A layered multi-module based process industry time series anomaly detection method

Abstract

This invention belongs to the field of process industry anomaly detection, specifically involving a hierarchical multi-module method for detecting time-series anomalies in the process industry. The method includes: dividing multivariate time-series data and inputting the divided data into hierarchical encoding layers through an input layer to obtain low-dimensional feature vectors; inputting the low-dimensional feature vectors into a masking process attention layer to obtain the coupling relationships between various sub-processes in the data; inputting the low-dimensional feature vectors with coupling relationships into a process information memory layer to mine the coupling relationships between features of adjacent sub-processes; inputting the low-dimensional feature vectors with coupling relationships between various sub-processes and between features of adjacent sub-processes into a hierarchical decoder to obtain reconstructed features; and inputting the reconstructed features into an output layer to obtain the recognition result. This invention uses a masking attention layer and a process information memory layer to capture the global and local coupling relationships between various sub-processes, thereby improving the accuracy of the detection results.

Description

Technical Field

[0001] This invention belongs to the field of process industry anomaly detection, specifically relating to a process industry timing anomaly detection method based on hierarchical multi-module. Background Technology

[0002] Process industries, primarily encompassing chemicals, energy, and building materials, are a vital pillar of the national economy and a crucial support for the sustained economic growth of major global manufacturing nations. They not only drive economic growth but also promote the efficient utilization of resources and the high-efficiency use of energy. However, with the continuous expansion of production scale and the development of production technologies in process industries, their production processes have become increasingly sophisticated and complex. This makes process industries more prone to anomalies, inevitably impacting their operational efficiency and potentially causing significant economic losses and safety accidents. Therefore, it is necessary to conduct research on anomaly detection in process industries. Data mining from publicly available time-series data in process industries can enable precise anomaly detection and provide anomaly location information, which is of great significance for ensuring safe and intelligent production in these industries. Numerous scholars have conducted related research, and existing methods can be categorized into two types: methods based on statistical theory and methods based on machine learning.

[0003] Statistical theory-based methods include moving averages, exponential smoothing, and ARIMA models. These methods offer good interpretability. However, statistical methods typically solve problems based on the distribution characteristics of data. In reality, the distribution of multivariate, high-dimensional process industry data is often extremely complex, making it difficult for statistical methods to quickly and effectively determine the true distribution characteristics. This leads to challenges such as difficult model construction and heavy computational burden. In contrast to statistical theory-based methods, machine learning methods can start from the inherent characteristics and structure of the data, effectively handling high-dimensional variables and mining data feature information. Machine learning methods include clustering and Bayesian networks, but due to their limited ability to mine feature information, they struggle to effectively consider the coupling between multivariate variables in process industries, resulting in insufficient anomaly detection accuracy.

[0004] With the rapid development of deep learning technology in recent years, its advantages, such as strong feature extraction capabilities and the ability to avoid complex mechanism analysis, have made it the mainstream research direction for anomaly detection in process industries. Since the number of anomaly samples in actual production processes is extremely small, existing anomaly detection methods are usually based on unsupervised learning, using deep neural networks to mine the reconstruction error between normal and anomalous data to detect anomalies. However, existing deep learning-based anomaly detection technologies mainly focus on feature mining based on the temporal characteristics of the data. They rarely address the high dimensionality and coupling of multi-dimensional data in process industries, resulting in a need for further improvement in anomaly detection accuracy. More importantly, existing methods struggle to achieve precise anomaly localization. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a hierarchical multi-module method for detecting time-series anomalies in the process industry. The method includes: acquiring data to be detected, inputting the data to be detected into a trained process industry time-series anomaly detection model, and obtaining detection results; wherein the process industry time-series anomaly detection model includes an input layer, a hierarchical encoder, a masking process attention layer, a process information memory layer, a hierarchical decoder, and an output layer.

[0006] The process industry time-series anomaly detection model processes the data to be detected in the following ways: the data to be detected is multi-dimensional time-series data, which is divided into segments. The segmented data is then input into a hierarchical encoding layer through an input layer to obtain low-dimensional feature vectors. These low-dimensional feature vectors are then input into a masked process attention layer to obtain the coupling relationships between various sub-processes in the data. The low-dimensional feature vectors with coupling relationships are then input into a process information memory layer to mine the coupling relationships between features of adjacent sub-processes. The low-dimensional feature vectors with coupling relationships between various sub-processes and between features of adjacent sub-processes are then input into a hierarchical decoder to obtain reconstructed features. Finally, the reconstructed features are input into an output layer to obtain the recognition result.

