CN-119512041-B - Layered multi-module-based flow industrial time sequence anomaly detection method

CN119512041BCN 119512041 BCN119512041 BCN 119512041BCN-119512041-B

Abstract

The invention belongs to the field of flow industrial anomaly detection, and in particular relates to a flow industrial time sequence anomaly detection method based on layered multi-module, which comprises the steps of dividing multi-element time sequence data, and respectively inputting the divided data into a layered coding layer through an input layer to obtain a low-dimensional feature vector; inputting the low-dimensional feature vector into a masking flow attention layer to acquire the coupling relation among all sub-flows in the data; inputting the low-dimensional feature vector with the coupling relation into a flow information memory layer, and excavating the coupling relation between adjacent sub-flow features; the invention adopts a masking attention layer and a flow information memory layer to capture the overall and local coupling relation among all sub-flows, thereby improving the accuracy of the detection result.

Inventors

ZHANG XIAOXIA
ZHOU YIDONG
YANG YAN
YU HONG
WANG GUOYIN

Assignees

重庆邮电大学

Dates

Publication Date: 20260505
Application Date: 20241115

Claims (9)

1. The flow industrial time sequence abnormality detection method based on the layered multi-module is characterized by comprising the steps of obtaining data to be detected, and inputting the data to be detected into a trained flow industrial time sequence abnormality detection model to obtain a detection result, wherein the flow industrial time sequence abnormality detection model comprises an input layer, a layered encoder, a flow attention masking layer, a flow information memory layer, a layered decoder and an output layer; the process industrial time sequence anomaly detection model is used for processing data to be detected and comprises the steps of dividing the data to be detected into multiple time sequence data, respectively inputting the divided data into a layered coding layer through an input layer to obtain low-dimensional feature vectors, inputting the low-dimensional feature vectors into a masking flow attention layer to obtain coupling relations among all sub-flows in the data, inputting the low-dimensional feature vectors with the coupling relations into a flow information memory layer to mine the coupling relations among adjacent sub-flow features, inputting the low-dimensional feature vectors with the coupling relations among all the sub-flows and the coupling relations among adjacent sub-flow features into a layered decoder to obtain reconstructed features, and inputting the reconstructed features into an output layer to obtain identification results.
2. The method for detecting industrial time sequence abnormality of the process based on the layered multi-module according to claim 1 is characterized in that the layered encoder consists of a space-time attention layer and an LSTM layer, wherein the processing of the data by the layered encoder comprises the steps of processing the data of the sub-process by a time attention mechanism to obtain a time attention calculation result, processing the data of the sub-process by a space attention mechanism to obtain a space attention calculation result, setting a time weight and a space weight, fusing the time attention calculation result and the space attention calculation result according to the time weight and the space weight to obtain a final output of the space-time attention, and inputting the final output of the space-time attention into the LSTM layer to obtain a low-dimensional feature vector.
3. The method for detecting the industrial time sequence abnormality of the process based on the layered multi-module according to claim 1, wherein the process masking attention layer processes the low-dimensional feature vector comprises the steps of calculating a process masking attention score matrix of each sub-process, and updating global information of the low-dimensional feature vector by adopting the process masking attention score matrix to obtain a coupling relation among the sub-processes.
4. A hierarchical multi-module based process industry timing anomaly detection method as set forth in claim 3, wherein calculating a masked process attention score matrix comprises: ; Where s is a mask operation, K m r is a flow key vector, Q m is a flow query vector, and d m is the dimension of K m r .
5. A hierarchical multi-module based process industrial timing anomaly detection method according to claim 3, wherein updating global information of low-dimensional feature vectors comprises: ; ; Where W i ' represents a trainable weight parameter matrix, b i ' represents a bias, M represents the number of sub-flows, j i,j represents an adaptive weight coefficient, M i ' is an output result of masking flow attention, and S i,j represents the attention weight of sub-flow i to sub-flow j.
6. The method for detecting the industrial time sequence abnormality of the process based on the hierarchical multi-module according to claim 1 is characterized in that the process information memory layer processes the low-dimensional feature vector and comprises the steps of initializing, memorizing and outputting a calculation vector for the process information memory layer of each sub-process, calculating a calculation result of the process information memory layer of the last sub-process and a forgetting variable to obtain historical state information which the current sub-process should memorize, calculating the low-dimensional feature vector of the current sub-process and the memory variable to obtain current state information which the current sub-process should keep, and obtaining an output result of the process information memory layer based on the calculated historical state information and the current state information, wherein the output result is transversely transmitted to the process information memory layer of the next sub-process for calculation while being output.
7. The hierarchical multi-module based process industrial timing anomaly detection method of claim 1, wherein the processing of the input features by the hierarchical decoder comprises: ; Wherein, the The reconstructed data representing the sub-process i, Representing the input data of the layered decoder.
8. A computer-readable storage medium having stored thereon a computer program, wherein the computer program is executed by a processor to implement the hierarchical multi-module based flow industrial timing anomaly detection method of any one of claims 1 to 7.
9. The flow industrial time sequence abnormality detection device based on the layered multi-module is characterized by comprising a processor and a memory, wherein the memory is used for storing a computer program, and the processor is connected with the memory and is used for executing the computer program stored in the memory, so that the flow industrial time sequence abnormality detection device based on the layered multi-module can execute the flow industrial time sequence abnormality detection method based on the layered multi-module according to any one of claims 1 to 7.

