A method for industrial process fault detection based on spatiotemporal feature decoupling

Abstract

This invention relates to an industrial process fault detection method based on spatiotemporal feature decoupling. A temporal branch and a spatial branch are constructed after the encoder of the MemAE model, with independent memory modules configured for each branch to store the memories of temporal dynamic relationships and spatial correlations, respectively, achieving effective decoupling of spatiotemporal features. DSVDD modules are introduced after the memory modules of the temporal and spatial branches, constructing an independent, highly compact feature space for each branch, enhancing the distinguishability between normal and faulty samples. Simultaneously, a fault detection statistic is designed based on the DSVDD-MemAE model, combining the reconstruction error with the DSVDD deviation error of the spatiotemporal branch, resulting in a more comprehensive and integrated judgment of abnormal data. This invention effectively decouples spatiotemporal features, expands the distinguishability between normal and abnormal samples, and effectively improves the fault detection rate and the reliability of industrial process monitoring.

Description

Technical Field

[0001] This invention belongs to the field of industrial process fault detection technology, specifically relating to an industrial process fault detection method based on spatiotemporal feature decoupling. Background Technology

[0002] With the rapid development of technologies such as Industry 4.0, the Industrial Internet of Things (IIoT), and artificial intelligence (AI), the industrial sector is showing a significant trend of deep integration and intelligentization, with industrial processes becoming increasingly large-scale and systems increasingly complex. In actual operation, a malfunction can lead to anything from a decrease in production efficiency or production interruption across the entire supply chain to severe equipment damage or even personal injury. Therefore, real-time fault detection in industrial processes is crucial for ensuring the safe and stable operation of modern industrial systems.

[0003] The widespread application and continuous upgrading of sensing and storage technologies in industrial processes have led to increasingly abundant process data, laying the foundation for deep model-based fault detection methods. Deep models can effectively learn complex features in data, and are particularly suitable for handling high-dimensional and nonlinear relationships common in industrial processes, thus achieving excellent detection performance. Autoencoders, as a typical deep learning model, can map high-dimensional industrial data to a low-dimensional feature space. By learning the data distribution characteristics under normal operating conditions, they can effectively identify abnormal operating conditions that deviate from the normal distribution, making them a classic model widely used in current industrial process fault detection research and applications. However, the powerful generalization ability of autoencoders may lead to the accurate reconstruction of some fault samples, making the reconstruction error difficult to distinguish from normal samples, thus causing false negatives.

[0004] To address the aforementioned issues, Memory-augmented Deep Autoencoders (MemAEs) effectively constrain their generalization ability by introducing a memory mechanism, thereby significantly enhancing the distinguishability between normal and fault samples. The original MemAE model primarily focuses on capturing the spatial correlations between process variables. However, chemical processes exhibit significant continuity and dynamism. To overcome the limitations of the original MemAE model in time-series modeling, existing research has improved the model architecture by introducing attention mechanisms. While these improvements enhance the processing capability for spatiotemporally coupled data, the learned feature representations are often spatiotemporally mixed. This strong coupling makes it difficult for the model to clearly distinguish between spatial and temporal anomalous patterns: temporal dynamics may overwhelm local faults in space, while spatial correlations may obscure temporal trend anomalies. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes an industrial process fault detection method based on spatiotemporal feature decoupling. A temporal branch and a spatial branch are constructed after the encoder of the MemAE model, and independent memory modules are configured for each branch structure to decouple the temporal dynamic features from the spatial correlation features. Based on this, independent Deep Support Vector Data Description (DSVDD) modules are configured for the decoupled spatiotemporal features, forming a DSVDD-MemAE model architecture. Finally, the reconstruction error of MemAE and the deviation error of DSVDD are fused to achieve comprehensive and reliable fault detection in industrial processes.

[0006] The technical solution adopted in this invention is as follows:

[0007] An industrial process fault detection method based on spatiotemporal feature decoupling includes the following steps:

[0008] (1) Obtain the multivariate time series dataset of industrial process, divide the dataset into training set and test set, and standardize the training set and test set respectively;

[0009] (2) Use time windows to serialize the training set data, construct the DSVDD-MemAE model, configure the hyperparameters of the model, and initialize its network parameters;

[0010] (3) Optimize the model loss function to obtain the completed DSVDD-MemAE model, thereby decoupling the temporal dynamic features and spatial correlation features of the training data;

[0011] (4) Based on the reconstruction error of the training samples and the deviation error of the DSVDD module in the spatiotemporal branch, obtain the statistics of the training samples and determine the control limits of the statistics.