[0007] To achieve the above objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any of the above-described hierarchical multi-module process industry timing anomaly detection methods.

[0008] To achieve the above objectives, the present invention also provides a process industry timing anomaly detection device based on hierarchical multi-module, comprising a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory and is used to execute the computer program stored in the memory, so that the process industry timing anomaly detection device based on hierarchical multi-module executes any of the above-mentioned process industry timing anomaly detection methods based on hierarchical multi-module.

[0009] The beneficial effects of this invention are:

[0010] This invention proposes a hierarchical, multi-module framework for multivariate temporal anomaly detection in the process industry based on unsupervised learning methods. This framework first decomposes process industry data into multiple low-dimensional sub-process data, which are then input into different encoding layers for hierarchical encoding learning. Because there is unidirectional data flow between sub-processes in the process industry—that is, feature information from upstream sub-processes is passed to downstream sub-processes—a masking attention layer and a process information memory layer are designed to capture the global and local coupling relationships between the sub-processes. This invention reconstructs the sub-process data by inputting it into different decoding layers, and finally detects anomalies based on the magnitude of the reconstruction error of the process data. Attached Figure Description

[0011] Figure 1 This is a framework diagram of the hierarchical multi-module-based process industry timing anomaly detection method of the present invention;

[0012] Figure 2 This is a structural diagram of the masking process attention layer of the present invention;

[0013] Figure 3 This is a structural diagram of the process information memory layer of the present invention;

[0014] Figure 4 This is a visualization experiment diagram for locating anomalies in the process industry according to the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] The technical solution adopted to achieve the objectives of this invention is as follows: First, a hierarchical autoencoder for the process industry is built based on the idea of ​​unsupervised learning. This hierarchical autoencoder reduces learning complexity and achieves anomaly localization by decomposing process industry data, and utilizes LSTM and spatiotemporal attention mechanisms to acquire temporal feature information of the time-series data. Next, to consider the global and local coupling relationships between various sub-processes in the process industry, this patent proposes a masking attention layer and a process information memory layer. Combining the hierarchical autoencoder, masking attention layer, and process information memory layer, a hierarchical multi-module framework for multivariate temporal anomaly detection in the process industry is proposed. Finally, experiments using three publicly available process industry datasets verify the effectiveness of this invention in anomaly detection and localization within the process industry.

[0017] A hierarchical multi-module method for detecting time-series anomalies in process industries is disclosed. The method includes: acquiring data to be detected; inputting the data into a trained process industry time-series anomaly detection model to obtain detection results; wherein the process industry time-series anomaly detection model includes an input layer, a hierarchical encoder, a masking process attention layer, a process information memory layer, a hierarchical decoder, and an output layer; the process industry time-series anomaly detection model processes the data to be detected by: the data to be detected is multi-dimensional time-series data; the multi-dimensional time-series data is divided, and the divided data is input into the hierarchical encoder layer through the input layer to obtain low-dimensional feature vectors; the low-dimensional feature vectors are input into the masking process attention layer to obtain the coupling relationships between various sub-processes in the data; the low-dimensional feature vectors with coupling relationships are input into the process information memory layer to mine the coupling relationships between features of adjacent sub-processes; the low-dimensional feature vectors with coupling relationships between various sub-processes and between features of adjacent sub-processes are input into the hierarchical decoder to obtain reconstructed features; and the reconstructed features are input into the output layer to obtain the recognition result.

[0018] This invention proposes a hierarchical, multi-module framework for multi-element time-series anomaly detection in process industries. For example... Figure 1 As shown, the hierarchical multi-module framework can be divided into four parts: a hierarchical encoder, a masked process attention layer, a process information memory layer, and a hierarchical decoder. Process industry operational data can be collected in real time through various monitoring systems and sensors. In the process industry, the entire process flow consists of many sub-processes, and each sub-process contains many sensors. Therefore, the multivariate time-series dataset collected in the process industry can be described by the formula:

[0019]

[0020] in P i Indicates the first i The time-series data for each sub-process is a subset of the original process industry dataset. M Indicates the number of subprocesses. , T This indicates the length of the time series. After dividing the process industry dataset according to formula (1), the data of the sub-processes... P i The input is fed into the hierarchical encoder for hierarchical coding learning.

