The automotive industry is witnessing its next phase of transformation. The vehicles are getting defined by software, becoming intelligent, connected and more complex to design, develop and analyze. For these complex vehicles, prognostics and proactive maintenance has become ever more critical than before.
OEMs and suppliers analyze probable failures that a vehicle component is likely to encounter, define fault codes to identify those failures, and provide procedure or guided steps to resolve them. For smarter vehicles, it is required that vehicles be capable to catch potential problems as soon as the component’s condition starts to deteriorate and becomes a failure. These failures could be known (defined) or new (undefined). Given the vehicle development timelines and increasing complexity, many problems are not analyzed at design stage and remain undetected before production. Hence, no fault code or test case exist for them. Diagnosing such problems become very difficult, postproduction.
The aim of this paper is to propose a Machine Learning (ML) based framework which utilizes minimally labelled or unlabeled sensor data generated from a vehicle system at a given frequency. The framework utilizes an ML model to identify any anomalous behavior or aberration, and flag it for further review. This framework can be adopted on large amount of real time or time series data to identify known as well as undefined failures early. These models could be deployed on cloud or on edge (on vehicles) for analyzing real-time sensor data for a given system/component and flag any anomaly. It could further be utilized to create a part specific Predictive Maintenance (PM) model to provide proactive warnings and prevent downtime.
anomaly detection, automotive, machine learning
Chandola, V., Banerjee, A. & Kumar, V. (2009). “Anomaly Detection: A Survey”. ACM Computing Surveys, 1-72
Telgaonkar Archana H, Deshmukh Sachin (2015). “Dimensionality Reduction and Classification through PCA and LDA”. International Journal of Computer Applications (0975 – 8887) Volume 122 – No.17, July 2015
Theissler, Andreas (2013). Detecting anomalies in multivariate time series from automotive systems. Doctoral dissertation. Brunel University, London, UK.