Remaining useful life classification of ECUs in trucks using a transformer encoder model

Abstract

Originally developed for natural language processing, transformer models have achieved state-of-the-art results in tasks such as machine translation and text classification. This has led to increasing interest in applying the transformer architecture to sequential data across multiple other domains. This thesis takes a binary classification approach to investigate whether a transformer encoder model can be used to classify the remaining useful life of electronic control units (ECUs) in Volvo trucks. The model is trained on operational data and faults related to the ECU, to predict whether an ECU is likely to fail within the following three years. The performance of the transformer model is evaluated against traditional machine learning classifiers, including logistic regression, LGBM, Extra Trees, and Random Forest. In addition to standard metrics, a custom cost metric is introduced to reflect the real-world impact of false positives and false negatives. Results show that the transformer encoder outperforms traditional models across all evaluation metrics, particularly when used with ensemble methods. However, the transformer encoder still underperformed compared to a naive classifier on the custom cost metric. This work serves as a starting point for improving the decision-making process in ECU refurbishment.