[0021] The hierarchical encoder consists of a spatiotemporal attention layer and an LSTM layer. The data processing steps include: processing the sub-process data using a temporal attention mechanism to obtain temporal attention calculation results; processing the sub-process data using a spatial attention mechanism to obtain spatial attention calculation results; setting temporal and spatial weights, and fusing the temporal and spatial attention calculation results based on these weights to obtain the final spatiotemporal attention output; and inputting the final spatiotemporal attention output into the LSTM layer to obtain a low-dimensional feature vector. The calculation formula for the hierarchical encoder is as follows:

[0022]

[0023]

[0024]

[0025]

[0026] in P i It is a sub-process i Data, t Represents the number of samples. K and Q Represents key and query. A T and A S These represent the calculation results for temporal and spatial attention, respectively. It is the final output of spatiotemporal attention. α and β These represent the weights of temporal attention and spatial attention, used to balance the degree to which data focuses on temporal and spatial information. Through LSTM layers, the hierarchical encoder maps each sub-process data into a low-dimensional feature vector. z i As shown in formula (6):

[0027]

[0028] In the code, HEN(*) indicates a layered encoder. P i Indicates the first i Input data for each sub-process z i express P i The low-dimensional feature vector after mapping by the hierarchical encoder. Then the low-dimensional feature vector... z iThe input is fed into the masking process attention layer to capture the coupling relationships between various sub-processes from a global perspective. The structure of the masking process attention layer is as follows: Figure 2 As shown. In the masking process attention layer, the first step is to calculate the masking process attention score matrix for each sub-process. The calculation formula is as follows:

[0029]

[0030] in S It is a masking process attention score matrix. S i,j Subprocess i Pair process j The attention weights are defined using `mask(*)`, which represents a masking operation to extract the lower left triangular portion of the attention score matrix, setting the rest to 0. The goal is to assign each subprocess a weight score based on its influence on itself and its upstream subprocesses. This is followed by the low-dimensional feature vectors of each subprocess. z i This matrix will be used to update the global process information. The calculation formula is as follows:

[0031]

[0032]

[0033] in, W i ' represents the trainable weight parameter matrix, b i ′ represents bias. M Indicates the number of subprocesses. j i,j Indicates the adaptive weighting coefficient. This is the output of masking process attention. The masking attention layer ensures that the current subprocess only focuses on the information of the upstream process by setting the attention weight score of the downstream subprocess to 0, and at the same time uses adaptive weights to adaptively increase the attention to the upstream subprocess from far to near.

[0034] After passing through the masking attention layer, the low-dimensional feature vector z i These features will be input into the process information memory layer to mine the coupling relationships between features of adjacent sub-processes. The structure of the process information memory layer is as follows: Figure 3 As shown. Information from adjacent subprocesses is selectively remembered and forgotten in the process information memory layer and passed to the next subprocess for continued memory-forgetting calculation. Before performing memory-forgetting, the weighting coefficient for memory-forgetting must be calculated, and the calculation formula is as follows:

[0035]

[0036]

[0037]

[0038] in, r i , f i , o i These represent the memory coefficient, forgetting coefficient, and output coefficient, respectively. c i-1 and h i-1 This represents the process status and output of the previous sub-process. U and V It is a randomly initialized weight matrix. b It's a bias. Then the low-dimensional feature vectors... z i The system uses calculated coefficients to aid in information retention and forgetting. The calculation formula is as follows:

[0039]

[0040]

[0041]

[0042] Formula (14) utilizes information from the previous subprocess. c i-1 With current process information g i Update process status c i Then formula (15) uses the output coefficients o i Current process status information c i Calculate the memory forgetting results h i After passing through formulas (13)-(15), the process information memory layer will cause the current subprocess to remember or forget the information of the current subprocess and the previous subprocess and output it, while passing the result to the next subprocess.

