Comparison of multiple AI models for failure classification in smart plug top case study

Smart plug tops (SPTs) with sensing capabilities are increasingly important for real-time monitoring and diagnostics in internal combustion engines. However, the deployment of numerous electronic devices and the increasing system complexity can lead to multiple types of failures that must be accurately investigated and categorized. This research presents a machine learning (ML) based approach for the categorization and prediction of various failure modes occurring in SPTs. The method involves systematic collection of sensor data during the testing phase of SPTs, which is then linked to failures identified through lifetime analysis. The ML model is trained using relevant features extracted from the acquired data, such as voltage levels, charge times, current levels, and other electrical parameters characterizing the SPT's operating behavior. The model is refined using a training and validation method to accurately predict various types of failures, such as electric discharge on the transformer secondary winding, damping diode breakdown, and short circuits between windings. A major challenge addressed in this work is the limited number of failure samples, since the device predominantly operates under normal conditions and only occasionally exhibits faulty behavior. Hence, an upsampling technique was applied to improve this imbalanced dataset. Various Artificial Intelligence (AI) models, including Machine Learning and Deep Learning were compared with each other to find out the most appropriate one for this particular case. The best classification algorithm achieves high accuracy along with good precision, recall, and F1-score on the test data. The results demonstrate the potential of ML-based analysis to enable the early identification of problem symptoms during acceptance testing and to provide a probabilistic classification of different failure types, thereby supporting predictive maintenance and reliability assessment of SPTs. • Using electrical test parameters enables identification of specific failure modes in Smart Plug Tops • Class imbalance strongly limits fault detection and requires upsampling of minority classes • Random Forest achieves the most reliable fault classification compared to other ML models • Cross-validated training confirms the robustness of the selected model • ML-based fault prediction supports improved diagnostics and reduces production returns