Diesel engines serve as critical power sources across transportation and industrial fields, and their fault diagnosis is essential for ensuring operational safety and system reliability. However, acquiring sufficient and effective operational data remains a significant challenge due to the high complexity of the systems. As a modeling method that incorporates expert knowledge, the belief rule base (BRB) demonstrates strong potential in resolving such challenges. Nevertheless, the reliance on expert knowledge constrains its practical application, particularly in complex engineering scenarios. To overcome this limitation, this study proposes a reliability fault diagnosis method for diesel engines based on the belief rule base with data-driven initialization (DI-BRB-R), which aims to improve modeling capability under conditions of limited expert knowledge. Specifically, the approach first employs fuzzy c-means clustering with the Davies–Bouldin index (DBI-FCM) to initialize attribute reference values. Then, a Gaussian membership function with Laplace smoothing (LS-GMF) is developed to initialize the rule belief degrees. Furthermore, to guarantee the reliability of the model optimization process, a group of reliability guidelines is introduced. Finally, the effectiveness of the proposed method is validated through an example of fault diagnosis of the WD615 diesel engine.

The accurate diagnosis of bearing faults is paramount for the reliable operation of industrial machinery. The belief rule base (BRB), recognized for its interpretability, has consequently been extensively applied in the domain of fault diagnosis. Nevertheless, practical industrial scenarios present considerable challenges, including noise interference, the demand for detailed fault information, and the limited availability of labeled samples, which collectively constrain the applicability of BRB methods. To address these limitations, the present study introduces a semisupervised multilayer belief rule base (SM-BRB) approach. Initially, the superlet transform (SLT) is leveraged to extract salient structural features and to enhance robustness against noise. Drawing inspiration from real-world diagnostic practices, a hierarchical framework is constructed to enable refined fault identification. Furthermore, a semisupervised learning paradigm is employed to capitalize on latent information contained within a surplus of unlabeled data, thereby enabling the model to sustain high diagnostic accuracy despite the scarcity of labeled examples. Experimental findings under noisy conditions substantiate the efficacy and interpretability of the proposed SM-BRB model.