Complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder

Abstract

The present invention relates to the technical field of industrial process fault detection and specifically discloses a complex industrial process fault detection method based on a spatiotemporal variational graph attention autoencoder. The method includes the following steps: Step 1: Temporal feature extraction based on slow feature analysis: Slow feature extraction is performed on measured data to obtain m slow features; Step 2: Spatial graph topology construction based on metric learning and prior knowledge: A graph topology is constructed for the obtained slow features; Step 3: Spatiotemporal feature representation learning based on a variational graph attention autoencoder: The variational graph attention autoencoder can process the temporal features of Step 1 and the spatial structure of Step 2 to generate a spatiotemporal representation of the latent features; Step 4: Training is performed by reconstructing the loss of the temporal features, spatial relationships, and the distribution of the latent representations; Step 5: Based on threshold calculation, the trained loss is input into kernel density estimation to obtain a fault detection threshold. The present invention can process spatiotemporal information of industrial processes, improve the fault detection rate, and reduce the false alarm rate.

Description

Technical Field

[0001] The present application relates to the technical field of industrial process fault detection, and specifically discloses a complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder. Background Art

[0002] With the advancement of modern industrial technology, industrial processes have become highly complex, leading to increasingly complex system structures. Such complex industrial processes often include multiple interconnected units that are tightly coupled and interdependent, resulting in high maintenance costs. The failure of one unit can trigger abnormalities in other units or even the collapse of the entire industrial process, resulting in huge economic losses. Therefore, it is imperative to effectively improve the fault detection performance of modern industrial systems.

[0003] Modern industrial plants have installed a large number of sensors in each operating unit to collect and store a large amount of data reflecting the operating status to ensure system safety and product quality. Therefore, data-driven methods are the preferred method for complex industrial process fault detection. It can be divided into two categories, multivariate statistical analysis and advanced artificial intelligence methods. In Example [1], researchers designed a hybrid first principle and data-driven process monitoring method, whose core algorithm is principal component analysis (PCA), which is a classic multivariate statistical analysis method. Specifically, in response to the problem that there are always large differences and imbalanced data in different working conditions, the researchers first designed an integrated principal component analysis and common subspace learning (PCA-CSL) method. Among them, a PCA algorithm was established for working conditions with sufficient data. On the contrary, for working conditions with insufficient data, a CSL algorithm was proposed, which uses working conditions with sufficient data to assist in modeling working conditions with insufficient data in the common subspace, thereby achieving accurate fault detection. Finally, based on the integrated first principle and data-driven method, a working condition decision rule library was established, and the decision rule parameters were optimized using the particle swarm optimization method to achieve abnormal state detection based on rule reasoning;

[0004] In example [2], researchers designed a full-graph autoencoder for fault detection of large-scale industrial IoT systems. The core technology of the designed model is graph autoencoder (GAE), which is an advanced artificial intelligence method. Specifically, the researchers used normal input data as input during training, and only included a few normalized graph convolution layers, which is simple and fast and can handle various irregular structures at the same time. In order for GAE to learn representations, multivariate time series are converted into graph data fused with prior knowledge. Then, the complete graph data is reconstructed to obtain reliable fault detection scores, and graph enhancement operations are used to improve accuracy and robustness;

[0005] The above defects:

[0006] The estimation method given in Example [1] has certain limitations. The method does not consider the spatial dependencies between multivariate time series data. In fact, there are spatial interactions of matter and energy between process units, which follow the laws of physical and chemical reactions and are affected by the topological structure of process connections. Due to this interaction, the failure of one unit will inevitably propagate to other units, triggering cascading failures. This means that a failure can be caused by the simultaneous superposition of multiple interrelated unit variables. Multiple interrelated process variables evolve together to maintain collaborative behavior. This interactive collaborative relationship contains process dynamics, topological structure and energy-quality coupling information, and can be regarded as a spatial relationship between multivariate time series variables. Therefore, it is crucial to mine the spatial interactions between multivariate time series data in complex industrial processes and fuse fault detection information;

[0007] The method given in Example [2] considers the spatial relationship of industrial process fault detection, but still lacks in-depth mining of temporal information features. In addition, the spatial information is not considered comprehensively. It only relies on the production data collected by the process itself, rather than multiple interrelated unit processes, which makes it very difficult to apply in practice.

