Multi-Variate Time Series Anomaly Detection Based On Behavioral Patterns In Electric Vehicle Battery

With the gradual popularization of electric vehicles,the safety of electric vehicles has received much attention in recent years.How to ensure the safety of electric vehicle batteries has become a key concern for the industry and consumers,and today’s research on this issue mainly has the following shortcomings:(1)many studies expect to build an electrochemical model of the normal operation of lithium-ion batteries for anomaly detection.However,such methods are mainly based on laboratory data for modeling,and are often ineffective in real operating environments due to external factors such as geographic location and user driving habits.(2)Related studies consider the sensor data collected from EV battery as multivariate time series and then apply the time series anomaly detection method to it,but this type of research fails to make full use of the small amount of anomalous data and is ineffective,and the output results are poorly interpretable and the false alarm rate is high,which affects the user experience.This thesis also adopts the idea of multivariate time series anomaly detection for EV battery anomaly detection,and the main content and innovation points of this thesis are as follows to solve the above problems.First,to address the problem of underutilized anomaly labels,this thesis proposes a weakly supervised anomaly detection algorithm called WSPL based on pseudo labeling,which learns the gradual accumulation trend of data anomalies while learning the normal distribution pattern of data so as to provide effective warning of faults.In addition,the WSPL algorithm adopts an optimized time series regression fitting model to enhance the ability to capture time-space correlation information in multivariate time series.Experiments on real EV battery datasets and public datasets of industrial sensors show that the WSPL algorithm not only provides effective warning of electric vehicle battery faults with an F1 score of 0.94,which is 14.6% higher than the comparison model,but also performs well in the multivariate time series anomaly detection problem of traditional industrial sensors due to its ability to learn the gradual evolution of potential faults.Second,to address the problem of poor interpretability of traditional multivariate time series anomaly detection algorithm,this paper proposes the multivariate time series behavior pattern mining algorithm called MTSAPM,and conducts electric vehicle battery fault detection with strong interpretability based on the mined behavior patterns.The experiments show that the anomalous patterns mined by MTSAPM can effectively warn the electric vehicle battery thermal runaway problem which is difficult to be solved by traditional time series anomaly detection algorithm,and the F1 score reaches 0.62,which is 72.2% higher than that of the comparison model.In addition,more than 120 representative anomalous patterns obtained by the algorithm can be matched with the root cause of the battery mechanism,so that the constructed anomaly pattern library can guide the design and production process of electric vehicle batteries,which is not available in traditional anomaly detection algorithms.