Deep Convolutional Autoencoder for Assessment of Drive-Cycle Anomalies in Connected Vehicle Sensor Data

TL;DR

The paper tackles unsupervised fault detection in hybrid-electric vehicle drive-cycle sensor data by training a fully convolutional autoencoder on normal drive cycles to learn typical powertrain patterns. Anomalies are identified via reconstruction error, with the model demonstrating strong discrimination between normal and fault-related drive cycles and outperforming traditional outlier detectors and non-convolutional baselines; MSCRED underperforms on this dataset. Experiments on drive-cycle data show high accuracy in both per-cycle and per-window evaluations, supporting the approach's viability for on-board diagnostics and predictive maintenance. The work also includes cross-dataset comparisons (Kaggle anomaly data) to illustrate generalization, and discusses practical implications for automotive engineering and fleet monitoring.

Abstract

This work investigates a practical and novel method for automated unsupervised fault detection in vehicles using a fully convolutional autoencoder. The results demonstrate the algorithm we developed can detect anomalies which correspond to powertrain faults by learning patterns in the multivariate time-series data of hybrid-electric vehicle powertrain sensors. Data was collected by engineers at Ford Motor Company from numerous sensors over several drive cycle variations. This study provides evidence of the anomaly detecting capability of our trained autoencoder and investigates the suitability of our autoencoder relative to other unsupervised methods for automatic fault detection in this data set. Preliminary results of testing the autoencoder on the powertrain sensor data indicate the data reconstruction approach availed by the autoencoder is a robust technique for identifying the abnormal sequences in the multivariate series. These results support that irregularities in hybrid-electric vehicles' powertrains are conveyed via sensor signals in the embedded electronic communication system, and therefore can be identified mechanistically with a trained algorithm. Additional unsupervised methods are tested and show the autoencoder performs better at fault detection than outlier detectors and other novel deep learning techniques.

Figures (3)

• Figure 1: The reconstruction error of our trained autoencoder is shown for different sampling regimes. The left figure shows the results of training and validation reconstruction error of our autoencoder using a batch-size of 256 and 128 time-step cropped drive cycle samples. The test reconstruction error is calculated on a separate data set sampled the same way. The result shows a similar range of reconstruction error of the normal drive cycles used for training and validation, and a higher reconstruction error of the faulty drive cycles. In the middle plot, we show the average reconstruction error using the same function in Equation 1, this time testing subsequent 128 time-step samples of entire drive cycles for each of the three data sets . On the right, the plot shows the range of reconstruction error calculated per individual 128-observation sample in entire data set without the cropping used previously.
• Figure 2: Reconstruction error $J(X,X')$ calculated using our autoencoder trained on a Kaggle anomaly data setalexander_scarlat_anomaly_2021 showing results for batched data used in training (left) and re-calculated using individual samples (right).
• Figure 3: Detecting outliers in the drive cycle data set using the methods and performance metrics proposed for various classifiers in the PyOD library: ABOD (left), PyOD autoencoder (middle), and KNN (left), drive cycles.