A Physics-Informed Discrepancy Diffusion Framework for Few-Shot Fault Diagnosis

In industrial fault diagnosis of rotating machinery, bearings often operate under strict reliability requirements, making labeled data under different health conditions difficult to accumulate and leading to the challenge of few-shot fault diagnosis. Generative approaches have been explored to mitigate this data scarcity. However, without explicit physical constraints, they may suffer from hallucination and produce physically implausible samples, which can undermine diagnostic reliability. Accordingly, a physics-informed discrepancy diffusion framework (PIDDF) is proposed, coupling physics-based modeling with diffusion-based generative learning to bridge the simulation-to-real gap under severe label scarcity. Specifically, PIDDF first constructs a physics-based dynamic model to synthesize physically consistent fundamental signals for different fault types. A discrepancy-aware diffusion model is then designed to learn environment-dependent residual discrepancies from a few real samples, where a spectral–temporal residual attention mechanism is embedded to enhance time–frequency feature extraction. In this way, environment-dependent uncertainty is modeled as residual discrepancies, which are sampled and superposed onto the physically consistent fundamentals to generate high-fidelity pseudoreal signals from only a few real samples. The enriched dataset enables reliable few-shot fault diagnosis. Experiments on a public bearing benchmark dataset and a self-collected long-shaft industrial bearing dataset for typical support-bearing scenarios demonstrate that PIDDF achieves consistent improvements under diverse few-shot conditions.