Meta-Learning Method with Improved Relation-Distribution Module for Industrial Equipment Fault Diagnosis Using Small-Sample Acoustic Signals

Abstract

: Compared to traditional machine learning and signal processing techniques, deep learning has achieved remarkable results in industrial fault diagnosis. However, the difficulty of obtaining sufficient labeled fault acoustic data in real industrial environments has led to increased interest in few-shot learning methods. Existing methods are typically constrained by their reliance on convolutional architectures that mainly emphasize local patterns, limiting their ability to capture task-specific relationships across classes. To address this challenge, a novel meta-learning method with an improved relation-distribution module is proposed for industrial acoustic signals under small-sample data conditions. The framework integrates a hierarchical convolutional feature backbone to extract multi-scale spectral representations, improving robustness to variations in operating conditions and noise. A relation-aware metric alignment module further models task-specific semantics and cross-sample dependencies, while a distribution-aware Wasserstein metric module refines class separability by emphasizing local discriminative structures. Finally, a classification module combines contextual and distribution-aware cues to ensure reliable decision making with very few training examples. Extensive experiments and ablation studies on the MIMII and PT Gear Fault datasets, together with additional validation on the CUMTB dataset, demonstrate that the proposed framework consistently outperforms compared methods under different signal-to-noise ratios and working conditions, confirming its effectiveness, robustness, and generalization capability in practical industrial scenarios.