[0008] Example [1] Source: Liang H, Yang C, Huang K, et al. A hybrid first principles and data-driven process monitoring method for zinc smelting roasting process [J]. IEEE Transactions on Instrumentation and Measurement, 2021,70:1-14.

[0009] Example [2] Source: Feng Y, Chen J, Liu Z, et al. Full Graph Autoencoder for One-Class Group Anomaly Detection of IIoT System [J]. IEEE Internet of ThingsJournal, 2022, 9(21): 21886-21898.

[0010] In general, it is very important to consider both temporal and spatial characteristics. In terms of time, the interaction mechanism between related unit processes is unclear, and there are multiple decision variables and dynamic correlations between them. As the material state, process mechanism and operating condition parameters change, the process variable changes show complex dynamic time-varying characteristics. In space, multiple interconnected unit processes are distributed in series or parallel at different locations according to a specific process connection structure, and interact through energy-mass flow. These interactions show structured characteristics and are easily affected by false correlations. It is very difficult to construct a complete process coupling model by simply linearly superimposing the mechanism models of each process, which brings challenges to process coupling modeling. In view of this, the inventors proposed a complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder. Summary of the Invention

[0011] The present invention provides a complex industrial process fault detection method based on a spatiotemporal variational graph attention autoencoder, which overcomes the limitation of existing methods that only extract features from the time or space of their own unit processes for fault detection. This invention can process multivariate time data coupled with multiple interrelated unit process structures and can learn to represent spatiotemporal interactive dependencies, thereby providing a reliable fault detection rate and significantly reducing the false alarm rate.

[0012] In order to achieve the above object, the present invention provides the following basic scheme:

[0013] The complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder includes the following steps:

[0014] S01: Temporal feature extraction based on slow feature analysis (SFA) extracts the slow component that carries important information and finds a conversion function that makes the feature S(t) = g(X(t)) the slowest. After the sensor data passes through SFA, m SFs can be obtained.

[0015] S02: Spatial graph topology construction based on metric learning and prior knowledge, constructing graph topology for the obtained SFA based on metric learning and prior knowledge;

[0016] S03: Learning spatiotemporal feature representations based on variational graph attention autoencoders: The variational graph attention autoencoder combines the graph attention autoencoder (GATE) and the variational graph autoencoder (VGAE), replacing the original GCN encoder of VGAE with the graph attention network (GAT) encoder of GATE to generate high-quality latent spatiotemporal feature representations;

[0017] S04: Comprehensive loss function training: Calculate the loss of the reconstructed graph spatial structure and the original graph spatial structure, the reconstructed graph node features and the original graph node features and the distribution of potential representations, and then optimize. The optimized training objects include node features and graph topology structure.

[0018] S05: Threshold calculation based on kernel density estimation. For the comprehensive training of the loss function in S04, the fault detection threshold is specifically determined by the distribution of the trained reconstruction loss function L. A non-parametric estimation method, kernel density estimation, is used to calculate the threshold of L.

[0019] The principles and effects of this basic solution are:

[0020] 1. Compared to existing methods, this method proposes an unsupervised framework for complex industrial process fault detection based on a spatiotemporal variational graph attention autoencoder, addressing the fact that existing methods fail to fully consider the temporal and spatial aspects of industrial processes. By extracting temporal dynamics, constructing spatial interaction graph topology, and learning and training integrated spatiotemporal feature representations, this method enables accurate fault detection for multiple interrelated unit processes. The detection threshold integrates temporal feature reconstruction, spatial relationship reconstruction, and the distribution loss of latent representation vectors. Consequently, this method can meticulously reflect the operational status of industrial processes and perform accurate fault detection.