[0012] (5) Input the test sample into the DSVDD-MemAE model trained in step (3). After decoupling the spatiotemporal features, obtain its corresponding statistics and compare them with the control limits to determine whether the sample is normal or abnormal, thus realizing industrial process fault detection.

[0013] As a preferred option, the specific process of step (1) is as follows: obtain a multivariate time series dataset of industrial processes, divide it into a training set and a test set, the training set contains only normal data, the test set contains normal data and abnormal data, and standardize it with Z-Score.

[0014] As a preferred embodiment, the specific process of step (2) is as follows: using a sliding window, sampling is performed along the sequence direction of the training set to divide the complete time series into a series of subsequence samples, and the subsequence samples are input into the DSVDD-MemAE model; the DSVDD-MemAE model contains five key modules: encoder, temporal branch, spatial branch, decoder and DSVDD module.

[0015] As a preferred embodiment, the specific process of step (3) is as follows: the DSVDD-MemAE model achieves dual constraints on normal data features by jointly optimizing the total loss function of MemAE and DSVDD, thereby improving the performance of process fault detection.

[0016] Preferably, the specific process of step (4) is as follows:

[0017] After training the DSVDD-MemAE model using the loss function in step (3), firstly, a statistic is constructed for fault detection. This statistic combines the reconstruction error of the training samples with the deviation error of the DSVDD module in the spatiotemporal branch. Then, based on this statistic, the statistic values ​​of all normal samples in the training set are obtained, and the upper quantile of its 95% confidence level is taken as the final control limit.

[0018] Preferably, the specific process of step (5) is as follows:

[0019] (5-1) Input the test samples into the trained DSVDD-MemAE model, extract its temporal dynamic features and spatial correlation features through the spatiotemporal feature decoupling module, and then obtain the corresponding statistics;

[0020] (5-2) Compare the statistic value with the control limit in step (4). If the statistic exceeds the control limit, the sample is considered abnormal; otherwise, it is considered normal.

[0021] As a preferred method, the process of inputting subsequence samples into the DSVDD-MemAE model is as follows:

[0022] (a) Input the subsequence samples into the encoder to obtain spatiotemporal hybrid features, which are then input into the temporal and spatial branches respectively to decouple the features;

[0023] (b) In the time branch, the features of adjacent time steps are first concatenated to transform instantaneous information into temporal dynamic information, and one-dimensional convolution is used to further extract features. Then, the extracted features are input into the memory module of the time branch to obtain the final temporal dynamic features.

[0024] (c) In the spatial branch, the fully connected neural network performs nonlinear transformation and feature extraction on different process variables within the same time step, and then inputs the extracted features into the memory module of the spatial branch to obtain the final spatial correlation features;

[0025] (d) The temporal dynamic features and spatial correlation features are fused and input into the decoder to obtain the reconstructed output;

[0026] (e) After outputting features from the memory modules of the temporal and spatial branches, the temporal dynamic features and spatial correlation features are compressed into a hypersphere to construct the feature space boundary for normal samples.

[0027] Preferably, the memory module of the temporal branch maintains a set of trainable temporal memory vectors, where each vector encodes a typical temporal dynamic feature contained in the training data. The module first calculates the similarity between the input feature and each temporal memory vector as the initial weight, then applies a hard contraction strategy to the weight and performs normalization processing, and finally re-represents the input feature as a weighted combination of all temporal memory vectors. The memory module of the spatial branch contains a set of trainable spatial memory vectors, where each vector is used to capture typical spatial correlation features in the training data. The module generates weights by calculating the similarity between the input feature and each spatial memory vector, and after hard contraction and normalization processing, reconstructs the input feature as a weighted combination of these spatial memory vectors.

