High-dimensional time series analysis and anomaly detection: A case study of vehicle behavior modeling and unhealthy state detection

Abstract

With the growing complexity of modern vehicles in the digital era, ensuring the reliability of vehicle system becomes a challenge that needs a responsive framework to monitor and alert the unhealthy states. An unhealthy state refers to a timestamp in which vehicle operates off the expected range of performance. Being able to detect an unhealthy state early not only helps to identify imminent performance failures but also indicates the necessity of deploying preventive maintenance. It is almost impossible to be aware of an imminent failure without modeling the behavior of various subsystems of vehicle. This study proposes a behavior modeling and anomaly detection approach to accurately analyze the operation of various vehicle subsystems and trigger early alerts of unhealthy states. We propose a Multi-Layer Long Short Term Memory (LSTM) network that is integrated with an Autoencoder architecture (ML-LSTMAE) to monitor and predict the operation of different subsystems of vehicle. Modern vehicles involve various subsystems with respective sensors that record various operating signals. The proposed model leverages multivariate time series data sequences that are recorded by different sensors to train an encoding–decoding scheme. We use One-Class Support Vector Machine (OCSVM) to analyze the reconstruction errors and build a support boundary. Then, the trained support boundary evaluates the prediction errors in test data to distinguish the healthy and unhealthy states of vehicle. We validate the efficiency of the proposed behavior modeling and anomaly detection approaches via evaluating two case study problems. First, a case study of an operating vehicle data that consists of 15 months time series records of 101 data channels. Second, a NASA bearing case study data that includes 20,110,378 observational records of four bearings time series sequences. The results of both case studies confirm the high accuracy of the proposed ML-LSTMAE network for learning the latent behavior of different operating subsystems. Moreover, the proposed OCSVM algorithm classifies the healthy and unhealthy states in both case studies, precisely.