A fault diagnosis framework for autonomous vehicles with sensor self-diagnosis

Fault diagnosis for autonomous vehicles aims to provide available information about the operation status of the vehicle to avoid potential risks, and sensor data provide the observations of the system only when sensors are proven to function adequately. Therefore, in the present work a fault diagnosis framework for autonomous vehicles with sensor self-diagnosis is proposed. It uses a residual consistency checking algorithm based on sensor redundancy to detect and isolate failed sensors in sensor self-diagnosis. Then, the denoising shrinkage autoencoder (DSAE) is put forward to address anomaly detection, where a shrinkage block with soft thresholding is embedded into the denoising autoencoder for feature representation enhancement, improving the anomaly detection performance. Several experiments with data collected from an autonomous vehicle in a real test field are implemented, and the results show that the proposed residual consistency checking algorithm can effectively detect and isolate the failed sensor, and the DSAE achieves relatively the best anomaly detection performance in terms of AUC_ROC and F1-score compared with several other machine learning based anomaly detectors studied.

Automated summary

Fault diagnosis for autonomous vehicles is important to provide information about the vehicle's operation status and avoid potential risks. This study proposes a fault diagnosis framework with sensor self-diagnosis, using a residual consistency checking algorithm and a denoising shrinkage autoencoder (DSAE) for anomaly detection. Experimental results show that the proposed algorithm effectively detects and isolates failed sensors, and the DSAE achieves the best anomaly detection performance compared to other machine learning methods.