A Novel Unsupervised Sound Based Vehicle Fault Anomalous Detection Method with Adversarial Defense Mechanisms

Abstract

Safety, as one of the key consumer issues in modern vehicles, benefits significantly from the new hardware components. There are three main factors in the new perception-based vehicle safety paradigm: perception ability, working environment, and application scenario. Traditionally, the safety problem is generally tackled by considering expertise knowledge of problems is well known. In this paper however, the research problem is formulated as an unsupervised anomalous sound detection (ASD) problem with unknown problem knowledge. In this paper, a novel four-step procedure is presented to tackle such problem including multiple streams of signal representation, onboard anomaly event candidates list generation, cloud-based anomaly event recommendation and bigdata driven anomaly event detection accordingly. The detection system's robustness is enhanced against adversarial examples (adv), which pose a growing threat to audio perception systems. Our approach not only detects anomalies in vehicle operation sounds but also bolsters the model's reliability against adversarial attacks, offering a comprehensive solution for modern vehicle safety.