Prediction of Failures in Air Pressure System: A Semi-Supervised Framework Based on Transformers

• Published in: IEEE International Conference on Industrial Informatics (INDIN) pp. 1 - 5
• Main Authors: Farea, Shawqi Mohammed, Mumcuoglu, Mehmet Emin, Unel, Mustafa, Mise, Serdar, Unsal, Simge, Cevik, Enes, Yilmaz, Metin, Koprubasi, Kerem
• Format: Conference Proceeding
• Language: English
• Published: IEEE 18.08.2024
• Subjects: air pressure system (APS); Atmospheric modeling; Costs; Data mining; Data models; fault detection; Feature extraction; heavy-duty vehicles (HVDs); Informatics; Maintenance; Predictive maintenance; Roads; semi-supervised transformer; Transformers
• ISSN: 2378-363X
• DOI: 10.1109/INDIN58382.2024.10774220

The air pressure system (APS) plays a prime role in pressurizing various subsystems of heavy-duty vehicles (HVDs). However, its reliability is crucial to ensure uninterrupted operation where failures in APS lead to HVDs being stranded on the road with the manufacturers and operators incurring associated high costs. This paper addresses the problem of predicting failures in APS using a semi-supervised transformer-based framework. The proposed framework commences with important preprocessing steps including data segmentation followed by sliding windows to handle the big raw data, and subsequent extraction of distinctive features. Using these features, the transformer model was trained to reconstruct data from healthy vehicles (i.e., vehicles without any APS failures) to capture the normal behavior of the healthy vehicles. At inference, the trained model distinguished the faulty vehicles with detected APS failure from the healthy ones based on their reconstruction errors. This semi-supervised formulation of APS failure detection overcomes limitations such as the imbalanced data issue and anomaly heterogeneity that are associated with the conventional supervised formulation. The model demonstrated robust performance with an F1 score of approximately 0.76, an accuracy of about 85%, and a high recall of 0.833, indicating successful detection of most faulty vehicles. Such advancements promise significant improvements in vehicle diagnostics and predictive maintenance.