An Anomaly Detection Study on Automotive Sensor Data Time Series for Vehicle Applications

Anomaly detection in automotive systems has been a strong challenge: first, during the development phase, then after the manufacturing approval in ramp-up production and finally during the vehicles life cycle management. The numerous sensors positioned inside a vehicle generate more than a gigabyte of data at each second timeframe. These sensors are connected through the vehicle network, which comprises Electronic Control Units (ECUs) and Controller Area Network (CAN) buses. Each ECU gets input from its sensors, executes specific instructions and aims to monitor the vehicle's normal state detecting any irregular action corresponding to its observed behavior. The aggregator of all sensor data and control actions detects the anomalies in vehicle systems, which poses a multi-source big data problem. Detecting anomalies during manufacturing has turned out to be another research challenge after the introduction of Industry 4.0. This paper presents a performance comparison of different anomaly detection algorithms on time series originating from automotive sensor data. Interquartile range, isolation forest, particle swarm optimization and k-means clustering algorithms are used to detect outlier data in the study.