Coursey, Austin.; Diaz-Gonzalez, Abel.; Quinones-Grueiro, Marcos.; Biswas, Gautam. (2025).听.听OpenAccess Series in Informatics, 136, 15.听
As modern machines become more complex, it is increasingly important to detect faults early and identify exactly where they are coming from. To support new methods that can work even when data are noisy, limited, and the mathematical model is incomplete, a fault-detection competition called DX 2025 LiU-ICE was created for diagnosing problems in the air path of an internal combustion engine. This paper describes the team鈥檚 winning solution. Their system first uses a semi-supervised Transformer Autoencoder, a type of machine-learning model trained to reconstruct normal engine behavior, to spot unusual patterns that may indicate a fault. To reduce false alarms, the detected signals are then passed through a rule-based filter that checks whether the problem persists long enough to be considered real. After a fault is confirmed, four neural networks estimate features from a partial system model, and the resulting residuals, meaning the differences between expected and observed behavior, are fed into a supervised classification network that estimates which fault is most likely. On the competition data, the system detected faults correctly 87% of the time with no false alarms, and it identified the correct fault with 73.8% probability on average. When tested on new driving data not seen during training, it still detected all faults and assigned the correct fault a 66.2% probability on average, although the autoencoder produced many false alarms because it did not transfer well to the new driving conditions. The authors discuss how future work could improve this weakness.
