Anomaly Detection And Root Cause Analysis For Time Series Based Industrial Processes

With the rapid development of modern technology,industrial systems are becoming more and more sophisticated,and the widespread use of intelligent sensors makes industrial processes generate and store a large number of complex time series data,which present multi-dimensional,coupled,non-linear and strong causal characteristics,consequently the mining and analysis of industrial process time series data has become a hot research.The industrial production system requires stable and efficient operation process,and how to analyze and solve the emergence of faults is one of the key issues.Industrial system fault analysis includes anomaly detection and root cause analysis.Anomaly detection requires timely and accurate anomaly judgment of the time series object,and root cause analysis requires variable analysis and root cause tracing of industrial faults.In this paper,based on time series data,we propose an integrated model for anomaly detection and a diagnosis framework for root cause analysis for the analysis of faults in industrial processes,and extend the fixed-lag Granger causality analysis method to variable-lag.The content of this paper is instructive for industrial faults analysis.The main research of this paper is as follows.（1）In the task of industrial time series anomaly detection with the reconstruction model,the reconstruction error differentiation is not sufficient.To address this problem,based on the long short term memory-autoencoder（LSTM-AE）model,according to the principle of integration idea,an LSTM autoencoder ensembles（LAEE）framework is proposed.In the training stage,LSTM-AEs with different hidden layer dimensions are trained,the base detectors are selected according to the detection performance of the training stage,and their weights were calculated.In the phase of anomaly detection,a new reconstruction error matrix is obtained by weighted integration of the reconstruction errors generated by each base detector for anomaly identification.Experimental results on two kinds of data sets show that the LAEE improves the accuracy of time series anomaly detection in industrial process.（2）For the root cause analysis problem of new types of faults without accurate historical data,a root cause analysis diagnosis framework is proposed.First,for the variable analysis problem,based on the principle of extreme gradient boosting tree,the binary-extreme gradient boosting（BiXgboost）model is proposed,which can analyze the importance of variables in real time when a fault occurs.Second,for the problem of fault-related variable selection,based on the number of variables,the mean contribution threshold（MCT）method is proposed,which can keep the selected number of fault-related variables within a suitable range.Finally,for the problem of causal analysis of faultrelated variables,the temporal causal analysis network（TCDN）is introduced,which is based on the prediction and attention mechanism to determine the causal logic between multi-dimensional time series.The validation results in the industrial simulation process show that the proposed framework is able to perform timely and accurate root cause analysis for different types of faults in industrial processes.（3）Previous time series causal analysis methods always assuming fixed causal lags,to address this problem,the variable lag Granger causality analysis（VLGC）algorithm is proposed.First,the adjustment between time series is performed by dynamic time warping,and the optimally aligned time series is reconstructed,thereafter,based on the principle of Granger causality analysis,vector autoregressive models were established,and the true causal relationships between the time series were obtained by comparing the predicted residuals among the models.The proposed algorithm is validated using the synthetic and actual datasets,respectively.The results show that the accuracy of the variable-lag Granger causal analysis method in performing causal analysis is higher than the baseline causal analysis methods with fixed-lag assumptions.