Reliability‐Aware Framework for Fault‐Tolerant Multi‐Sensor Fusion Using Neuromorphic Hardware‐Native Learning

ABSTRACT

As decision‐making in autonomous systems becomes increasingly decentralized, with data processing moving closer to the sensing and execution layers, the pursuit for low‐latency and hardware simplicity must not compromise reliability and fault tolerance. Memristor‐based neuromorphic hardware‐native learning offers a promising approach in this direction for handling data locally and enabling fast decisions in safety‐critical autonomous driving scenarios. However, since learning in such circuits is driven by sensor data‐generated pulses in real time that update memristive weights, the reliability of these inputs must be verified before association occurs. Otherwise, noisy or corrupted sensor inputs may produce incorrect weight updates, leading to false activations. In this work, we demonstrate the need for a reliability‐aware framework that monitors the credibility and correctness of sensor inputs arriving in the training phase. A major contribution is the development of FaultynuScenes , an imperfect multimodal dataset that captures the effects of real‐time errors in truthful sensor data points. Together with rule‐based reliability criteria, this dataset provides a low‐level concept demonstration of the proposed reliability layer and clarifies its role in fault‐tolerant hardware‐native learning. Presented PSPICE simulation results clearly show that the reliability layer reduces false activations by approximately 25%–50% depending upon various scenarios.