-SSAE submodel under statistical conditions Q represents the global statistics of the TS-SSAE sub-model under the Q statistic condition, B represents the number of samples, and b represents the b-th sample, that is, the b-th sub-block.

[0253] Optionally, the output module is further configured to:

[0254] According to the process data after feature enhancement, the spatiotemporal-sparse autoencoder TS-SSAE sub-model and the canonical variable analysis CVA sub-model, the control limits of the TS-SSAE sub-model and the global statistics of the TS-SSAE sub-model are obtained.

[0255] If the global statistics of the TS-SSAE sub-model is less than or equal to the control limit of the TS-SSAE sub-model, it is judged that the hot strip rolling process is running normally; if the global statistics of the TS-SSAE sub-model is greater than the control limit of the TS-SSAE sub-model, it is judged that a fault occurs in the hot strip rolling process.

[0256] In an embodiment of the present invention, the spatial neighborhood and temporal neighborhood of the sample at the current moment can be used to expand the neighborhood features of the current sample. The addition of temporal neighborhood information will be beneficial for processing dynamic problems. The introduction of spatial similarity allows the reconstructed sample to consider the correlation between time and space at the same time, which can more fully reflect the actual situation of the production process and reflect the fault information contained therein.

[0257] By utilizing the feature expression capability of SSAE, the learned high-dimensional sparse features are helpful for fault detection and are more suitable for practical application processes.

[0258] CVA is used to further extract nonlinear dynamic characteristics and construct monitoring statistics and control limits in the state subspace and residual subspace to achieve ideal monitoring effects.

[0259] By using Bayesian decision making to fuse local models, the state change information of the hot rolling process can be obtained more intuitively.

[0260] Figure 7 7 is a schematic diagram of the structure of an electronic device 700 provided by an embodiment of the present invention. The electronic device 700 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 701 and one or more memories 702, wherein the memory 702 stores at least one instruction, and the at least one instruction is loaded and executed by the processor 701 to implement the following industrial process fault detection method with enhanced spatiotemporal neighborhood features of data:

[0261] S1. Obtain process data of hot rolling of the strip steel to be tested, and perform feature enhancement on the process data.

[0262] S2. Input the feature-enhanced process data into the constructed spatiotemporal-stacked sparse autoencoder-canonical variable analysis TS-SSAE-CVA model.

[0263] Among them, the spatiotemporal-stacked sparse autoencoder-canonical variate analysis TS-SSAE-CVA model includes a spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model and a canonical variate analysis CVA sub-model.

[0264] S3. Based on the process data after feature enhancement, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model and the canonical variable analysis CVA sub-model, the fault detection results of the hot strip rolling process are obtained.

[0265] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions, which can be executed by a processor in a terminal to complete the industrial process fault detection method for enhancing the spatiotemporal neighborhood features of data. For example, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0266] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0267] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An industrial process fault detection method based on enhanced spatial and temporal neighborhood features of data, characterized in that: The method comprises: S1. Acquire process data of hot rolling of a steel strip to be detected, and perform feature enhancement on the process data; S2, input the feature-enhanced process data into the constructed spatiotemporal-stacked sparse autoencoder-canonical variable analysis TS-SSAE-CVA model; Wherein, the spatiotemporal-stacked sparse autoencoder-canonical variate analysis TS-SSAE-CVA model includes a spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model and a canonical variate analysis CVA sub-model; S3, obtaining the fault detection result of the hot strip rolling process according to the process data after the feature enhancement, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model and the canonical variable analysis CVA sub-model; The construction process of the spatiotemporal-stacked sparse autoencoder-canonical variable analysis TS-SSAE-CVA model in S2 includes: S21, obtaining training data of hot strip rolling; S22, performing feature enhancement on the training data to obtain feature-joined training data; S23, inputting the training data after the feature splicing into the constructed spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model to obtain the dimension reduction data corresponding to the training data after the feature splicing; S24, inputting the dimension reduction data into the constructed canonical variate analysis CVA sub-model to obtain the control limits and global statistics of the TS-SSAE sub-model; The step of performing feature enhancement on the training data in S22 to obtain feature-joined training data includes: S221, for each sample in the training data, using the k-nearest neighbor algorithm, respectively performing feature enhancement based on the time neighborhood and feature enhancement based on the space neighborhood, to obtain training data of each sample with spatial neighborhood feature enhancement and training data of each sample with temporal neighborhood feature enhancement; S222, standardizing the spatial neighborhood feature enhanced training data of each sample and the temporal neighborhood feature enhanced training data of each sample to obtain the standardized spatial neighborhood feature enhanced training data of each sample and the standardized temporal neighborhood feature enhanced training data of each sample; S223, performing feature splicing on each sample with the standardized spatial neighborhood feature enhanced training data of each sample and the standardized temporal neighborhood feature enhanced training data of each sample to obtain feature-spliced ​​training data; The step S24 inputs the dimension reduction data into the constructed canonical variate analysis CVA sub-model to obtain the control limits and global statistics of the TS-SSAE sub-model, including: S241, inputting the dimension reduction data into the constructed canonical variate analysis CVA sub-model to obtain the local statistics of the state subspace and the local statistics of the residual subspace of the TS-SSAE sub-model; S242, obtaining the control limit of the TS-SSAE submodel according to the local statistics of the state subspace, the local statistics of the residual subspace and the kernel density estimation KDE algorithm; S243. Obtain the global statistics of the TS-SSAE submodel according to the local statistics of the state subspace, the local statistics of the residual subspace, and Bayesian reasoning.