[0028] As a preferred option, the hyperparameters of the DSVDD-MemAE model include the number of model layers, the number of hidden layer nodes, the batch size, the learning rate, the sliding window size, the total number of memory vectors in the memory module, and the hard contraction threshold; the network parameters of the DSVDD-MemAE model include all weight parameters and bias term parameters.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] (1) Based on the MemAE model, this invention introduces two memory modules, time branch and spatial branch, to store the memories of temporal dynamic relationships and spatial correlations respectively, thus achieving effective decoupling of spatiotemporal features. Compared with the limitations of the traditional MemAE model, which mixes spatiotemporal features and is prone to mutual interference between features and obscuring key information, this invention enables the two types of features to learn and memorize independently in their respective modules. This not only accurately captures the dynamic evolution law in the temporal dimension, but also clearly preserves the details of the interrelationship between variables in the spatial dimension, effectively avoiding modeling bias caused by feature mixing.

[0031] (2) By introducing DSVDD modules after the memory modules of the time and spatial branches, an independent, highly compact feature space can be constructed for each branch, thereby enhancing the distinguishability between normal and faulty samples. Simultaneously, this invention constructs a statistic by combining the reconstruction error with the DSVDD deviation error of the two branches, effectively fusing reconstruction information and feature deviation information to form a more comprehensive basis for anomaly judgment. This invention effectively overcomes the limitations of a single detection index; even with good reconstruction results, it can keenly identify abnormal states through the deviation of DSVDD in the feature space, significantly improving fault detection performance. Attached Figure Description

[0032] Figure 1 This is a flowchart of the industrial process fault detection method based on spatiotemporal feature decoupling of the present invention;

[0033] Figure 2 This is a schematic diagram of the structure of the DSVDD-MemAE model constructed in this invention;

[0034] Figure 3 This is a fault detection effect diagram of the present invention and several other comparison methods on the TE process dataset for fault 5. Figure 3 (a) in the image shows the fault detection effect of MemAE. Figure 3 (b) in the diagram shows the effect of IAEMM fault detection. Figure 3 (c) in the image shows the fault detection effect of STMemAE. Figure 3 (d) in the figure represents the fault detection effect of the present invention;

[0035] Figure 4 This is a fault detection effect diagram of the present invention and several other comparison methods on the TE process dataset for fault 10. Figure 4 (a) in the image shows the fault detection effect of MemAE. Figure 4 (b) in the diagram shows the effect of IAEMM fault detection. Figure 4 (c) in the image shows the fault detection effect of STMemAE. Figure 4 (d) in the figure is the fault detection effect diagram of the present invention. Detailed Implementation

[0036] The technical solution of the present invention will be further described in detail below through embodiments. These embodiments are illustrative of the present invention and not intended to limit the present invention. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0037] This invention addresses the problem of missed fault detection caused by excessive generalization and mixed spatiotemporal features in traditional self-encoders in industrial process fault detection, and proposes an industrial process fault detection method based on spatiotemporal feature decoupling.

[0038] Reference Figure 1 A method for detecting industrial process faults based on spatiotemporal feature decoupling includes the following steps:

[0039] (1) Raw data acquisition and preprocessing: Obtain the industrial process multivariate time series dataset (collect raw data), divide it into training set and test set. The training set contains only normal data, and the test set contains normal data and abnormal data. Perform Z-Score standardization on it (training data preprocessing, test data preprocessing) to eliminate the influence of different variables in terms of units and numerical ranges.

[0040] (2) Construction of DSVDD-MemAE model:

[0041] The training set data is serialized using a time window to construct the DSVDD-MemAE model, the hyperparameters of the model are configured, and the network parameters are initialized.

[0042] The structure of the DSVDD-MemAE model is as follows: Figure 2 As shown, it comprises five key modules: encoder, temporal branch, spatial branch, decoder, and DSVDD module. The encoder is responsible for extracting basic features from the input data (constructing a multi-layer fully connected coding network); the temporal and spatial branches contain independent memory modules (temporal branch memory module and spatial branch memory module), which store the memories of temporal dynamic relationships and spatial correlation relationships respectively, achieving effective decoupling of spatiotemporal features; the decoder is responsible for reconstructing the input using spatiotemporal features (the decoder reconstructs the data using fused features); after the memory modules of the temporal and spatial branches, DSVDD modules (temporal branch DSVDD module and spatial branch DSVDD module) are introduced respectively, constructing an independent, highly compact feature space for each branch, enhancing the distinguishability between normal samples and faulty samples.

[0043] (3) DSVDD-MemAE model training: By jointly optimizing the total loss function of MemAE and DSVDD, the trained DSVDD-MemAE model is obtained, which realizes the dual constraint on normal data features, thereby improving the subsequent fault detection performance.

