Research On Anomaly Detection And Prediction Of Time Series Signals

The analysis of time series signals plays an important role in the maintenance of equipment status and early warning of failures.Today's complex equipment uses sensors to collect signals and detect the operating status of each component.However,in the process of using historical time series data to detect potential problems,due to factors such as equipment wear and device aging,there is uncertainty in the abnormal evolution,and the time series data is present.The output format is not uniform,the sample is unbalanced,and the fluctuations are disorderly.Common anomaly detection methods are difficult to model irregular fluctuations.When detecting anomalies in the potential relationship between multi-dimensional data,they will face the disaster of dimensionality.The abnormal period needs to be replaced by predicted values.Therefore,it is of great significance to study the anomaly detection and prediction of time series signals.In order to solve the above problems,this paper standardizes the original data,proposes the definition of anomaly types,and uses a combination of statistics and machine learning methods to propose an anomaly detection system based on the codec model that is adaptive and robust.,Can effectively suppress the impact of data without fluctuations.The main research work and innovative results of this article are summarized as follows:(1)Study the problem of detecting abnormal moments of unlabeled single-dimensional time series signals.Based on the bearing data and flight attitude two sets of sensor data,innovative anomalies and additional anomalies appearing in the signal are defined.In order to solve the decline in the adaptability of disordered oscillating signals in the codec model,and the failure problem of abnormal detection with predictive methods,a long and short-term memory network codec model based on time-series signal decomposition(Trend period remainder decomposition-Long short term memory)is proposed.,TPR-LSTM)multi-sensor adaptive anomaly detection model.The model performs statistical modeling and dynamic estimation of the periodic component through the periodic component,trend component and remainder of the time sequence signal of each sensor,and integrates the results of each component to realize the effective labeling of abnormal moments under irregular fluctuations.Finally,simulation experiments verify the effectiveness of the relevant methods.Compared with the original codec anomaly detection method,the proposed method can reduce the influence of disordered signal fluctuation on the system robustness.(2)Research the problem of abnormal relationship detection of multi-dimensional high-correlation time series signals.Multi-dimensional high-correlation time series signals have the characteristics of strong cross-correlation and strong auto-correlation.The abnormal duration is short and the state does not evolve over time.Through the analysis of the abnormal characteristics,the use of the difference change characteristics of adjacent periods,and the autocorrelation of the time series to make a preliminary location of the abnormal period.In addition,considering the coupling relationship among multi-dimensional variables,the potential relevance is modeled,the Pearson correlation coefficient is used to group the variables,and the expert experience is used to determine the threshold value to pave the way for the subsequent relationship model between variables.After verifying that the variables within the group show a linear relationship,a linear regression model is used to model the relationship between the variables.The two results are superimposed and fused,and the compensation and supplementary verification of the linear regression model reduce the impact of unknown disturbances on the robustness of the system.Finally,a simulation experiment verifies the effect and reliability of the differential linear regression abnormal detection model(Differential linear regression,DLR).(3)Study the prediction of multi-dimensional high-correlation time series signals.In order to consider the influence of other variables on the cross-correlation of predictors and improve the interpretability of the model,on the basis of the traditional convolutional neural network,residual network,causal convolution,and expansion convolution are added,and a temporal convolutional network model is established.Then,by changing the way of the convolution kernel,the historical information of other variables is introduced to synthesize the time convolutional network model(Multivariate compressed attention-Temporal convolutional network,MCA-TCN)to improve the attention mechanism.Finally,the comparison with the recurrent neural network verifies the prediction effect and applicability of the method.