TSAN: A New Deep Learning-Based Detection Method for Sensor Anomaly in Mobile Robots

In the area of robotic systems, the detection of anomalies is a crucial capability for achieving long-term autonomy (LTA) of robots, as this capability ensures the stable operation of robots over extended periods. Adversary can launch injection attacks to interfere with various sensors, including speed and acceleration, thereby inducing abnormalities into the robot's operation. To address this gap, this article proposes a novel attention mechanism network, namely temporal shuffle attention network (TSAN), which efficiently analyzes time-series data obtained from mobile robot's internal sensors. TSAN combines the strengths of global temporal attention (GTA) and external attention (EA) for effective temporal feature extraction. By integrating these attention mechanisms and adding positional encoding, TSAN enhances the representation of time-domain information, facilitating the effective extraction of temporal data features. By exploiting the benefits of time-domain feature extraction, TSAN aims to enhance the performance of anomaly detection. The efficacy of TSAN is verified by performing a real experimental study on the mobile robot in the lab. It is shown that TSAN exhibits excellent detection performance for various types of anomaly scenarios. Moreover, comparisons of the multitype anomaly detection performance with other methods are carried out, which demonstrate the superiority of the proposed TSAN method.