Logo: to the web site of Uppsala University

Efficient and accurate failure log analysis would accelerate product development and ensure reliability in heavy-duty vehicle development. Meanwhile, labeled data can be costly in an industrial environment. Therefore, it is imperative to construct an unsupervised automatic failure classification system. The thesis focuses on building an automated failure logs analysis pipeline using time-dependent logs from heavy-duty vehicles provided by Scania. To reach this goal, a comprehensive literature study on unsupervised learning and multivariate time series is performed. Besides, other techniques, such as feature subset selection, dimension reduction, and clustering validation, are also explored. The thesis implement an automatic analysis pipeline composed of relevant signal identification, failure instance clustering, and anomaly analysis. This automated analysis framework improves the analysis quality, shortens the software testing period, and reduces subjective bias in the analysis result. The robustness of the pipeline is validated by testing on a wide range of failure cases from different failure families. In the experiment, hundreds of a given type of failure instance can be grouped into dozens of groups with varying patterns. One or a few major groups contain 90% of the failure instances, and the rest have small group sizes and nontypical patterns.