[0044] (4) Construction of statistics:

[0045] The reconstruction error of the joint training samples and the deviation error of the DSVDD module in the spatiotemporal branch are used to calculate statistics, and the control limits of the statistics are determined.

[0046] (5) Fault detection: The test sample is input into the DSVDD-MemAE model trained in step (3). Its temporal dynamic features and spatial correlation features are extracted by the spatiotemporal feature decoupling module. Then, the statistics are calculated and compared with the control limit in step (4) to determine whether the control limit is exceeded, thereby realizing industrial process fault detection. At the same time, the fault detection rate is calculated to evaluate the performance of the model in fault detection.

[0047] Specifically:

[0048] The specific process of step (1) is as follows:

[0049] First, a multivariate time series dataset of an industrial process is obtained and divided into training and test sets. The training set contains only normal data, while the test set contains both normal and outlier data. Z-score standardization is then applied to eliminate the influence of different units and numerical ranges among the variables. The standardized training and test sets are as follows: and Where N is the number of variables in the training and test data. tr N is the sequence length of the training set. ts This indicates the sequence length of the test set.

[0050] Furthermore, the specific process of step (2) is as follows:

[0051] Using a sliding window of length T, along X train Sampling is performed along the sequence direction to divide the complete time series into a series of subsequence samples. , where x t-T+1 For the sample at time t-T+1, x t-T+2 For the sample at time t-T+2, x t This is the sample at time t. Subsequently, the subsequence samples... Input the DSVDD-MemAE model. When processing subsequence samples in the DSVDD-MemAE model, it sequentially goes through five core modules: encoder, temporal branch, spatial branch, decoder, and the DSVDD module. The specific process is as follows:

[0052] (a) Encoder

[0053] First, the subsequence samples are input into the encoder to obtain the spatiotemporal hybrid features. ,in, This represents the feature dimension of the encoder output. For the first Time sample The corresponding spatiotemporal hybrid characteristics, For the first Time sample The corresponding spatiotemporal hybrid characteristics, For the first Time sample The corresponding spatiotemporal hybrid features are then input into the temporal and spatial branches respectively to decouple the features.

[0054] (b) Time branch

[0055] In the time branch, features from adjacent time steps are first concatenated to transform instantaneous information into temporal dynamic information. Then, one-dimensional convolution is used to further extract features. The specific calculation formula is as follows:

[0056]

[0057] in, For one-dimensional convolution, The features representing adjacent time steps are calculated using the following formula:

[0058]

[0059] Where lag is the size of the time step. The first output of the encoder Time sample The corresponding spatiotemporal hybrid characteristics, The first output of the encoder Time sample The corresponding spatiotemporal hybridity characteristics.

[0060] The time branch ends with a memory module, which enhances the model's ability to distinguish between normal and abnormal patterns by constraining overgeneralization. The core idea is to maintain a set of learnable memory matrices in memory. , where each vector It encodes typical feature patterns from the training data. For a given input The module first calculates the similarity between the input and each memory vector as a weight, and then re-represents the input as a weighted combination of these memory vectors to obtain the final temporal dynamic features. The specific calculation formula is as follows:

[0061]

[0062] in, This is the weight coefficient matrix for the time branch. The memory matrix for time branches, The number of memory vectors in the memory matrix, and the weight coefficient vector of the time branch. From the input vector With each memory vector of the time branch cosine similarity Confirmed, the specific calculation formula is as follows:

[0063]

[0064]

[0065] Furthermore, despite the introduction of the memory module, some faulty samples may still be reconstructed with high quality by the decoder through a complex weighted combination of memory vectors, leading to faulty samples being misclassified as normal. To address this issue, a hard contraction strategy is required, with the specific calculation formula as follows:

[0066]

[0067] in, This is a threshold value; the attention weight is only effective when it is greater than this value. Otherwise, the attention weight is 0. These are positive numbers with very small values. After hard shrinkage, the weights are normalized at the end.

