Deep Learning Based Process Fluctuation Anomaly Detection Algorithm Research

At present,industrial automation in China is in a period of booming development.To ensure that industrial processes fluctuate normally within a reasonable range,high-performance,highprecision sensors need to be installed to track the operating status of industrial equipment and to detect abnormal process fluctuations in a timely manner.Abnormality detection of industrial processes based on the abnormal data captured by the sensors can provide timely warning of potential abnormalities,thereby improving the quality of industrial products,reducing production costs,and increasing productivity and optimising process management.Therefore,the task of time series anomaly detection for industrial processes is of great importance and significance in real industrial production.This paper investigates anomaly detection algorithms for multivariate time-series process data based on a deep learning approach.The main research in this paper includes the following aspects:1.In response to the strong coupling and correlation features present in current multivariate time series process data,this paper proposes an anomaly detection model based on feature fusion and attention mechanisms.The model adopts convolutional coding layer and convolutional long and short term memory network layer to capture the temporal and correlation features of multivariate time series.After reconstructing the feature matrix at different scales using the convolutional decoding layer,a hierarchical feature fusion module is proposed to fuse the features at different levels so that the network captures a more comprehensive and accurate feature representation.Afterwards,a convolutional attention module is further introduced to weight the fused features in terms of channel and spatial dimensions,enhancing the model’s focus on important features and suppressing the response to irrelevant or redundant features.After experimental validation on several datasets,the model proposed in this paper achieves an average score of 0.9233 on the F1 metric,which is a significant advantage in anomaly detection.2.To further enhance the model’s capability for long-range time dependence in multivariate time series process data,a novel anomaly detection model(Temporal Graph Gated Variationaldetector,TGGV)is proposed in this paper.In this paper,spatio-temporal feature extraction and temporal feature reconstruction networks are designed in the model.The temporal feature extraction network is used to capture the temporal features of time series and the correlation of different time series.Then,this paper proposes a variational self-encoder structure for temporal feature reconstruction based on gated cyclic units.Finally,a threshold setting strategy is used in the model to reasonably set the threshold values.In this paper,the TGGV algorithm is compared with five frontier algorithms on several data sets.The experimental results show that the TGGV algorithm improves the F1 score by 14.7% ～ 16.5% compared with other frontier algorithms and can effectively detect potential anomalies in multivariate time series.3.This paper constructs a production process fluctuation anomaly detection system based on battery production process data and combines this system with the multivariate time series anomaly detection algorithm proposed in this paper.By conducting experiments on the coating process,the results show that the algorithm proposed and the system designed in this paper can identify possible abnormalities in the process in a timely manner,so that measures can be taken to repair or optimise the equipment in time to improve the reliability and stability of the process.