Description

Layered multi-module-based flow industrial time sequence anomaly detection method Technical Field The invention belongs to the field of flow industrial anomaly detection, and particularly relates to a flow industrial time sequence anomaly detection method based on layered multi-module. Background The process industry mainly comprises industries of chemical industry, energy, building materials and the like, is an important support of national economy, and is an important supporting force for continuously increasing the national economy in world manufacturing. It not only promotes the economic growth, but also promotes the effective utilization of resources and the efficient use of energy. However, with the continuous expansion of the production scale of the process industry and the continuous development of the production technology, the production process becomes finer and more complicated, so that the process industry is more prone to abnormality, the working efficiency of the process industry is inevitably affected, and even serious economic loss and safety accidents are caused. Therefore, it is necessary to develop a process industry anomaly detection research, perform data mining from the currently disclosed process industry time sequence data, perform accurate anomaly detection and provide anomaly positioning information, and have important significance for guaranteeing the safe and intelligent production of the process industry. In this regard, many scholars have developed related studies, and existing methods can be classified into two types, a statistical theory-based method and a machine learning-based method. Methods based on statistical theory include moving average, exponential smoothing, ARIMA model, etc. The method based on the statistical theory is better in interpretation. However, the statistical method usually performs problem solving based on the distribution characteristics of the data, but in reality, the distribution of the multi-element high-dimensional flow industrial data is often quite complex, so that the statistical method is difficult to quickly and effectively determine the real distribution characteristics of the data, and the difficult problems of difficult model construction and heavy calculation load exist. Compared with a method based on a statistical theory, the machine learning method can effectively process high-dimensional variables and mine data characteristic information from the characteristics and the structure of the data. The machine learning-based method comprises clustering, bayesian networks and the like, but due to the limited energy content of the mining characteristic information, the coupling among the process industry multiple variables is difficult to consider effectively, so that the abnormality detection precision is insufficient. With the rapid development of deep learning technology in recent years, the method has the advantages of strong feature extraction capability, capability of avoiding complex mechanism analysis and the like, so that the method becomes a main flow research direction of current industrial anomaly detection. Because the number of abnormal samples in the process industry is very small in the actual production process, the existing abnormality detection method is generally based on an unsupervised learning method, and the abnormality is detected by utilizing a depth neural network to mine the reconstruction error between normal data and abnormal data. However, the existing anomaly detection technology based on deep learning is mainly used for carrying out feature mining on time sequence characteristics of data, aiming at high dimensionality and coupling property of process industry multi-element data, the existing anomaly detection method of time sequence is focused on the fact that the anomaly detection precision of the existing anomaly detection method in the process industry data is still to be further improved, and more importantly, the existing method is difficult to realize accurate positioning of anomaly detection. Disclosure of Invention In order to solve the problems in the prior art, the invention provides a process industrial time sequence abnormality detection method based on layered multi-module, which comprises the steps of obtaining data to be detected, and inputting the data to be detected into a trained process industrial time sequence abnormality detection model to obtain a detection result, wherein the process industrial time sequence abnormality detection model comprises an input layer, a layered encoder, a process attention masking layer, a process information memory layer, a layered decoder and an output layer; the process industrial time sequence anomaly detection model is used for processing data to be detected and comprises the steps of dividing the data to be detected into multiple time sequence data, respectively inputting the divided data into a layered coding layer through an input layer to obtain low-d