2. The method according to claim 1, characterized in that The spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model comprises a first TS-SSAE sub-model and a second TS-SSAE sub-model; The step S23 of inputting the training data after feature concatenation into the constructed spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model to obtain the dimension reduction data corresponding to the training data after feature concatenation includes: Inputting the standardized spatial neighborhood feature enhanced training data into the first TS-SSAE sub-model to obtain the dimension reduction data corresponding to the standardized spatial neighborhood feature enhanced training data; The standardized training data with enhanced temporal neighborhood features is input into the second TS-SSAE sub-model to obtain the dimensionality reduction data corresponding to the standardized training data with enhanced temporal neighborhood features.

3. The method according to claim 1, characterized in that The construction process of the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model in S23 includes: According to the training data after the feature splicing and the back propagation algorithm, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model is trained to obtain a constructed TS-SSAE sub-model.

4. The method according to claim 1, characterized in that: The CVA sub-model includes a first CVA sub-model and a second CVA sub-model; The step S241 of inputting the dimension reduction data into the constructed canonical variate analysis CVA sub-model to obtain the local statistics of the state subspace and the local statistics of the residual subspace of the TS-SSAE sub-model includes: Inputting the reduced-dimensionality data corresponding to the standardized spatial neighborhood feature-enhanced training data into the first CVA sub-model to obtain the spatial neighborhood local state subspace and the spatial neighborhood local residual subspace of the reduced-dimensionality data corresponding to the standardized spatial neighborhood feature-enhanced training data, and calculating the local statistics of the spatial neighborhood local state subspace and the local statistics of the spatial neighborhood local residual subspace according to the spatial neighborhood local state subspace and the spatial neighborhood local residual subspace of the reduced-dimensionality data corresponding to the standardized spatial neighborhood feature-enhanced training data; The reduced-dimensional data corresponding to the standardized training data with enhanced temporal neighborhood features are input into the second CVA sub-model to obtain the temporal neighborhood local state subspace and the temporal neighborhood local residual subspace of the reduced-dimensional data corresponding to the standardized training data with enhanced temporal neighborhood features, and the local statistics of the temporal neighborhood local state subspace and the local statistics of the temporal neighborhood local residual subspace are calculated based on the temporal neighborhood local state subspace and the temporal neighborhood local residual subspace of the reduced-dimensional data corresponding to the standardized training data with enhanced temporal neighborhood features.

5. The method according to claim 1, characterized in that: The step S243 obtains the global statistics of the TS-SSAE submodel according to the local statistics of the state subspace, the local statistics of the residual subspace, and Bayesian reasoning, including: The failure probability of the local statistic of the state subspace and the failure probability of the local statistic of the residual subspace are calculated, and the calculation formula is shown in the following formula (1) (2): in, represents the failure probability of the local statistics of the state subspace, that is, T in the bth subblock 2 The probability of process abnormality under statistical conditions, F represents the abnormal conditions of the process, represents the T of the b-th sub-block 2 The control limit of the statistic, exp() represents an exponential function with the natural constant e as the base, represents the T of the b-th sub-block 2 Statistics, T represents the b-th sub-block 2 The prior probability of statistical failure, P(N) represents α, α is the confidence level, N represents the normal condition of the process, P(F) represents 1-α, P(F|Q b ) represents the probability of process abnormality under the Q statistic condition of the b-th sub-block, P(Q b |F)=exp(-Q b,lim / Q b ), Q b,lim represents the control limit of the Q statistic of the b-th sub-block, Q b represents the Q statistic of the b-th sub-block, represents the prior probability of failure under the Q statistic of the b-th sub-block, P(Q b |N)=exp(-Q b / Q b,lim ); According to the failure probability of the local statistics of the state subspace, the failure probability of the local statistics of the residual subspace and Bayesian reasoning, the global statistics of the TS-SSAE submodel are obtained, and the calculation formula is shown in the following formula (3) (4): in, Indicates that in T 2 Global statistics and BIC of TS-SSAE submodel under statistical conditions Q represents the global statistics of the TS-SSAE sub-model under the Q statistic condition, B represents the number of samples, and b represents the b-th sample, that is, the b-th sub-block.

