Unsupervised Data-Driven Automotive Diagnostics with Improved Deep Temporal Clustering

• Published in: 2019 IEEE 90th Vehicular Technology Conference (VTC2019-Fall) pp. 1 - 6
• Main Authors: Wolf, Peter, Chin, Alvin, Baker, Bernard
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.09.2019
• Subjects: Automotive engineering; Data models; Decoding; Fault detection; High frequency; Time measurement; Time series analysis
• ISSN: 2577-2465
• DOI: 10.1109/VTCFall.2019.8891120

The majority of data-driven fault detection approaches for automotive diagnostics are trained in a supervised manner. However, gathering diagnostics datasets to train supervised fault detection models is still challenging. Precise correlation of input data and defect events is non-trivial and requires extensive support of domain experts. Additionally, a strong imbalance caused by rare faulty events results in models which are biased towards non-faulty events. Hence, in this work, we propose a fully unsupervised data- driven diagnostics approach to detect faults in high frequency in-vehicle data. We transfer the concept of deep embedded clustering for static data to multivariate in- vehicle time series. We extend the approach by modifying the neural network architecture and comparing three similarity measures in the clustering layer, i.e., soft dynamic time warping, complexity invariant distance, and Euclidean distance. We further introduce an adapted target distribution to tackle imbalanced datasets. Our approach is evaluated on multivariate high frequency electronic control unit data of a test vehicle to detect pre-ignitions in high pressure turbocharged petrol engines. Current state-of-the-art time series clustering approaches are used as baselines for performance comparison. The results show that our approach is able to identify pre-ignitions without labels and outperforms the baselines by 10 percent in terms of accuracy.