[0043] After the low-dimensional feature vectors of each sub-process are processed by the masking process attention and process information memory layers, they are input into the hierarchical decoder for hierarchical decoding. The hierarchical decoder gradually restores the dimensions of the low-dimensional feature vectors of each sub-process to the original feature vector dimensions. The internal calculation formula of the hierarchical decoder is the same as that of the hierarchical encoder, therefore its calculation formula is as follows:

[0044]

[0045] Where HDE(*) represents a layered decoder. Indicates the first i The reconstructed data for each sub-process. Since the training objective of unsupervised learning-based temporal anomaly detection is to minimize the difference between the sub-process data and its reconstructed data, the loss function for each sub-process is calculated as follows:

[0046]

[0047] in P i,j and Indicates the first i Sub-process number j Input data and reconstructed data for each feature vector. m This represents the number of features in this subprocess. l i Representing the i The loss value of each subprocess. T This represents the length of the time series. Since the original time series dataset has been divided into multiple sub-process data streams, the loss value of each sub-process data stream relative to its reconstructed data can be calculated using a loss function. l i During the training phase of the proposed framework, the model updates its parameters through backpropagation using the loss values ​​from each sub-process. The calculation formula is as follows:

[0048]

[0049] After calculating the anomaly score at each time point, a threshold is needed as the criterion for judging anomalies. To this end, this invention designs a threshold calculation formula based on the mean and variance, as follows:

[0050]

[0051]

[0052]

[0053] in, This represents the mean of the outlier scores over all time points. S The standard deviation of outlier scores over all time periods. k These are the hyperparameters that are set. If the first... j If the anomaly score calculated at a given time is greater than the threshold, the predicted value for that time is 1 (abnormal); otherwise, it is 0 (normal).

[0054] In this embodiment, the process information memory layer processes the low-dimensional feature vector as follows: for each sub-process's process information memory layer, initialize the memory, forget, and output calculation vectors; calculate the historical state information that the current sub-process should remember by combining the calculation result of the previous sub-process's process information memory layer with the forget variable; calculate the current state information that the current sub-process should retain by combining the low-dimensional feature vector of the current sub-process with the memory variable; and obtain the output result of the process information memory layer based on the calculated historical state information and current state information. This output result is then horizontally passed to the process information memory layer of the next sub-process for calculation.

[0055] In this embodiment, the present invention used three publicly available datasets in the experiment. These three datasets are all process industry time series datasets. Table 1 describes the characteristic information of these datasets.

[0056] Table 1 Introduction to Process Industry Datasets

[0057]

[0058] Because datasets have different feature types, their units and measurement standards may vary. Therefore, preprocessing to standardize the data units and measurement standards is usually necessary. This invention employs data normalization to unify the data units and measurement standards. The training sets of these three datasets do not contain outliers, while the test sets do.

[0059] To verify the superiority of the method proposed in this invention, it is compared with six current mainstream multivariate time-series anomaly detection methods:

[0060] USAD: This model is based on unsupervised learning and uses two decoders as discriminators and generators. These two decoders are trained adversarially in an autoencoder and finally detect anomalies through reconstruction error.

[0061] GDN: GDN is a multivariate temporal anomaly detection model based on graph neural networks. It identifies anomalies by combining structure learning methods with graph neural networks.

[0062] OmniAnomaly: This model learns feature representations of multivariate time series through techniques such as random variable concatenation and planar normalization, captures the normal pattern of multivariate time series, reconstructs the input data using these representations, and uses the reconstruction probability to identify anomalies.

[0063] MAD-GAN: This model is trained based on the generative adversarial approach. The discriminator and generator are composed of LSTM-RNN, and anomalies are detected by using the DR score as the anomaly score.

[0064] MTAD-GAT: This model uses two graph attention layers to learn the complex correlations of multivariate time series in the time dimension and feature dimension, respectively, and finally detects anomalies in time series data through two methods: reconstruction and prediction.

[0065] FuSAGNet: This model uses a combination of sparse autoencoders and graph neural networks to learn the dependencies between features, and then identifies anomalies through reconstruction and prediction of time-series data.

[0066] During the experiments, the proposed framework's hierarchical configuration for the original process industry dataset is shown in Table 2. This configuration is based on the sub-process division information provided in the official dataset documentation. Taking the SWaT dataset as an example, the SWaT dataset contains 50 features, which together form 6 sub-processes. The first 5 features belong to the first sub-process, the 6th to 11th features belong to the second sub-process, and so on.

