Wei He

Fault diagnosis is essential for the safe and reliable operation of industrial equipment. Deep learning models provide strong fault recognition capabilities, but offer limited transparency, while belief rule bases (BRBs) provide explicit reasoning, but face challenges in representing complex fault relationships and meaningful intermediate states. Existing knowledge distillation methods mainly focus on final outputs and do not explicitly align intermediate semantic states and diagnostic paths. Moreover, conventional classification metrics do not directly evaluate hierarchical reasoning consistency. To address these issues, this paper proposes a hierarchical semantic knowledge-distilled belief rule base (HSD-BRB). Expert-guided semantic modeling assigns explicit meanings to intermediate states, while output-, state-, and path-level knowledge distillation transfers teacher knowledge to the corresponding levels of the BRB. An expert-deviation constraint further preserves the initialized rule knowledge, and a consistency evaluation mechanism quantifies state- and path-level alignment. Experiments on the Southeast University bearing and gearbox datasets achieve mean Accuracies of 95.76% and 96.81% and Macro-F1 scores of 0.9575 and 0.9680, respectively. In the gearbox small-sample setting, the BRB is optimized using 90 labelled samples, while the teacher uses the full 890-sample training partition. The results show improved state- and path-level consistency together with traceable semantic states, rule activations, and final fault beliefs. Overall, HSD-BRB provides an interpretable framework for transferring neural knowledge into hierarchical rule-based fault diagnosis.

Safety assessment of battery thermal runaway (BTR) in new energy vehicles is crucial for ensuring operational reliability. Although the belief rule base (BRB) effectively handles complex uncertainties, its application in BTR assessments encounters two primary challenges: (1) Existing interpretability criteria and constraint strategies fail to match BTR characteristics, resulting in deteriorating parameter interpretability during optimization. (2) The absence of an effective rule hierarchy mechanism diminishes model transparency. To address these issues, an interpretable safety assessment method with hierarchical characteristics based on a belief rule base (BRB-IH) is proposed for BTR. First, interpretability criteria specific to BTR scenarios are established to regulate and preserve interpretability throughout the model's construction and inference processes. Second, rule dominance is developed to quantify the structural significance of individual rules, enabling effective hierarchical structuring. This mechanism suppresses interference from non-critical rules to enhance reasoning transparency while dynamically amplifying high-risk rules to effectively improve early warning capabilities. Third, grounded in rule dominance, an asymmetric interpretability constraint strategy is proposed to protect dominant rules while freezing dispensable rules, streamlining the optimization process. Case studies based on real BTR data prove that BRB-IH achieves prominent advantages in interpretability and prediction precision.

Health assessment is important for ensuring the safety and stability of systems. However, due to the many uncertainty challenges, there are still three problems with the robustness of belief rule base models: first, the fuzziness of expert knowledge affects the reliability of the assessment results; second, the fuzziness of input information increases the uncertainty in input transformations; and third, effective methods to measure the impact of fuzziness on robustness are still lacking. In this paper, a robust nonlinear transform BRB (RN‐BRB) health assessment method is proposed to address these issues. First, the influence of expert knowledge ambiguity on the input is mitigated by reference value optimization, which improves the reliability of the results. Second, the nonlinear input transformation is introduced to ensure the flexibility of input transformation, based on which the RN‐BRB assessment model is constructed to enhance further the robustness of the health assessment method. In addition, a robustness metric is invoked to realize the measure of model robustness under the nonlinear input transformation method, which helps to identify and improve the weak points of the model. Finally, the validity and superiority of the RN‐BRB model are verified through the health assessment case studies.

The precise diagnosis and scientific management of diabetes are highly important for improving patients’ quality of life and reducing the risk of complications. However, in actual clinical settings, diagnostic processes often face challenges, including significant individual differences among patients, complex and diverse parameters, and heterogeneity in disease progression. These challenges not only impose greater requirements on the adaptability and precision of diagnostic and therapeutic models but also highlight the need for explainable disease mechanisms and rational treatment strategies. To address these issues, this study proposes an Extended Belief Rule Base (EBRB) model based on neighborhood covering reduction, abbreviated as NCR-EBRB, for diabetes prediction and diagnosis. During the model construction phase, the Extreme Gradient Boosting (XGBoost) method is first employed for feature importance evaluation to reasonably screen key features and effectively reduce model dimensionality. In the model inference phase, the Neighborhood Covering Reduction (NCR) method is adopted to implement rule reduction in the rule base, combined with a threshold-based rule activation strategy to filter out inefficient rules, ensuring efficient reasoning processes and effective result output. During the model optimization phase, the Projection Covariance Matrix Adaptive Evolution Strategy (P-CMA-ES) is applied to optimize the parameters of the streamlined rule base, aiming to identify optimal parameter configurations for further improving model performance. Through this meticulous parameter tuning, the diagnostic accuracy is enhanced, and the robustness of the model is improved.