[0021] 2. Compared to existing technologies, this paper addresses the difficulty of converting real-world industrial structure knowledge and measurement data into graph data. This paper proposes a graph topology construction paradigm based on data features and prior knowledge. This approach, combining metric learning with prior knowledge, effectively converts the interactive information of time series data into graph-structured data. This ensures that both temporal and spatial information are effectively preserved.

[0022] 3. Compared with the existing technology, the training process of the present invention only requires normal sample data and can handle practical problems where unlabeled data is available.

[0023] 4. This method achieves efficient fault detection for multiple interconnected unit processes through temporal dynamic feature extraction, spatial interaction information transformation, and integrated spatiotemporal feature learning. First, slow feature analysis (SFA) is used to extract the underlying driving forces that cause process dynamic time variation, namely slowly varying variables, from time series data. These extracted slow features represent the fundamental dynamic trends of system operation. Then, metric learning is combined with prior knowledge to construct a graph topology based on slow features and interactive structural information. This combines temporal features with entity dependencies. Based on this, a novel variational graph attention autoencoder (VGATE) is proposed, which can automatically represent temporal and spatial structural feature information with high quality. Subsequently, a comprehensive training loss function is designed that integrates the reconstruction, edge interaction relationships, and latent representations of all time-varying nodes. Finally, a robust fault threshold is derived from the training loss function using kernel density estimation.

[0024] Furthermore, in step S01, the linear SFA learning paradigm is used to learn slow changes, which can be expressed as s(t)=Wx(t), where W is the weight matrix. The input measurement data can be solved, and the solution can obtain m slow features (SFs) arranged from large to small according to the slowness. The slowly changing features in (SFs) are the features that can properly represent the essential characteristics of the data, and the fast-changing features are regarded as noise signals, and the fastest-changing features are removed.

[0025] Furthermore, in step S02, each SFs variable is defined as a node, and the metric similarity between node i and node j can be expressed as A ij =M(H i ,H j ), using the metric method defined on the vector space, a complete norm vector space D is defined Chebyshev (H i ,H j )=max n (|H i -H j |).

[0026] Furthermore, by combining the metric similarity between nodes i and j and learning the spatial structure through the threshold parameter k, the spatial structure matrix information can be expressed as: Then, the adjacency matrix A′ can be constructed based on prior knowledge ij , where if it is determined that node i and node j have an inherent connection relationship, then A′ ij =1, otherwise, A′ ij =0, similarity matrix of comprehensive metric learning and prior knowledge matrix A′ ij, we can get the final spatial structure E: E contains feature information and prior knowledge information. Based on E, graph structure data G(V,E) can be constructed, where V is a slow feature variable and E is the structural information between features.

[0027] Further, in step S03, the graph attention encoder is in VGATE, and the industrial process variable node S after SFA processing i In the application of k th The output of a GAT encoding layer can be expressed as: S represents the set of all industrial nodes, which can be obtained by formula (6), E i The neighborhood spatial relationship set of node i is obtained by (10), Represents the original feature S of the industrial variable node i (t),W (k) represents the weight matrix of the hidden layer of the k-th encoding layer feature map. σ(·) represents the activation function. is the kth th The correlation attention coefficient between neighbor node j and node j.

[0028] Further, the correlation attention coefficient: represents the correlation coefficient between node i and its neighbor node j.

[0029] further, in:

[0030] It is k th The trainable parameters of the encoder layer, θ(·) represents the activation function, and σ(·) represents the activation function of the kth encoder layer.

[0031] Furthermore, in step S05, through comprehensive training of the loss function, the spatiotemporal relationship between multiple interrelated unit processes is effectively learned and trained. The loss function value is an ideal detection indicator, and the fault detection threshold is specifically determined by the distribution of the reconstructed loss function L, which is trained on normal samples.

