Data reduction and clustering techniques for fault detection and diagnosis in automotives

In this paper, we propose a data-driven method to detect anomalies in operating Parameter Identifiers (PIDs) and in the absence of any anomaly, classify faults in automotive systems by analyzing PIDs collected from the freeze frame data. We first categorize the operating parameter data using automotive domain knowledge. The dataset thus obtained is then analyzed using Principal Component Analysis (PCA) and Independent Component Analysis (ICA) for finding coherence among the PIDs. Then we use clustering algorithms based on both linear distance and information theoretic measures to assign coherent PIDs to the same class or cluster. A comparative analysis of the behavior of PIDs belonging to the same cluster can now be made for detecting anomaly in PIDs. Since a system fault is characterized by the values by of all PIDs across all the clusters, we use the joint probability distribution of the independent components of all PIDs to characterize the fault and find the divergence between the joint distributions of training and test data to classify faults. The proposed method can analyze available parameter data, categorize PIDs into informative or non-informative category, and detect fault condition from the clusters. We demonstrate the algorithm by way of an application to operating parameter data collected during faults in catalytic converters of vehicles.