كاربرد شبكه‌هاي بيزي در تشخيص عيب موتورهاي احتراق‌ جرقه‌اي

In modern world, alongside production, one of the most important tasks is the establishment an‎d implementation of maintenance measures to ensure user safety, service continuity, an‎d reduction of damage caused by failures. In car manufacturing, due to the close interaction between the user an‎d the product, maintenance practices are pursued with a distinct approach, relying increasingly on predictive strategies to prevent incidents that could impose severe financial an‎d human costs. Machine learning methods play a growing role in predictive maintenance. Among them, Bayesian networks (BNs) are attractive because their causal structure not only supports fault detection but also models fault propagation an‎d enables root cause analysis. In this research, a method was developed to extend a Bayesian Network for fault detection using the EngineFaultDB car-engine dataset. A domain expert first specified an initial causal skeleton; however, due to sensor limitations, no direct parents were assigned to fault nodes. To complete the causal BN, a range of statistical an‎d sensitivity-based methods were applied, including Ran‎dom Forest, XGBoost, SHAP, LASSO, Morris, mutual information, permutation importance, Spearman Correlation, MRMR, FAST, RBD, an‎d Sobol. In addition, following recommendations from prior studies, a Noisy-OR parameterization with bootstrap aggregation was implemented. Beyond this, four BN structure-learning algorithms including PC, Tabu Search, GFlowNet, an‎d MCMC were implemented an‎d eva‎luated in terms of fault-detection accuracy. These models were included both as high-accuracy, causality-agnostic references an‎d to test whether applying expert-defined causal constraints results in an accuracy trade-off. Comparative results showed that enforcing causal relations specified by the expert did not reduce accuracy relative to structure-learning algorithms. Among the sensitivity-based approaches, Sobol5, Morris3, an‎d FAST4 achieved the highest reported accuracy. Furthermore, in comparison with the published baseline on EngineFaultDB, which reported 75% accuracy, the proposed approach achieved approximately 72.8% accuracy while still preserving causal relations, very close to the peak accuracy of purely data-driven methods.

زينب مالكي , ايمان ايزدي نجف آبادي

محسن دوازده امامي

حسين فلسفين , محمدرضا حيدرپور