[0032] Furthermore, a non-parametric estimation method, kernel density estimation, is used to calculate the threshold of L. The training vector li,i=1,2,...,n is obtained. The kernel density estimation at point l is defined as: Where K(·) is the kernel function and h is the bandwidth. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present application, 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 application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0034] Figure 1 A schematic diagram of the modeling strategy of the complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder proposed in an embodiment of the present application is shown;

[0035] Figure 2 A schematic diagram of the roasting process in the complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder proposed in an embodiment of the present application is shown;

[0036] Figure 3 A schematic diagram of slow features of 400 test samples in the complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder proposed in an embodiment of the present application is shown;

[0037] Figure 4 The adjacency matrix heat map of the spatial graph topology in the complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder proposed in an embodiment of the present application is shown;

[0038] Figure 5 A schematic diagram of the spatiotemporal features of the test samples extracted by the model in the complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder proposed in an embodiment of the present application is shown;

[0039] Figure 6 A comparison diagram of online tests of a complex industrial process fault detection method based on a spatiotemporal variational graph attention autoencoder proposed in an embodiment of the present application is shown;

[0040] Figure 7 A schematic diagram of features extracted from online testing of various methods in the complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder proposed in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0041] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0042] Implementation example Figure 1-Figure 7 As shown:

[0043] Complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder,

[0044] The modeling process of the present invention is shown in the attached Figure 1 It is shown that it consists of five parts: temporal feature extraction based on slow feature analysis, spatial graph topology construction based on metric learning and prior knowledge, spatiotemporal feature representation learning based on variational graph attention autoencoder, comprehensive training of loss function, and threshold calculation based on kernel density estimation.

[0045] Step 1: Temporal feature extraction based on slow feature analysis.

[0046] Slow feature analysis is an unsupervised time domain feature extraction algorithm. Its purpose is to extract the slow component that carries important information from time-varying data, thereby eliminating noise information. For the information X(t) = [x1(t), x2(t), ..., x m (t)] T , the goal of slow feature analysis is to find a conversion function Make the characteristic S(t)=g(X(t)) change the slowest, as shown in the following formula.

[0047]

[0048] Constraints

[0049] i >t=0 (2)

[0050]

[0051]

[0052] in, is the first derivative of the slow feature, <·> t is the average value of s over time, which can be calculated as follows:

[0053]

[0054] There are different solutions to the above optimization problem. In order to maintain generality, the present invention adopts the linear SFA learning paradigm to learn slow changes, which can be expressed as:

[0055] s(t)=Wx(t) (6)

[0056] ​Where W is a weight matrix. The input measurement data can be solved to obtain m slow features (SFs), ranked from most to least slow. It is generally believed that slowly varying features in SFs are the ones that accurately represent the essential characteristics of the data. Fast-changing features are considered noise signals, so this invention removes the fastest-changing features. Note that this is not limited to removing the fastest-changing features; the number of fast-changing features to be removed depends on the specific characteristics of the industrial process, and this invention is merely an example.

[0057] Step 2: Spatial graph topology construction based on metric learning and prior knowledge:

[0058] This paper constructs graph topology based on metric learning and prior knowledge. Metrics are a commonly used method for learning spatial structures. To ensure the accuracy of structural information, it is known that after the sensor data undergoes SFA, m SFs can be obtained. Therefore, we can define each SF variable as a node, and the metric similarity between node i and node j can be expressed as:

[0059] A ij =M(H i ,H j ) (7)

[0060] Among them, H i and H j Denote nodes i and j, respectively. The feature representation of a node is SFs. M(·) represents the metric learning method. This paper uses, as an example, the Chebyshev distance, a metric defined on a vector space. It defines a complete normed vector space, known in mathematics as a Banach space, as shown below.