[0068] (c) Spatial branching

[0069] In the spatial branch, the fully connected neural network performs nonlinear transformations and preliminary feature extraction on different process variables within the same time step. The extraction and specific calculation formula are as follows:

[0070]

[0071] The spatial branch also employs a memory module at its end, with the core principle being consistent with the temporal branch. This module utilizes a set of trainable memory vectors to encode typical spatial feature patterns and represents the input features as a weighted combination of these memory vectors, yielding the final spatial correlation features. The specific formula is as follows:

[0072]

[0073] in, This is the weight coefficient matrix for the spatial branch. The memory matrix for the spatial branch, and the weight coefficient vector for the spatial branch. From the input vector With each memory vector of the spatial branch cosine similarity Confirmed, the specific calculation formula is as follows:

[0074]

[0075]

[0076] Furthermore, similar to the memory modules in the temporal branch, the spatial branch also employs a hard contraction strategy, with the specific calculation formula as follows:

[0077]

[0078] in, For the threshold, These are positive numbers with very small values. After hard shrinkage, the weights are normalized at the end.

[0079] (d) Decoder

[0080] Sample at time t Corresponding temporal dynamic characteristics Spatial association features The components are fused, input into the decoder, and the reconstructed output is obtained. Therefore, the original input sequence The corresponding decoder output is ,in, for The reconstructed output, for The reconstructed output.

[0081] (e) DSVDD module

[0082] To enhance the distinction between normal and abnormal samples and suppress false negatives caused by MemAE's excessive generalization ability, a DSVDD module was introduced after the memory module in both the temporal and spatial branches. This module's function is to convert the output features from the memory modules of the temporal and spatial branches into temporal dynamic features. Spatial association features Each sample is compressed into a compact hypersphere, thus constructing a well-defined and compact feature space boundary for normal samples.

[0083] The loss function for the time-branched DSVDD module is:

[0084]

[0085] The loss function for the spatial branch DSVDD module is:

[0086]

[0087] in, For time-series dynamic characteristics, Spatial correlation features, The center of the hypersphere in the time-branch feature space. The center of the hypersphere in the spatial branch feature space. This is the weight matrix of the l-th layer of the neural network in the time branch. Let L be the weight matrix of the l-th layer of the neural network in the spatial branch. The regularization coefficient for the time branch is . is the regularization coefficient for spatial branching. The number of network layers for time branches. The number of network layers for spatial branching.

[0088] Based on the aforementioned DSVDD-MemAE model, core hyperparameters such as the number of model layers, the number of hidden layer nodes, the batch size, the learning rate, the sliding window size, the total number of memory vectors in the memory module, and the hard shrinkage threshold are set. All weight parameters and bias terms are initialized, thus completing the model construction and laying the foundation for subsequent model training.

[0089] Furthermore, the specific process of step (3) is as follows:

[0090] The DSVDD-MemAE model achieves dual constraints on normal data features by jointly optimizing the total loss function of MemAE and DSVDD, thereby improving the performance of process fault detection. The specific formula is as follows:

[0091]

[0092] in, , , These are MemAE error, time branch DSVDD error, and spatial branch DSVDD error, respectively. , for and The weight. It consists of the mean squared error and the sparse regularization loss of the attention weights, and the specific formula is as follows:

[0093]

[0094] in, The entropy of the weight coefficients of the time branch memory module. The entropy of the weight coefficients of the spatial branch memory module. and express and Weights in the loss function.

[0095] Furthermore, the specific process of step (4) is as follows:

[0096] In order to achieve real-time monitoring of the operating status, after training the DSVDD-MemAE model using the loss function in step (3), the following statistics are constructed for fault detection:

[0097]

[0098] This statistic combines the reconstruction error of the training samples with the deviation error of the DSVDD module in the spatiotemporal branch. Compared with methods that rely solely on the MemAE reconstruction error to distinguish between normal and abnormal samples, it can more sensitively capture abnormal information in temporal dynamics or spatial correlation features, thus providing richer and more comprehensive judgment criteria.

[0099] Then, based on the calculation method of the statistic, the statistic values ​​of all normal samples in the training set are calculated, and the upper quantile of the 95% confidence level is taken as the final control limit.

