Time-Frequency Fault Diagnosis for Health-Aware Predictive Control of Rotating Machinery

Rotating-machinery diagnosis is often evaluated as an isolated classification task, although the operational value of a diagnostic decision depends on how it is translated into health-aware actions. This study develops a diagnosis-to-control framework that integrates multi-sensor time and frequency-domain representations, machine-leaming-based fault identification, a continuous relative health index, and fault-aware model predictive control. The public Southeast University drivetrain dataset was organized into ten bearing and gear health classes under two operating conditions. Eight synchronous channels were processed into ۲۰۰ non-overlapping ۱۵-s runs; each run was represented by ۱۰۸۰ frequency-domain and ۱۵۶۰ time-domain descriptors. The fused ۲۶۴۰-dimensional representation was augmented only within training subsets, standardized, and reduced to ۲۰ principal components. Six classifiers were evaluated under five-fold cross-validation, with the deployment model selected by cross-validation macro-F۱ rather than the locked test set. Random forest achieved the highest validation macro-F۱ (۹۳.۷۳%) and a locked-test accuracy of ۹۷.۵% with a macro-F۱ of ۹۷.۴۶%. A family-specific Mahalanobis-distance health index, fused with diagnostic fault confidence, separated most faulty runs from the healthy reference states. In closed-loop simulation, the AI-driven health-aware controller reduced mean control effort, maximum torque, and maximum torsional response by ۷.۱۳%, ۸.۳۶%, and ۹.۶۱%, respectively, relative to nominal MPC, while increasing speed-tracking RMSE by ۴۶.۱۶%. The results demonstrate an explicit and quantifiable health-performance trade-off and support diagnosis-driven control adaptation as a system-level extension of conventional fault classification.

کلیدواژه ها:

rotating machinery ، fault diagnosis ، sensor fusion ، health-aware MPC