Research On Fault Prediction Based On Multivariate Time Series Analysis

In many application domains,ranging from industry and home appliances to transportation,fault prediction has always been a hot topic.A sudden failure of equipment will cause incalculable loss.A reliable fault prediction system can give an early warning before the equipment failure,alerting people to do maintenance to prevent the occurrence of failure.However,the current fault prediction methods are still not mature enough,and the loss caused by equipment failure is untold every year:Therefore,this work needs to design a reliable fault prediction model to help operators do timely maintenance to ensure the safety of equipment.With the development of the Internet of Things technology and the Big Data technology,the running data of equipment can be recorded and saved to the remote storage system.And these data consist of multivariate time series,which can reflect the state of the equipment.We can make full use of the multivariate time series to analyze the state of equipment to predict the fault of equipment.This paper will study the problem of fault prediction based on multivariable time series analysis.However,the existing fault prediction methods based on multivariate time series analysis are usually based on prediction.The idea of these methods is to predict the future value of the equipment in a certain time step according to the historical time series of the equipment.These methods predict the fault based on the difference between the predicted value and the preset normal value.However,it is difficult to accurately predict the value at the future moment for the random equipment whose running state varies randomly,which makes the fault prediction difficult to achieve.According to observations,there are some symptoms before the failure of equipment,such as the value of one component is not consistent with that of other components.We call these symptoms anomalies in this paper.These anomalies can't actually lead the equipment to breakdown,but the anomalies continue to aggravate with time going by and eventually cause the equipment to breakdown.Based on this observation,we propose a two-stage fault prediction model based on anomaly detection and anomaly accumulation.This model mainly includes two stages.The first stage is the anomaly detection on multivariable time series.We first divide the multivariable time series into multiple sub-sequences by sliding windows,and then design an anomaly detection model to detect anomalies on these sub-sequences.The second stage is anomaly accumulation.We model the anomaly score sequence to mine the anomaly accumulation pattern of the fault equipment and predict the fault.This research includes two challenges.First,how to detect anomaly in such a complex and random equipment?Second,how to mine the anomaly accumulation pattern of equipment?The main works and contributions of this paper are summarized as follows:1.In order to predict fault,this paper proposes a novel two-stage fault prediction method based on anomaly detection and anomaly accumulation,called IFP-ADAC.Based on the cumulative effect of anomalies,the method starts from anomaly detection and then accumulates the detected anomalies to predict fault.Our model can not only predict fault on random equipment accurately,but also provide an explainable result compared with the end to end methods,which can provide instructions for maintenance personnel while ensuring the accuracy of results.2.In order to solve the problems of small samples and complex relationships among components of equipment,this paper proposes an anomaly detection model based on Generative Adversarial Networks,called CT-GAN.The input is normal sample,because there are few anomaly samples.The generator and discriminator are trained in an adversarial way,which improves the ability of the generator to generate normal samples and the discriminator to distinguish normal and abnormal samples.In addition,in order to model the complex relationships among multiple components,this paper uses Pearson Correlation Coefficient to model the correlations among multiple variables.A correlation map is calculated to represent the state of equipment.Finally,the reconstruction score of correlation map generator and multivariate time series generator,and the discriminator score of discriminator are jointly taken as the anomaly score,which improves the anomaly detection accuracy.3.In order to mine the anomaly accumulation pattern,this paper proposes an anomaly accumulation model based on LSTM and Attention Mechanism,called AttLSTM.To mine the temporal pattern of anomalies,we use LSTM to model the temporal feature of anomaly sequences.And the attention mechanism is introduced to model the severity of anomalies.Finally,MLP is used to map the fault features of equipment to prediction result.Att-LSTM not only captures the temporal feature of anomaly sequences but also takes the anomaly severity into consideration,which improves the accuracy of fault prediction.4.In order to verify the performance of the proposed model IFP-ADAC,this paper conducts a series of experiments on two real world datasets,the Internet of Vehicles(IOV)dataset and the subway Air Conditioning(AIC)dataset.We use Pre,Rec and F1score as the metrics to evaluate and compare the experimental results.The experimental results show that the overall performance of IFP-ADAC is better than the baselines.In addition,a fine-grained experimental analysis is also carried out to verify the effectiveness of different components and the influence of parameters.We also present the interpretability of our method.