[0100] Furthermore, the specific process of step (5) is as follows:

[0101] The test sample is input into the trained DSVDD-MemAE model, and its temporal dynamic features and spatial correlation features are extracted by the spatiotemporal feature decoupling module, and then the corresponding statistics are calculated. The statistical value is compared with the control limit in step (4). If the statistical value exceeds the control limit, the sample is judged to be abnormal; otherwise, it is considered normal. Based on this discrimination mechanism, real-time fault detection and abnormal alarm of industrial process can be realized. At the same time, the performance of the method of the present invention is evaluated using the Fault Detection Rate (FDR), and its formula is as follows:

[0102]

[0103] TP stands for True Positive, which is the number of samples that are actually faulty but are correctly detected as faulty, and FN stands for False Negative, which is the number of samples that are actually faulty but are considered normal.

[0104] This invention was applied to the Tennessee-Eastman (TE) process to verify its fault detection performance. The TE process has 52 variables, including 11 manipulated variables and 41 measured variables. The dataset contains 21 typical faults, and the specific causes and types of faults are shown in Table 1. In the test set, each fault type contains 960 samples, with the fault introduced after the 161st sample. In the DSVDD-MemAE model, the Adam optimizer was used for model optimization, with a learning rate of 0.0001, a batch size of 32, a sliding window size of 40, a total number of memory vectors in the spatiotemporal branch memory module of 100, and a hard contraction threshold of 0.0125.

[0105] Table 1 Fault Types in TE Dataset

[0106]

[0107] To test the fault detection performance of this invention in industrial processes, the DSVDD-MemAE model of this invention was compared with the MemAE model, the IAEMM model, and the STMemAE model. The MemAE model is the basic model; the IAEMM model combines MemAE and DSVDD but does not decouple the spatiotemporal features; while the STMemAE model achieves decoupling of spatiotemporal features but does not introduce a DSVDD module. Figure 3 The above model demonstrates its fault detection performance on fault 5 of the TE process dataset. Figure 4 The above model's fault detection performance on the TE process dataset, fault 10, is demonstrated. From... Figure 3 , Figure 4 As can be seen, this invention constructs a comprehensive fault detection statistic through the collaborative optimization of a dual-memory module with spatiotemporal feature decoupling and a dual-branch DSVDD, thereby achieving a significant improvement in fault detection performance.

[0108] To further quantify and compare the detection performance of different models, Table 2 shows the fault detection rate (FDR) of each model on different faults in the TE process dataset.

[0109] Table 2 Comparison of FDR fault detection performance on the TE process dataset

[0110]

[0111] As can be seen from the data in Table 2, the present invention exhibits a higher detection rate in the vast majority of faults, and significantly outperforms all comparative models with an average fault detection rate of 87.583%. This verifies that the dual memory modules of time branch and spatial branch in the present invention can promote the effective decoupling of spatiotemporal features. Furthermore, the introduction of DSVDD modules after the dual branches helps to expand the distinction between normal and abnormal samples. Together, these factors contribute to the accurate and reliable fault detection effect.

[0112] This invention constructs temporal and spatial branches after the MemAE encoder, and configures independent memory modules for the two branch structures to store the memories of temporal dynamic relationships and spatial correlation relationships, respectively, achieving effective decoupling of spatiotemporal features. DSVDD modules are introduced after the memory modules of the temporal and spatial branches, constructing an independent, highly compact feature space for each branch, enhancing the distinguishability between normal and faulty samples. Simultaneously, a fault detection statistic is designed based on the DSVDD-MemAE model, combining the reconstruction error with the DSVDD deviation error of the spatiotemporal branches, making the judgment of abnormal data more comprehensive. This invention effectively decouples spatiotemporal features, expands the distinguishability between normal and abnormal samples, and effectively improves the fault detection rate and the reliability of industrial process monitoring.

[0113] Finally, it should be noted that the above examples are merely some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the disclosure of this invention should be considered within the scope of protection of this invention.

