Postprocessing of Autoencoder Reconstruction Error for Detection and Diagnostics of Faults in Infrequently-driven Ground Vehicles

We investigated the detection and classification of engine and transmission faults in infrequently-driven ground vehicles using data-driven methods based on neural network autoencoders. The data came from seventeen vehicles, each with an engine-related or a transmission-related maintenance event. The vehicles had months to years of sensor controller area network (CAN) bus data sampled at 1Hz. Separate autoencoder models were trained for each vehicle to improve detection sensitivity. The paper investigates several condition indicators (CIs) derived from autoencoder reconstruction error, each computed from a sequence of the reconstruction’s mean absolute error (MAE). These CIs were compared using a performance metric computed as the area under the Pareto front with respect to normalized detection horizon and normalized baseline-relative CI margin. A novel detection procedure, consistent detection, effectively filtered out short-duration isolated spikes, likely false positives, while also increasing sensitivity to more plausible anomalies. In addition, the initial development of data-driven diagnostics, based on a novel approach of classifying full reconstruction error vectors associated with the fault state, showed promise but failed our robustness checks.