A New Supervised Triple Deep Learning Strategy for Fault Isolation and Tolerant Cruise Control in Connected Autonomous Vehicles

Abstract

Fault-tolerant cruise control technologies are crucial for autonomous vehicles to ensure continuous operation and safety by effectively managing unexpected system failures. This article introduces a new supervised triple deep learning method for fault isolation and tolerant cruise control in connected autonomous vehicles (CAVs). First, the dynamic model of CAVs is captured using an autoencoder via measurable signals. Adaptive thresholds are then derived from the dynamic model to facilitate fault detection. Afterward, for fault isolation, a bidirectional long short-term memory (BiDLSTM) network is employed to classify the fault types. Later, three different supervised neural networks, including BiDLSTM, long short-term memory (LSTM), and fully connected networks, are trained to reconstruct the faulty signals. The outputs from these networks are integrated using a fusion method to generate robust reconstructed signals. Subsequently, these signals are forwarded into an adaptive neural network-based controller to mitigate the fault effects. The main contribution is that the proposed dynamic model with adaptive thresholding improves fault isolation with lower false alarm rates (FARs). In addition, the tolerant cruise control provided a robust control structure that can ensure connectivity and continuous operation. The comparative analysis with traditional methods shows that the proposed fault isolation and fault-tolerant method significantly improves efficiency and reliability. The proposed method leads to a missed alarm rate (MAR) of 3.7%, an FAR of 1.7%, and a correct detection ratio (CDR) of 97.0%. In addition, the fault classification process by using the BiDLSTM network achieves 98.1% accuracy.