6. The method according to claim 1, characterized in that The S3 obtains the strip hot rolling process fault detection results according to the feature-enhanced process data, the spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model and the canonical variable analysis CVA sub-model, including: According to the feature-enhanced process data, the spatiotemporal-sparse autoencoder TS-SSAE sub-model, and the canonical variable analysis CVA sub-model, the control limits of the TS-SSAE sub-model and the global statistics of the TS-SSAE sub-model are obtained; If the global statistics of the TS-SSAE sub-model is less than or equal to the control limit of the TS-SSAE sub-model, it is judged that the hot strip rolling process is operating normally; if the global statistics of the TS-SSAE sub-model is greater than the control limit of the TS-SSAE sub-model, it is judged that a fault occurs in the hot strip rolling process.

7. An industrial process fault detection device with enhanced data spatiotemporal neighborhood features, characterized in that: The device comprises: An acquisition module, used for acquiring process data of hot rolling of a steel strip to be detected, and performing feature enhancement on the process data; An input module is used to input the feature-enhanced process data into the constructed spatiotemporal-stacked sparse autoencoder-canonical variable analysis TS-SSAE-CVA model; Wherein, the spatiotemporal-stacked sparse autoencoder-canonical variate analysis TS-SSAE-CVA model includes a spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model and a canonical variate analysis CVA sub-model; An output module is used to obtain a fault detection result of a strip hot rolling process according to the process data after feature enhancement, a spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model, and a canonical variable analysis CVA sub-model; The construction process of the spatiotemporal-stacked sparse autoencoder-canonical variable analysis TS-SSAE-CVA model includes: S21, obtaining training data of hot strip rolling; S22, performing feature enhancement on the training data to obtain feature-joined training data; S23, inputting the training data after the feature splicing into the constructed spatiotemporal-stacked sparse autoencoder TS-SSAE sub-model to obtain the dimension reduction data corresponding to the training data after the feature splicing; S24, inputting the dimension reduction data into the constructed canonical variate analysis CVA sub-model to obtain the control limits and global statistics of the TS-SSAE sub-model; The step of performing feature enhancement on the training data in S22 to obtain feature-joined training data includes: S221, for each sample in the training data, using the k-nearest neighbor algorithm, respectively performing feature enhancement based on the time neighborhood and feature enhancement based on the space neighborhood, to obtain training data of each sample with spatial neighborhood feature enhancement and training data of each sample with temporal neighborhood feature enhancement; S222, standardizing the spatial neighborhood feature enhanced training data of each sample and the temporal neighborhood feature enhanced training data of each sample to obtain the standardized spatial neighborhood feature enhanced training data of each sample and the standardized temporal neighborhood feature enhanced training data of each sample; S223, performing feature splicing on each sample with the standardized spatial neighborhood feature enhanced training data of each sample and the standardized temporal neighborhood feature enhanced training data of each sample to obtain feature-spliced ​​training data; The step S24 inputs the dimension reduction data into the constructed canonical variate analysis CVA sub-model to obtain the control limits and global statistics of the TS-SSAE sub-model, including: S241, inputting the dimension reduction data into the constructed canonical variate analysis CVA sub-model to obtain the local statistics of the state subspace and the local statistics of the residual subspace of the TS-SSAE sub-model; S242, obtaining the control limit of the TS-SSAE submodel according to the local statistics of the state subspace, the local statistics of the residual subspace and the kernel density estimation KDE algorithm; S243. Obtain the global statistics of the TS-SSAE submodel according to the local statistics of the state subspace, the local statistics of the residual subspace, and Bayesian reasoning.

Citation Information

Cited By

  • Abnormity detection method and device based on dynamic graph bidirectional transformation

    CN119862515A