Physics-Guided Compositional Diagnosis of Unseen Compound Faults in Variable-Speed Induction Motors

Compound faults in electric motors are difficult to diagnose because simultaneous-fault recordings are scarce and fault multiplicity is unknown at inference. We propose a compound-sample-free, cardinality-free framework for induction motors running continuously varying speed profiles at several load levels. Synchronized key-phase, triaxial vibration and three-phase current signals are converted to order-domain representations by anti-aliased computed order tracking, augmented by a band-pass envelope order spectrum and normalized with a scale-invariant, noise-floor-removed representation. Nine fault primitives are evaluated by modality-specific experts, combined through physics-regularized routing, and decoded by maximum a posteriori inference over 19 feasible machine states. Six leave-one-speed-load-combination-out folds and three seeds evaluate every held-out recording. Exact condition accuracy counts a recording as correct only when the predicted set of fault primitives matches the true set exactly; exact compound recovery applies the same criterion to the unseen compound recordings, requiring both constituent primitives and nothing else. Without a fault-count prior, the pipeline reaches 76.7% and 45.7%, against 62.5% and 11.1% for a conventional order-domain front end and 59.2–62.0% and 0.0–2.5% for three re-implemented baselines. A source-domain modality-selection control reduces compound recovery from 59.3% to 27.8%. Physically aligned representation, sensing specialization and cardinality-aware decoding are therefore critical to compositional motor-fault diagnosis.