Uncertainty-Aware Artificial Intelligence for Gear Fault Diagnosis in Motor Drives

TL;DR

The paper tackles uncertainty in AI-based fault diagnosis for motor-drive gear systems by adopting Bayesian neural networks (BNN) with Bayes-by-Backprop to quantify both epistemic and aleatoric uncertainty. It emphasizes probabilistic outputs and uncertainty decomposition to distinguish between data noise and model/data limitations, and validates the approach on a Gearbox Dynamics Simulator using extrinsic sensor data under noisy and unseen fault conditions. The study compares BNNs with deterministic CNN/ResNet baselines, showing that BNNs produce calibrated uncertainty estimates that reveal when new faults fall outside learned categories, and demonstrates incremental learning (BNN-2, -3, -4) to incorporate new fault classes while maintaining reliable predictions. Overall, the work provides a principled, uncertainty-aware framework for trustworthy fault diagnosis in power electronics, with implications for data collection, interpretability, and decision support in industrial settings.

Abstract

This paper introduces a novel approach to quantify the uncertainties in fault diagnosis of motor drives using Bayesian neural networks (BNN). Conventional data-driven approaches used for fault diagnosis often rely on point-estimate neural networks, which merely provide deterministic outputs and fail to capture the uncertainty associated with the inference process. In contrast, BNNs offer a principled framework to model uncertainty by treating network weights as probability distributions rather than fixed values. It offers several advantages: (a) improved robustness to noisy data, (b) enhanced interpretability of model predictions, and (c) the ability to quantify uncertainty in the decision-making processes. To test the robustness of the proposed BNN, it has been tested under a conservative dataset of gear fault data from an experimental prototype of three fault types at first, and is then incrementally trained on new fault classes and datasets to explore its uncertainty quantification features and model interpretability under noisy data and unseen fault scenarios.

Figures (9)

• Figure 1: (a) Schematic view of the categorical differences between aleatoric and epistemic uncertainty -- the former is aimed at noisy diverging data and the latter focused on missing information, (b) Overfitting issue caused by NNs due to training over limited data -- as the training data corresponding to the polynomial $y = 0.5x^2-2x+3$ is collected aimed at regressing over the true data, overfitting over minimal points can cause a large deviation from the actual model.
• Figure 2: Out of distribution (OOD) samples correspond to the unseen data/conditions, that ultimately aggravates the uncertainty in deep learning predictions.
• Figure 3: Gearbox Dynamics Simulator used for collecting fault data. Detailed setup specifications can be found in agni.
• Figure 4: Fault signatures for the same loading profile -- surface fault vs. no fault. The overlapping region for both torque as well as DC voltages can lead to over-confident decisions from AI models due to conventional point-estimate deterministic learning approaches.
• Figure 5: (a) Probabilistic assignment of neuronal weights to formalize a variational inference approach, (b) Model specifications of the designed bayesian neural network (BNN).