Fusion-Based Deep Learning Techniques for Fault Isolation in Connected Vehicles

An accurate and robust fault isolation technique is essential to guarantee the safety of Connected Autonomous Vehicles (CAVs) systems. This study presents a hybrid FDI approach that utilizes nonlinear system modeling, adaptive thresholding, and deep learning-based fault isolation techniques. First, the dynamic behavior of CAVs is modeled using Bidirectional Long Short-Term Memory (BiLSTM), Orthonormal Basis Function (OBF), and Multi-Layer Perceptron (MLP) networks to mitigate system uncertainties and disturbances. Adaptive thresholds are then designed based on these models to improve fault detection by reducing false alarms. To isolate faults, four deep learning-based classifiers are trained, and their results are integrated using multiple decision fusion methods to enhance robustness. The main contribution is that an ensemble of several deep learning models improves feature extraction diversity and enhances the robustness of fault diagnosis. Another benefit is that the dynamics of the plant are accurately captured using Laguerre filters, MLP, and a BiLSTM-based error model, and significantly mitigates the effects of uncertainties, which boosts accuracy in both fault detection and fault isolation. Finally, implementing fusion strategies improves the reliability and accuracy of the fault isolation process. Test results demonstrate that the presented FDI method is more sufficient than conventional approaches and provides a reliable fault detection and isolation structure under varying operating conditions.