ACS-DM: Adaptive conditional sampling diffusion model for few-shot machinery fault diagnosis

Abstract

Deep learning has advanced machinery fault diagnosis, yet performance remains constrained by scarce and imbalanced labeled vibration data. We present ACS-DM, an adaptive conditional sampling diffusion framework that synthesizes frequency-faithful yet temporally diverse signals for few-shot regimes. ACS-DM couples a Nested U-Net denoiser with 1D Omni-Dimensional dynamic convolution and shuffle attention to capture multiscale temporal–spectral patterns while reducing parameters by 15–20% without loss of fidelity. Lightweight conditioning via classifier-free guidance and time-step embeddings steers generation toward class-discriminative attributes without auxiliary classifiers. Trained with the standard DDPM noise-prediction objective, ACS-DM is evaluated under 5/10/20-shot settings on CWRU, PU, and a field BJTU-RAO metro-bogie dataset. Classifiers trained solely on ACS-DM-augmented datasets and tested on held-out real signals achieve consistent gains. In the experiments on the PU dataset, with five real samples per class, accuracy reaches 98.8% under our configurations. The synthesized signals preserve diagnostic spectral characteristics while maintaining time-domain diversity, yielding robust generalization across machines and operating conditions and offering a practical route to mitigate scarcity and imbalance in industrial fault diagnosis.