An Integrated Safety Architecture for Learning-Based Control Systems: Experimental Evaluation with Verification and Fault Detection

Abstract

Safety guarantees are essential for learning-based controllers for autonomous systems before putting them into practice in the real world. This paper introduces an integrated safety architecture that integrates safe reinforcement learning, fault detection models and formal verification tools for autonomous vehicle control. The architecture is based on the concepts of Proximal Policy Optimization and Deep Q-Networks with safety-constrained reward shaping. Running time-based anomaly detection models like Isolation Forest, Autoencoders and Deep Neural Networks to watch the behavior of the system in real time. The Z3 Solver ensures that the safety requirements, like collision avoidance and speed limits, are fulfilled. Experimental results indicate that the safe reinforcement learning model obtains 98.2 per cent collision avoidance rate and 3.1 per cent rate of safety violation. The best fault detection accuracy (97.4%) and detection latency (45 milliseconds) is achieved by the isolation forest model. Ablation analysis shows that this is a step-wise decrease of constraint violation with each safety layer. The entire design has the lowest violation rate of 3.1 percent. The results from statistical analysis of five independent runs show that the results are well converged with standard deviation of 0.6 percent collision avoidance. The safety properties are given and a formal verification is performed that the learnt policy fulfills all these properties. The results in this work suggest that the combination of safe reinforcement learning and fault detection formal verification can enhance the reliability of autonomous systems.