This paper investigates stable and compact diagnostic feature signatures for faults related to demagnetization in brushless direct-current (BLDC) and permanent-magnet synchronous motor (PMSM) drives. Discovery analysis on the public DUDU-BLDC v1 benchmark—where DUDU is the project name derived from the Polish Diagnostyka Uszkodzeń i Degradacji Urządzeń—compares current, speed, and combined representations under explicit top-k budgets using ReliefF, minimum-redundancy maximum-relevance (mRMR), least absolute shrinkage and selection operator (LASSO), and Bayesian automatic relevance determination (ARD) logistic ranking. The revision is accompanied by DUDU-BLDC 1.5, a new and previously unpublished March 2026 dataset comprising 50 recordings from five physical motors under altered acquisition conditions. On this second campaign, leakage-free nested five-fold recording-grouped validation with three deterministic repeats reached balanced accuracies of 0.760 and 0.758 for the two confirmatory within-motor tasks. Motor-held-out balanced accuracies fell to 0.500 and 0.554, with four of five physical-motor estimates at chance and one estimate at 0.663, exposing substantial between-motor heterogeneity. A paired experiment that quantized the original raw current signals to the approximately 0.08 A resolution of DUDU-BLDC 1.5 produced a mean absolute balanced-accuracy change of 0.0046, although the maximum change was 0.0725 and individual ranking-stability changes were larger. The results support compact signatures for repeated monitoring within an established or calibrated motor population and show aggregate robustness to reduced current resolution. They do not establish universal transfer to unseen motor instances; broader deployment requires motor-specific calibration or more diverse multi-motor training data.
Industrial diagnostic systems increasingly combine machine-learning classifiers, temporal models, uncertainty surrogates, and explanation methods. Their operational use requires a separate mechanism that determines whether the available evidence is valid, mutually consistent, sufficient for a recommendation, or should instead lead to abstention or human escalation. This paper introduces the Diagnostic Evidence Governance Agent (DEGA), a deterministic agent that separates evidence generation from workflow governance. DEGA operates on immutable evidence objects, follows an explicit finite-state workflow, applies a higher-priority SafetyGuard, and records a hash-linked audit trace supporting deterministic replay. The architecture is demonstrated in an offline case study using eight DUDU-BLDC acquisitions, 200 non-overlapping windows, 28 current- and speed-based features, three classifiers, and four spline representations. Across 16 dependent acquisition-disjoint assignment views, mean acquisition-level macro-F1 was 0.875 for Logistic Regression and 0.667 for both tree-based classifiers. The pipeline produced 768 complete EvidenceBundles with reproducible scientific hashes. Explanation rankings were assignment-sensitive, with a mean top-10 Jaccard similarity of 0.318. Twelve representative DEGA executions ended in three escalations and nine no-decision outcomes, with complete audit chains and deterministic replay. The study demonstrates auditable, fail-closed governance of AI-supported diagnostic evidence without delegating workflow control to an AI agent.
This paper investigates stable and compact diagnostic feature signatures for faults related to demagnetization in BLDC/PMSM drives. Discovery analysis on the public DUDU-BLDC benchmark compares current, speed, and combined representations under explicit top-k budgets using ReliefF, mRMR, LASSO, and Bayesian automatic relevance determination logistic ranking. A second March 2026 experimental campaign, released as DUDU-BLDC 1.5, comprises 50 recordings from five physical motors under altered acquisition conditions. On this second campaign, five-fold recording-grouped validation with three deterministic repeats reached balanced accuracies of 0.806 and 0.779 for the two confirmatory within-motor tasks. Motor-held-out balanced accuracies fell to 0.512 and 0.604, with most physical-motor estimates near chance, exposing substantial between-motor heterogeneity. A paired experiment that quantised the DUDU-BLDC raw current signals to the approximately 0.08 A resolution of DUDU-BLDC 1.5 produced a mean absolute balanced-accuracy change of 0.0046, although the maximum change was 0.0725 and individual ranking-stability changes were larger. The results support compact signatures for repeated monitoring within an established or calibrated motor population and show aggregate robustness to reduced current resolution. They do not establish universal transfer to unseen motor instances; broader deployment requires motor-specific calibration or more diverse multi-motor training data.
Industrial diagnostic systems require a governance mechanism that determines whether heterogeneous evidence is admissible, mutually consistent, and sufficient for an automatic action. This paper presents the Diagnostic Evidence Governance Agent (DEGA), a deterministic governance layer built from explicit finite-state machine states, replaceable routing policies, an authoritative SafetyGuard, and hash-linked audit with deterministic replay. The bounded DUDU-BLDC case study retains diagnostic evidence from eight acquisitions and extends the evaluation to governance-profile sensitivity and state coverage. Across 15,360 retained real-evidence case-policy routes, no automatic recommendation was issued; 9216 routes reached Decision Check; and the mandatory-explanation rule triggered 7904 times. Under predeclared contract-satisfied simulated profiles, 2924 recommendations and 12,436 escalations were produced, with no recommended route violating the internal governance-consistency proxy. The results show that explanation availability is a binding governance condition and that recommendation-path reachability can be evaluated without weakening the remaining SafetyGuard gates; the simulation is restricted to explanation availability and admissibility metadata and does not validate feature attribution, recommendation correctness, or industrial deployment.