[0061] D Chebyshev (H i ,H j )=max n (|H i -H j |)(8)

[0062] Where n represents the number of samples. If, D Chebyshev (H i ,H j ) is closer to 0, indicating that the similarity between the two vectors is higher. Then, the spatial structure is learned by the threshold parameter k. The spatial structure matrix information can be expressed as:

[0063]

[0064] Then, the adjacency matrix A′ can be constructed based on prior knowledge ij , where if it is determined that node i and node j have an inherent connection relationship, then A′ ij =1, otherwise, A′ij = 0. Similarity matrix for comprehensive metric learning and prior knowledge matrix A′ ij , we can get the final spatial structure E:

[0065]

[0066] E contains feature information and prior knowledge. Based on E, we can construct graph structure data G(V,E), where V is the slow feature variable and E is the structural information between features.

[0067] Step 3: Spatiotemporal feature representation learning based on variational graph attention autoencoder:

[0068] The variational graph attention autoencoder designed by the invention combines the algorithmic advantages of the graph attention autoencoder (GATE) and the variational autoencoder (VGAE). Specifically, the original GCN encoder of VGAE is replaced with the graph attention network (GAT) encoder of GATE to generate high-quality spatial representations of latent features. The GATE and VGAE algorithms show complementary characteristics in industrial fault detection. The graph attention mechanism of the GATE encoder and decoder is conducive to the extraction of interactive perception features and improves the overall performance of the VGATE model. In addition, the variational autoencoder strategy of VGAE helps to improve the quality of the representation of the spatiotemporal latent feature distribution. The variational graph attention autoencoder consists of three basic components: a graph attention encoder, a variational inference process, and a decoder.

[0069] (1) Graph Attention Encoder:

[0070] In VGATE, the industrial process variable node S after SFA processing i In the application of k th The output of a GAT encoding layer can be expressed as:

[0071]

[0072] Where S represents the set of all industrial nodes, which can be obtained by formula (6). i The neighborhood spatial relationship set of node i is obtained by (10), Represents the original feature S of the industrial variable node i (t),W (k) represents the weight matrix of the hidden layer of the k-th encoding layer feature map. σ(·) represents the activation function. is the kth th The correlation attention coefficient between neighbor node j and node j is obtained by the following formula.

[0073]

[0074] in, It represents the correlation coefficient between node i and its neighbor node j, which can be calculated by the following formula.

[0075]

[0076] in, are the trainable parameters of the kth encoder layer. θ(·) represents the activation function, and σ(·) represents the activation function of the kth encoder layer.

[0077] (2) Variational inference process

[0078] Based on the variational idea, the above GAT coding layer is used to fit μ and σ, as shown in the following formula.

[0079] μ=GAT μ (S (k) ,E) (14)

[0080] logσ=GAT σ (S (k) ,E) (15)

[0081] Where μ is the mean vector matrix and σ is the variance matrix. logσ and μ share the weight W. S (k) is k after (11) th GAT encoding is obtained. Subsequently, the spatiotemporal hidden vector representation can be obtained.

[0082] Z=q(S|N(μ , σ 2 )) (16)

[0083] Where Z is the potential vector, and its posterior distribution P(Z|S,E) is the variational posterior distribution q(z i |S,E) approximation. q(z i |S,E) is set to a Gaussian distribution, and its diagonal covariance matrix is ​​shown below.

[0084]

[0085]

[0086] (3) Decoding

[0087] The purpose of the decoder is to reconstruct the features S and spatial relations E.

[0088]

[0089]

[0090] Among them, GAT (k) Indicates that the number of encoding layers is consistent with the number of decoding layers. It is the inner product of any two node representation vectors to satisfy the adjacency relationship, thereby reconstructing whether any two points in the graph are connected:

[0091]

[0092]

[0093] Step 4: Comprehensive training of loss function:

[0094] For a well-trained STVGATE model, the optimized training objects include node features and graph topology. In order to enable STVGATE to learn high-quality representations of both, the loss function is designed as follows.

[0095]

[0096] Among them, MSE is the mean square error loss function, n is the sample size. E q(Z∣S,E) [logp(E|Z)] represents the difference between the spatial structure of the reconstructed graph and the original graph, which is minimized using the binary cross entropy loss. KL[q(·)‖p(·)] is the Kullback-Leibler divergence between q(·) and p(·).