Claims

1. A method for detecting industrial process faults based on spatiotemporal feature decoupling, characterized in that... Includes the following steps: (1) Obtain the multivariate time series dataset of industrial process, divide the dataset into training set and test set, and standardize the training set and test set respectively; (2) Use time windows to serialize the training set data, construct the DSVDD-MemAE model, configure the hyperparameters of the model, and initialize its network parameters; (3) Optimize the model loss function to obtain the completed DSVDD-MemAE model, thereby decoupling the temporal dynamic features and spatial correlation features of the training data; (4) Based on the reconstruction error of the training samples and the deviation error of the DSVDD module in the spatiotemporal branch, obtain the statistics of the training samples and determine the control limits of the statistics. (5) Input the test sample into the DSVDD-MemAE model trained in step (3), and after decoupling the spatiotemporal features, obtain its corresponding statistics and compare them with the control limits to determine whether the sample is normal or abnormal, thereby realizing industrial process fault detection. The specific process of step (2) is as follows: using a sliding window, sampling is performed along the sequence direction of the training set to divide the complete time series into a series of subsequence samples, and the subsequence samples are input into the DSVDD-MemAE model; the DSVDD-MemAE model contains five key modules: encoder, temporal branch, spatial branch, decoder and DSVDD module; The process of inputting subsequence samples into the DSVDD-MemAE model is as follows: (a) Subsequence samples are input into the encoder to obtain spatiotemporal hybrid features, which are then input into the temporal and spatial branches respectively to decouple the features; (b) In the time branch, the features of adjacent time steps are first concatenated to transform instantaneous information into temporal dynamic information, and one-dimensional convolution is used to further extract features. Then, the extracted features are input into the memory module of the time branch to obtain the final temporal dynamic features. (c) In the spatial branch, the fully connected neural network performs nonlinear transformation and feature extraction on different process variables within the same time step, and then inputs the extracted features into the memory module of the spatial branch to obtain the final spatial correlation features; (d) The temporal dynamic features and spatial correlation features are fused and input into the decoder to obtain the reconstructed output; (e) After outputting features from the memory modules of the temporal and spatial branches, the temporal dynamic features and spatial correlation features are compressed into a hypersphere to construct the feature space boundary for normal samples.

2. The industrial process fault detection method based on spatiotemporal feature decoupling according to claim 1, characterized in that, The specific process of step (1) is as follows: obtain the industrial process multivariate time series dataset, divide it into training set and test set, the training set contains only normal data, the test set contains normal data and abnormal data, and perform Z-Score standardization on it.

3. The industrial process fault detection method based on spatiotemporal feature decoupling according to claim 1, characterized in that, The specific process of step (3) is as follows: The DSVDD-MemAE model achieves dual constraints on normal data features by jointly optimizing the total loss function of MemAE and DSVDD, thereby improving the performance of process fault detection.

4. The industrial process fault detection method based on spatiotemporal feature decoupling according to claim 1, characterized in that, The specific process of step (4) is as follows: After training the DSVDD-MemAE model using the loss function in step (3), firstly, a statistic is constructed for fault detection. This statistic combines the reconstruction error of the training samples with the deviation error of the DSVDD module in the spatiotemporal branch. Then, based on this statistic, the statistic values ​​of all normal samples in the training set are obtained, and the upper quantile of its 95% confidence level is taken as the final control limit.

5. The industrial process fault detection method based on spatiotemporal feature decoupling according to claim 1, characterized in that, The specific process of step (5) is as follows: (5-1) Input the test samples into the trained DSVDD-MemAE model, extract its temporal dynamic features and spatial correlation features through the spatiotemporal feature decoupling module, and then obtain the corresponding statistics; (5-2) Compare the statistic value with the control limit in step (4). If the statistic exceeds the control limit, the sample is considered abnormal; otherwise, it is considered normal.

6. The industrial process fault detection method based on spatiotemporal feature decoupling according to claim 1, characterized in that, The temporal and spatial branches each have a corresponding memory module at their ends. The memory module of the temporal branch maintains a set of trainable temporal memory vectors, where each vector encodes a typical temporal dynamic feature contained in the training data. This module first obtains the similarity between the input feature and each temporal memory vector as the initial weight, then applies a hard contraction strategy to the weight and performs normalization processing, finally re-representing the input feature as a weighted combination of all temporal memory vectors. The memory module of the spatial branch contains a set of trainable spatial memory vectors, each vector used to capture typical spatial correlation features in the training data. This module generates weights by obtaining the similarity between the input feature and each spatial memory vector, and after hard contraction and normalization processing, reconstructs the input feature as a weighted combination of these spatial memory vectors.

7. The industrial process fault detection method based on spatiotemporal feature decoupling according to claim 1, characterized in that, The hyperparameters of the DSVDD-MemAE model include the number of model layers, the number of hidden layer nodes, the batch size, the learning rate, the sliding window size, the total number of memory vectors in the memory module, and the hard shrinkage threshold; the network parameters of the DSVDD-MemAE model include all weight parameters and bias term parameters.

Images