Anomaly detection for sustainable automotive manufacturing

The automotive sector provides society with the means to move people and goods, however the increasing need for climate change mitigation places responsibility on automotive manufacturers to develop low-carbon technologies to meet these needs. These rapid technological developments are prone to high costs, skill gaps, and sub-optimal energy-intensive practices, particularly within the testing environments for automotive components. Automated safety thresholds are in place to detect large fluctuations in real-time data, whilst the detection of other, more nuanced faults is manual, but these have the risk of undermining vehicle development. As such, this process is resource intensive, such as the personnel time and energy consumed to rerun tests containing faults, and the financial costs associated with these. Here we show the ability of unsupervised, data-driven machine learning based anomaly detection methods to identify anomalous time-periods within an automotive test, including faults in either the test component or in the test facility equipment, without the need for training data. We compare the performance of three clustering algorithms –k-means, agglomerative clustering, and DBSCAN – based on their run time and ability to create defined anomalous clusters of the two anomalies. K-means was able to identify the two anomalies with eight total clusters in half the time of agglomerative clustering. DBSCAN clustered the data in half the time ofk-means however was unable to create defined anomalous clusters. These results illustrate the potential for unsupervised data-driven anomaly detection to operate within automotive manufacturer testing environments. These methods provide a low-cost digital solution to the resource demands associated with the traditional processes used by automotive manufacturers when developing sustainable transport options.