[0097] Step 5: Threshold calculation based on kernel density estimation:

[0098] Through comprehensive training of the loss function, the spatiotemporal relationship between multiple interrelated unit processes is effectively learned and trained. The loss function value is an ideal detection indicator. The fault detection threshold is specifically determined by the distribution of the reconstruction loss function L, which is trained on normal samples. In this study, a non-parametric estimation method, kernel density estimation, is used to calculate the threshold of L. Let the training result l i ,i=1,2,...,n vector, the kernel density estimate at point l is defined as:

[0099]

[0100] Where K(·) is the kernel function and h is the bandwidth. This study uses the radial basis function, and the estimated probability density function can be expressed as:

[0101]

[0102] Then, given the confidence level α, the detection limit d is expressed as:

[0103]

[0104] To make the purpose, technical solutions and advantages of the present invention clearer, the following is a Figure 2 The preferred embodiments provide a clear and complete description of the technical solutions of the present invention. It should be noted that the embodiments described are only a portion of the present invention and are intended only to illustrate the present invention and are not intended to limit the present invention. Furthermore, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion regarding the present invention.

[0105] Example 1: Take the roasting process of a smelter as an example, as shown in the attached Figure 2 As shown in Figure 1, the roasting process is a typical multi-unit process consisting of five distinct units: a feed unit, a blast unit, a furnace unit, a flue gas treatment unit, and a roasting sand treatment unit. The primary objective of the roasting process is to increase the soluble zinc content in the roasted product. Specifically, in the roaster, zinc concentrate (ZnS) fed by the feeder and air (O2) blown by the blower form a fluidized bed where they are thoroughly mixed and undergo a vigorous oxidation reaction. The resulting zinc roasted sand (ZnO) is transported from the roasting treatment unit to the next process, while the flue gas (SO2) is fed to the waste heat recovery and acid production processes. In this example, 25 variables were measured from the plant. A total of 4,000 normal state samples were used for training, and 200 normal and 200 faulty samples were used for testing.

[0106] Step 1: First, for the current time t, collect 25 process variables X(t) = [x1(t), x2(t), ..., x m (t)] T , 4000 normal samples are used for training, and 400 test samples (including 200 normal samples and 200 fault samples) are used for online testing. All training processes of this method only require normal samples. First, based on slow feature analysis, the time feature of the process variable is extracted to calculate the following formula W.

[0107] s(t)=Wx(t)

[0108] Computer solution process (brief description):

[0109] f(x)=W(x+b)

[0110] Get W and b by inputting sample x

[0111]

[0112]

[0113] in, is an orthogonal matrix, D is the eigenvalue diagonal matrix, and U is the feature matrix. After obtaining w and b, W is also calculated at the same time. Then, the time slow features of the test sample are extracted through W, and 25 slow features are obtained. The slow features with the fastest change at the end are removed, and 24 slow features of the test sample are obtained as shown in the attached figure. Figure 3 shown.

[0114] Step 2: Through step 1, we can obtain 24 slow features for training and testing, and then we build spatial topology as spatial relationship based on the trained slow features. Note that this part is still based on training data, and the results are visualized using heat maps as shown in the attached figure. Figure 4 Note that the heat map here is only an example, and the specific spatial connectivity is calculated based on the prior knowledge of relevant engineers and different data.

[0115] Step 3 and Step 4: Spatiotemporal feature representation learning and comprehensive training based on variational graph attention autoencoder:

[0116] After steps 1 and 2, temporal features and spatial structures have been obtained, so they are input into the variational graph attention autoencoder for comprehensive training. After training, the comprehensive training model l oss value can be obtained, which is used to input the threshold value in step 5. In addition, the trained model can be obtained. Here, the spatiotemporal hidden vector Z of the test set is visualized by the trained model to illustrate the model's feature extraction capability. The attached image is obtained by visualization through dimensionality reduction processing. Figure 5 As shown in the figure, it can be seen that the model has a strong extraction capability and can better distinguish between faulty and normal samples.

