Autoencoder and LSTM based fault diagnosis for intelligent vehicles

Abstract:

Fault diagnosis for intelligent vehicles is of great significance to ensure the safe driving. This paper proposes a series of fault diagnosis methods aiming at anomaly detection for sensor data and vehicle motion of intelligent vehicles. For the non-sequential sensor data, extreme learning machine based autoencoder is utilized to compress the normal data instances to learn the feature representation, and then reconstruct the data using the compressed feature. Whether an instance is normal or not is decided in accordance with the reconstruction error. To detect the anomaly in the sequential sensor data, multi-layer long-short time memory network is adopted to learn the time adherence of the sequential data to predict the current data value, and whether the data is normal or not is judged according to the prediction error. Besides, an adaptive threshold calculation method is proposed, where the threshold dynamically changes with the reconstruction error or prediction error and enable the decision variable sensitive to the anomaly. Furthermore, to detect whether the vehicle motion is abnormal, the vehicle bicycle kinematic model and Kalman filter are adopted and the normality of the residuals between the estimated and measured values is checked using Jarque-Berra test. The experiments verifies that the methods proposed in this paper can effectively detect the anomaly in the non-sequential or sequential sensor data, and detect the abnormality of the vehicle motion.