Waldemar Bauer

Fast and reliable detection of faults in electric motors is a critical element of properly functioning industrial systems. This paper presents a comprehensive comparative analysis of machine learning algorithms, probabilistic models and deep neural networks in the task of classifying mechanical and electrical faults in DC motors. Two different analytical approaches were proposed and verified in the study. The first, based on the extraction of structured statistical features from longer signal waveforms, was used to evaluate generative models (GNB, QDA), ensemble tree models (Random Forest, XGBoost) and Bayesian Logistic Regression. The second approach, dedicated to real-time systems, consisted of analysing raw 2-millisecond time windows using a two-channel LSTM recurrent network (analysing current and rotational speed signals) and a GNB reference model for flattened sequential data. The results of the experiments confirmed the high effectiveness of both approaches, showing the highest accuracy of XGBoost and QDA models (95%) for tabular data and LSTM networks (93%) for raw time signals. The evaluation and final selection of optimal architectures provide a foundation for further research on the rigorous quantification of epistemic and aleatoric uncertainty in industrial predictive diagnostics.

In safety-critical applications such as industrial fault diagnosis, classification accuracy alone is not sufficient — the reliability of a model’s confidence estimates is equally important. In this work, which is a continuation of research on electric motor fault classification presented in a companion paper, pre-trained probabilistic prediction models were applied to motor state classification, and their predictive uncertainty was subsequently modelled and compared using five strategies: (1) native probabilistic outputs of single models, (2) tree level variance decomposition for Random Forest, (3) bootstrap ensemble aggregation for frequentist classifiers, (4) Markov Chain Monte Carlo (MCMC) posterior sampling for Bayesian Softmax Regression, and (5) Monte Carlo Dropout for LSTM networks. These strategies were evaluated on two complementary data representations — statistical feature vectors (56 test samples) and raw signal time windows (22372 test windows) — using six metrics: accuracy, Negative Log-Likelihood (NLL), Brier Score, Expected Calibration Error (ECE), average predictive entropy, and average epistemic variance. Additionally, an Error Detection AUC metric was introduced to assess the practical utility of uncertainty for flagging misclassifications. Taking the best results from each evaluation set: on the statistical feature set, the Bayesian Softmax model achieved the best calibration (ECE = 0.060), the lowest NLL (0.257) and Brier Score (0.157), while Random Forest attained both the highest accuracy (89.5%) and the highest Error Detection AUC (0.909). On the time series set, LSTM with MC Dropout achieved the best results across all reported metrics simultaneously: accuracy of 95.0%, NLL of 0.117, Brier Score of 0.071, near-perfect calibration (ECE = 0.015), and an Error Detection AUC of 0.945. Convergence of the MCMC sampler was confirmed by R-hat values of 1.000 for all parameters and effective sample sizes exceeding 3000. The results indicate that bootstrap ensembles consistently improve uncertainty quality across all base classifiers, and that Bayesian methods provide better calibration even when accuracy is not maximal.

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.

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.

Battery engineers often work with repeated cycle-level monitoring signals that are related to ageing but are not direct capacity or resistance measurements. This paper studies how such a signal can be used to detect faster change and to decide whether the measured record is long enough to locate the transition reliably. We analyse 14 lithium-ion cell records from a processed Hawaii Natural Energy Institute (HNEI) cycle-level table, using the charging-to-discharge duration ratio (C/D) as a practical charge/discharge-duration indicator. The corresponding original HNEI measurement files were checked to improve the cell description and to examine whether a capacity-based comparison was possible. They confirm substantial capacity fade, but they also contain non-physical capacity entries near cycle 370; these entries are not used as validation of C/D transition cycles. We compare a linear reference, broken-line regression, a smoothing-spline curvature check, an aggregate Bayesian smooth-transition model, and a battery-level hierarchical Bayesian model. The hierarchical model estimates a mean battery-level transition cycle of 553.9 cycles (95% credible interval: 547.1–560.9), with substantial battery-to-battery variation (standard deviation about 142 cycles). All 14 batteries show a positive increase in the rate of change of C/D. Randomly removing about half of the measurements while retaining the full test span widens uncertainty but preserves the acceleration conclusion and nearly preserves battery ordering. In contrast, cutting off the late part of the test record strongly destabilizes transition timing and extrapolation. The approach is therefore useful as a retrospective screening and test-interpretation tool for a chosen ageing-related signal. It is not a direct capacity-knee detector, a mechanism diagnosis, or a remaining-useful-life predictor.

Reliable battery operation requires not only detecting accelerated degradation, but also knowing whether the available history is sufficient to localize that transition. We analyse a public cycle-level dataset comprising 14 reconstructed lithium-ion battery records, using the charging-to-discharge duration ratio as an empirical degradation indicator. A linear reference, continuous broken-stick regression, smoothing-spline curvature analysis, an aggregate Bayesian smooth-transition model, and a hierarchical Bayesian smooth-hinge model are compared. The hierarchical model estimates battery-specific transition midpoints and slope changes under partial pooling. Its population transition midpoint is 553.9 cycles (95% highest-density interval: 547.1–560.9), with substantial between-battery variability (posterior median standard deviation: 142.0 cycles); all batteries have posterior probability one of a positive slope increment. Removing approximately half the observations throughout the trajectories widens uncertainty but preserves the acceleration conclusion and nearly preserves battery ordering. In contrast, truncating late-life observations strongly destabilizes transition localization and extrapolative coverage. The practical implication is that distributed sparse monitoring can remain informative, whereas dense early-life data cannot replace continued observation through the post-transition regime. The framework supports retrospective, uncertainty-aware screening and confirmation of accelerated ageing in battery-monitoring data, but it is not a prospective changepoint or remaining-useful-life predictor.

This paper presents a robust framework for brushless DC (BLDC) motor fault diagnosis by comparing two data fusion strategies: decision-level stacking and data-level fusion. The study uses features extracted from current and rotational speed signals to classify four conditions: healthy, mechanical damage, electrical damage, and combined faults. Decision-level fusion employs a meta-learner to integrate predictions from specialized base models, while data-level fusion utilizes a single Random Forest classifier trained on all combined features. This comparative analysis aims to provide insights into the optimal fusion strategy for developing highly accurate BLDC motor diagnostic systems, with particular relevance for predictive maintenance applications, where early and reliable fault identification is essential.