• An interpretable cause tracing method is proposed for heavy rail defects. • Mechanism evidence is introduced and defect causes are organized into process-oriented categories. • A dimensional nonconformity case study and SHAP analysis validate the effectiveness and interpretability of the method. Heavy rail defects in hot rolling are characterized by complex causation, strong inter-stage coupling, and limited interpretability of diagnostic results. To address these issues, this paper proposes an interpretable cause tracing method for heavy rail defects guided by mechanism evidence. First, a mechanism-evidence feature is constructed from the relative deviation between the measured rolling force and the mechanism-based reference rolling force to characterize process-state deviations across multiple rolling stages. Second, defect causes are organized into process-oriented categories, and ME-LightGBM is developed to identify and rank candidate cause categories. Third, SHAP is introduced to decompose the model outputs and reveal the main evidence and contribution patterns underlying different cause determinations from the perspectives of global feature contribution distributions, single-feature response relationships, and representative misclassified samples. The proposed method is validated through a case study on dimensional nonconformity defects using data collected from a digital heavy rail production system. The results show that the method yields good overall performance in both cause ranking and category discrimination. The introduced mechanism-evidence feature improves the separability among different cause categories, while the interpretability analysis further extends the analysis from candidate-cause identification to key-parameter localization and evidence-based interpretation. The proposed method provides a process-oriented approach for the inspection and cause analysis of heavy rail defects.
Xiang et al. (Fri,) studied this question.