[0117] Step 5: Threshold calculation based on kernel density estimation:

[0118] By setting the confidence level, an accurate threshold is calculated, which is used to distinguish whether the input sample is a fault. The recommended confidence level is (95% to 99%), and the threshold calculated by the present invention is 1.2445.

[0119] Step 6: Online Test:

[0120] After steps 1-5, a trained STVGATE model and an accurate threshold are obtained. Then, 400 test samples can be directly input, and the STVGATE model will output the loss value. By comparing this loss with the threshold, it can be determined whether the input is a fault. In this step, the false alarm rate (FAR) and fault detection rate (FDR) are used as performance indicators of the evaluation method to verify the scheme of this embodiment, and the comparative effect of fault detection in the zinc smelting and roasting process is obtained as shown in the attached figure. Figure 6 shown.

[0121] As shown in the figure above, PCA achieves a FAR and FDR of 17.5% and 42%, respectively, while K-SVD achieves a FAR and FDR of 30.5% and 60%, respectively. These low fault detection rates indicate that these methods are ineffective in detecting faults. Furthermore, the detection curves show that in many cases, the normal state statistics are higher than the fault statistics for PCA and K-SVD. This suggests that PCA and K-SVD cannot effectively extract features from multiple interconnected unit processes, making it difficult to distinguish between normal and faulty states. AE achieves a FAR of 37.5% and an FDR of 100%, which improves detection performance somewhat. However, it still fails to extract high-quality features from normal data, resulting in a relatively high FAR. The invented spatiotemporal variational graph attention autoencoder, which combines spatial and temporal information, achieves a FAR of 2.5% and an FDR of 100%. This FAR is 93.33% lower than that of AE, demonstrating that the spatiotemporal variational graph attention autoencoder model is able to reflect the interactions between interconnected unit processes and exhibits robust fault detection.

[0122] In order to further study the underlying reasons for the superiority of the proposed fault detection method, the t-distributed stochastic nearest neighbor embedding (t-SNE) algorithm is used to visualize the feature representations learned by different methods. The results are shown in the attached figure. Figure 7 As shown in the figure, the features extracted by PCA are random and overlapping. Meanwhile, the features extracted by K-SVD are sparse but not sufficient to distinguish faults. Although the quality of features extracted by AE is improved, they still show significant overlap. These observations suggest that the effectiveness of PCA, K-SVD, and AE in feature extraction and knowledge representation may be limited in fault decision making. In contrast, the method proposed in this paper can easily distinguish between normal and faulty data, as shown in the figure. Figure 6 These results demonstrate that STVGATE can capture the interactive information of time variables processed by multiple interrelated units, learn high-level spatiotemporal feature representations, and significantly improve the model's detection performance. This is crucial for accurately detecting faults in complex industrial processes and ensuring the normal operation of industrial systems.

[0123] This method addresses the fact that existing methods fail to fully consider the temporal and spatial aspects of industrial processes. By proposing an unsupervised framework for complex industrial process fault detection based on a spatiotemporal variational graph attention autoencoder, this method accurately detects faults in multiple interrelated unit processes through temporal dynamics extraction, spatial interaction graph topology construction, and integrated spatiotemporal feature representation learning and training. The detection threshold integrates temporal feature reconstruction, spatial relationship reconstruction, and distribution loss of the latent representation vector. Consequently, this method can meticulously reflect the operational status of industrial processes and accurately detect faults.

