A diffusion model with adaptive dilated convolution and timestep-aware attention for fault signal generation

Abstract Real-time intelligent fault diagnosis is crucial for the early detection of incipient faults; however, its effectiveness is often severely hindered by the scarcity of fault data in predominantly healthy operating environments. To address this data imbalance, we propose a novel conditional denoising diffusion probabilistic model, designed to generate high-fidelity fault signals to assist in improving the fault diagnostic performance. Specifically, the proposed method incorporates an adaptive dilated convolution module that dynamically adjusts receptive fields based on the input signal content, thereby enhancing the capability to capture cross-scale temporal features. To further optimize signal quality, we introduce an adaptive timestep-aware attention mechanism. This component flexibly modulates the generation process, effectively distinguishing between the establishment of global temporal sequences and the refinement of local fault details. Furthermore, to tackle the challenge of mode collapse in small-sample tasks, an intra-batch diversity regularization loss is designed to guide the network in generating distinct samples by penalizing excessive similarity. Experimental results on three public datasets demonstrate that the proposed model significantly outperforms existing approaches in generating synthetic data in terms of signal quality and diversity, leading to improved fault diagnosis accuracy and highlighting its potential for robust industrial fault monitoring.