Uncertainty-aware Sensor Data Anomaly Detection for Autonomous Vehicles

Autonomous vehicles have stridden over the budding stage and are stepping into the phase of large-scale commercial deployment. Nonetheless, safety issues of autonomous driving remain to be fully solved. Sensor data provide the observations of the internal status and the driving environment of the autonomous vehicle, and sensor data anomaly detection is indispensable to ensure the safety since the occurrence of sensor data anomalies indicate potential safety risks. Tremendous works has contributed to the sensor data anomaly detection issue but most of them ignore the trustworthiness estimation of the anomaly detection results, leading to difficulties for decision-making in safety-critical systems. Therefore, this work proposes an uncertainty-aware sensor data anomaly detection method to enhance the trustworthiness of anomaly detection results. Specifically, this method includes a Bayesian LSTM prediction network that outputs both the predicted values and the distribution of the predicted values, an anomaly uncertainty quantification method, and an adaptive thresholding method to improve the anomaly detection performance. Anomaly detection is achieved by capturing the predicted values with high uncertainty. The efficacy and robustness of the proposed methodology have been substantiated through empirical field tests conducted with real-world autonomous driving vehicles. The evaluation yielded a recall of 0.893 and an F1-Score of 0.937, which underscores the superior anomaly detection capabilities of the approach within practical autonomous driving contexts.