Physics-Informed Neural Network for Detecting and Isolating Faults in Cooperative Adaptive Cruise Control

Abstract

This paper presents a Physics-Informed Neural Network (PINN) framework for detecting and isolating sensor and communication faults in Cooperative Adaptive Cruise Control (CACC). Unlike conventional data-driven methods, the proposed approach integrates the platoon dynamics into the neural network training process, enabling the model to learn from both measured data and physics-based constraints. Two learnable parameters are introduced to modulate the acceleration terms corresponding to communication and onboard sensors, allowing the network to identify the fault type and its source. By leveraging the inherent structure of the physical system, the PINN achieves reliable fault detection under limited and noisy measurements, outperforming data-driven techniques in accuracy, generalization, and interpretability. To validate the proposed framework, a comparative study was conducted using two existing fault detection methods, namely an autoencoder-based reconstruction approach and a Long Short-Term Memory (LSTM) predictor, demonstrating the superior effectiveness and capability of the PINN-based approach to identify faults, achieving substantially higher detection accuracy than both data-driven baselines. In addition to fault diagnosis, the framework also supports string stability analysis, ensuring that disturbance amplification along the vehicle platoon can be effectively evaluated. The proposed approach further demonstrates robustness under reduced sampling rates and measurement noise, maintaining clear discrimination between faulty and fault-free conditions. This approach demonstrates significant promise for real-world applications in system monitoring and fault identification in connected vehicles, even in the presence of noise and limited data.