[0067] Table 2 Introduction to Process Industry Datasets

[0068]

[0069] In terms of evaluation indicators, classification accuracy is adopted. A Classification accuracy P Category recall rate R , F A score is used to measure the accuracy of the model in detecting anomalies in the process industry. The classification accuracy is the ratio of correctly predicted anomalies to the total number of samples, calculated as follows:

[0070]

[0071] In the formula, A For classification accuracy; TP This represents the number of genuine outlier samples that were correctly classified. TN The number of real, normal samples that were correctly classified. FN This represents the number of genuine abnormal samples that were misclassified as normal samples. FP This represents the number of genuine normal samples that were misclassified as abnormal samples.

[0072] Classification precision is the ratio of correctly classified outliers to the number of outliers identified as outliers by the model. The formula is as follows:

[0073]

[0074] Classification recall is the ratio of correctly classified outliers to the number of truly outliers. The formula for calculating it is as follows:

[0075]

[0076] F The 1 value is an evaluation metric for model classification performance; it is the harmonic mean of accuracy and recall. F The value of 1 ranges from 0 to 1, with a larger value indicating better classification performance. The calculation formula is as follows:

[0077]

[0078] Table 3 presents the anomaly detection performance of the proposed framework compared to other methods. Overall, experiments on three real-world process industry datasets show that the proposed framework achieves the highest accuracy across all scenarios.

[0079] On the WADI dataset, the proposed framework achieved precision, recall, and F1 score of 97.12%, 96.86%, and 96.99%, respectively, with the F1 score being approximately 50% higher than the comparison methods. Due to the large number of features in WADI, the anomaly detection performance of the comparison methods was generally poor. On the HAI dataset, the proposed framework achieved the highest F1 score and precision, reaching 80.51% and 92.73%, respectively, with precision only 1.35% lower than the FuSAGNet model. Compared to the WADI dataset, the comparison methods showed improved detection performance, mainly because the HAI dataset has fewer features, reducing the computational burden on the comparison methods. On the SWaT dataset, the proposed framework achieved a precision of 99.77%, a 0.8% improvement over MAD-GAN, while also achieving a precision of 95.96%. Compared to the WADI dataset, the comparison methods showed a significant improvement in detection performance on SWaT, while the proposed anomaly detection framework performed comparably to current mainstream methods on the SWaT dataset. The main reason is that the number of features in the WADI dataset is significantly increased compared to the SWAT dataset. The high-dimensional coupled features make anomaly detection much more difficult, resulting in a deterioration in the detection performance of the comparison method on the WADI dataset. However, the proposed framework can perform decomposition learning, thus maintaining good performance in high-dimensional data.

[0080] Experimental data from three datasets reveals that high-dimensional and complex data features increase model computational complexity, leading to a decline in detection performance of comparative methods. Therefore, as the problem scale increases, it is necessary to design methods to simplify the problem and reduce computational burden. The anomaly detection framework proposed in this invention can decompose process features, transforming the original, relatively complex process industry anomaly detection problem into multiple sub-process anomaly detection problems, reducing the difficulty of model learning and thus improving anomaly detection performance. The results of the anomaly detection performance comparison are shown in Table 3.

[0081] Table 3 Comparison of Anomaly Detection Performance

[0082]

[0083] Since hierarchical autoencoders can decompose high-dimensional time series, the reconstruction error of each sub-process can be calculated and used for anomaly localization. Therefore, corresponding anomaly localization visualization experiments were conducted on three datasets, such as... Figure 4 As shown in the diagram. The horizontal axis of each subplot represents the sub-process, and the vertical axis represents the reconstruction error. Each subplot represents two reconstruction errors of the sub-process, representing the errors under normal and abnormal conditions, respectively. Figure 4 (a), (b), and (c) correspond to experimental data from the WADI, HAI, and SWAT datasets, respectively. Figure 4 It can be seen that there are certain differences in the reconstruction error values ​​of each subprocess under normal and abnormal conditions. Therefore, when the error value of a subprocess differs from the error value under normal conditions, it can be determined that the subprocess is in an abnormal state.

[0084] In one embodiment of the present invention, the present invention further includes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described hierarchical multi-module process industry timing anomaly detection methods.

[0085] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented using computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0086] A process industry timing anomaly detection device based on hierarchical multi-modules includes a processor and a memory; the memory is used to store computer programs; the processor is connected to the memory and is used to execute the computer programs stored in the memory, so that the process industry timing anomaly detection device based on hierarchical multi-modules can perform any of the above-mentioned process industry timing anomaly detection methods based on hierarchical multi-modules.

[0087] Specifically, the memory includes various media capable of storing program code, such as ROM, RAM, magnetic disk, USB flash drive, memory card, or optical disk.