[0124] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as above in terms of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can, without departing from the scope of the technical solution of the present invention, make some changes or modifications to equivalent embodiments using the technical contents disclosed above. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder, characterized by: The following steps are involved: Step 1: Based on the temporal feature extraction of slow feature analysis (SFA), the slow component carrying important information is extracted and a conversion function is found to make the feature S(t) = g(X(t)) the slowest. After the sensor data passes through SFA, m SFs can be obtained. Step 2: Spatial graph topology construction based on metric learning and prior knowledge: construct the graph topology of the obtained SFs based on metric learning and prior knowledge; Step 3: Learning spatiotemporal feature representations based on variational graph attention autoencoders: The variational graph attention autoencoder combines the graph attention autoencoder (GATE) and the variational graph autoencoder (VGAE), replacing the original GCN encoder of VGAE with the graph attention network (GAT) encoder of GATE to generate high-quality latent spatiotemporal feature representations; Step 4: Comprehensive training of loss function: Calculate the loss of the distribution of the reconstructed graph spatial structure and the original graph spatial structure, the reconstructed graph node features and the original graph node features and potential representation, and then optimize. The optimized training object includes node features and graph topology structure. Step 5: Threshold calculation based on kernel density estimation. For the comprehensive training of the loss function in step 4, the fault detection threshold is determined by the distribution of the trained reconstruction loss function L. A non-parametric estimation method, kernel density estimation, is used to calculate the threshold of L. Step 6: Online testing.

2. The complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder according to claim 1 is characterized in that In step 1, the linear SFA learning paradigm is used to learn slow changes, which can be expressed as s(t) = Wx(t), where W is the weight matrix. The input measurement data can be solved to obtain m slow features (SFs) arranged from large to small in terms of slowness. The slowly changing features in (SFs) are the features that can properly represent the essential characteristics of the data, and the fast-changing features are regarded as noise signals, and the fastest-changing features are removed.

3. The complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder according to claim 1 is characterized in that: In step 2, each SFs variable is defined as a node, then the metric similarity between node i and node j can be expressed as A ij =M(H i ,H j ), using the metric method defined on the vector space, a complete norm vector space D is defined Chebyshev (H i ,H j )=max n (|H i -H j |).

4. The complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder according to claim 3 is characterized in that: Combining the metric similarity between node i and node j, the spatial structure is learned through the threshold parameter k. The spatial structure matrix information can be expressed as: Then, the adjacency matrix A′ can be constructed based on prior knowledge ij , where if it is determined that node i and node j have an inherent connection relationship, then A′ ij =1, otherwise, A′ ij =0, similarity matrix of comprehensive metric learning and prior knowledge matrix A′ ij , we can get the final spatial structure E: E contains feature information and prior knowledge information. Based on E, graph structure data G(V,E) can be constructed, where V is a slow feature variable and E is the structural information between features.

5. The complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder according to claim 4 is characterized in that: In step 3, the graph attention encoder is in VGATE, and the industrial process variable node S after SFA processing i In the application of k th The output of a GAT encoding layer can be expressed as: S represents the set of all industrial nodes, which can be obtained by the formula s(t) = Wx(t), where W is the weight matrix, E i represents the neighborhood spatial relationship set of node i, Represents the original feature S of the industrial variable node i (t),W (k) represents the weight matrix of the hidden layer of the k-th encoding layer feature map, σ(·) represents the activation function, is the kth th The correlation attention coefficient between neighbor node j and node j.

6. The complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder according to claim 5 is characterized in that: Relevance Attention Coefficient: represents the correlation coefficient between node i and its neighbor node j.

7. The complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder according to claim 6 is characterized in that: in, It is k th The trainable parameters of the encoder layer, θ(·) represents the activation function, σ(·) represents k th Activation function of the encoding layer.

8. The complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder according to claim 7 is characterized in that: In step 5, through the comprehensive training of the loss function, the spatiotemporal relationship between multiple interrelated unit processes is effectively learned and trained. The loss function value is an ideal detection indicator, and the fault detection threshold is specifically determined by the distribution of the reconstruction loss function L, which is trained on normal samples.

9. The complex industrial process fault detection method based on spatiotemporal variational graph attention autoencoder according to claim 8 is characterized in that: A non-parametric estimation method, kernel density estimation, is used to calculate the threshold of L, and the training is performed to obtain l i ,i=1,2,...,n vector, the kernel density estimate at point l is defined as: Where K(·) is the kernel function and h is the bandwidth.

Images