Physics-Validated Explainable Ensembles for Rotating Machinery Fault Diagnosis

This registration pre-registers the fault-signature expectation table used to validate SHAP-based explainability results in a study of machine-learning fault diagnosis for rotating machinery (MaFaulDa and CWRU datasets). For each fault class, the table specifies, in advance of any model training or explainability analysis, the vibration-spectrum features (characteristic bearing frequencies, shaft-order harmonics, and their sidebands, tiered as primary or secondary) that a competent vibration analyst would expect a trustworthy model to rely on. The table was reviewed and approved on 27 July 2026 by a Chartered Engineer with over 30 years of experience in the oil and gas industry, including condition-monitoring system engineering for compressors and turbo-expanders. Registering this table before any SHAP attribution results are computed establishes an independent, verifiable timestamp demonstrating that the evaluation criteria were fixed prior to seeing model outputs, so that reported alignment between model explanations and physical fault signatures cannot be a post-hoc artifact of choosing expectations to match results. The accompanying experimental protocol document specifies the full evaluation design, including condition-aware train/test splitting protocols, the model suite, and the alignment-scoring rule applied against this table.