[0088] Preferably, the processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0089] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting timing anomalies in process industries based on hierarchical multi-module architecture, characterized in that, include: Acquire the data to be detected, input the data to be detected into the trained process industry time series anomaly detection model, and obtain the detection results; The process industry timing anomaly detection model includes an input layer, a layered encoder, a masked process attention layer, a process information memory layer, a layered decoder, and an output layer. The process industry time-series anomaly detection model processes the data to be detected in the following ways: the data to be detected is multi-dimensional time-series data, which is divided into segments. The segmented data is then input into a hierarchical encoding layer through an input layer to obtain low-dimensional feature vectors. These low-dimensional feature vectors are then input into a masked process attention layer to obtain the coupling relationships between various sub-processes in the data. The low-dimensional feature vectors with coupling relationships are then input into a process information memory layer to mine the coupling relationships between features of adjacent sub-processes. The low-dimensional feature vectors with coupling relationships between various sub-processes and between features of adjacent sub-processes are then input into a hierarchical decoder to obtain reconstructed features. Finally, the reconstructed features are input into an output layer to obtain the recognition result.

2. The method for detecting timing anomalies in process industries based on hierarchical multi-modules as described in claim 1, characterized in that, The hierarchical encoder consists of a spatiotemporal attention layer and an LSTM layer; The hierarchical encoder processes the data in the following ways: it uses a temporal attention mechanism to process the data of the sub-processes to obtain the temporal attention calculation results; it uses a spatial attention mechanism to process the data of the sub-processes to obtain the spatial attention calculation results; it sets temporal and spatial weights, and fuses the temporal and spatial attention calculation results based on the temporal and spatial weights to obtain the final spatiotemporal attention output; and it inputs the final spatiotemporal attention output into the LSTM layer to obtain a low-dimensional feature vector.

3. The method for detecting timing anomalies in process industries based on hierarchical multi-modules as described in claim 1, characterized in that, The masking process attention layer processes low-dimensional feature vectors by: calculating the masking process attention score matrix for each sub-process; and updating the global information of the low-dimensional feature vectors using the masking process attention score matrix to obtain the coupling relationship between the sub-processes.

4. The method for detecting timing anomalies in process industries based on hierarchical multi-modules according to claim 3, characterized in that, The calculation of the masking process attention score matrix includes: ； Where s is the attention score matrix of the masking process, and mask is the masking operation. K m r Q is the process key vector. m For the process query vector, d m for K m r The size of the dimension.

5. The method for detecting timing anomalies in process industries based on hierarchical multi-modules according to claim 3, characterized in that, Updating the global information of the low-dimensional feature vector includes: ； ； in, W i ' represents the trainable weight parameter matrix, b i ′ represents bias. M Indicates the number of subprocesses. j i,j m represents the adaptive weighting coefficient. i 'It is the output that masks process attention' S i,j Subprocess i Pair process j The weight of attention.

6. The method for detecting timing anomalies in process industries based on hierarchical multi-modules according to claim 1, characterized in that, The process information memory layer processes low-dimensional feature vectors as follows: For each sub-process, the process information memory layer initializes memory, forget, and output calculation vectors; the calculation results of the previous sub-process's process information memory layer are calculated with the forgetting variable to obtain the historical state information that the current sub-process should remember; the low-dimensional feature vector of the current sub-process is calculated with the memory variable to obtain the current state information that the current sub-process should retain; based on the calculated historical and current state information, the output result of the process information memory layer is obtained, and this output result is simultaneously passed to the process information memory layer of the next sub-process for calculation.

7. The method for detecting timing anomalies in process industries based on hierarchical multi-modules according to claim 1, characterized in that, The hierarchical decoder processes the input features in the following ways: ； in, Representative subprocess i Reconstructed data, This represents the input data for the layered decoder.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor to implement any one of the hierarchical multi-module process industry timing anomaly detection methods according to claims 1 to 7.

9. A process industry timing anomaly detection device based on hierarchical multi-module architecture, characterized in that, It includes a processor and a memory; the memory is used to store computer programs; the processor is connected to the memory and is used to execute the computer programs stored in the memory, so that the hierarchical multi-module process industry timing anomaly detection device performs any one of the hierarchical multi-module process industry timing anomaly detection methods according to claims 1 to 7.

Images