Entropy reduction for multi-channel feature representation: A VMD-integrated diffusion model for rotating machinery fault diagnosis

Intelligent fault diagnosis is pivotal to modern industrial systems, directly supporting equipment safety, operational reliability, and efficiency. However, existing deep learning models face two major challenges in practical deployment: (1) the scarcity of fault data and (2) the complexity of multiple fault patterns, such as compound faults and varying severity levels. Although generative models have been employed to mitigate data scarcity, most existing studies directly utilize nonlinear and nonstationary raw signals as inputs. This high entropy distribution hinders generative models from effectively capturing deep latent fault features and results in computational inefficiency. To address these challenges, this article proposes a novel multichannel fault diagnosis framework, named VMD-DDPM, which integrates variational mode decomposition (VMD) with denoising diffusion probabilistic models (DDPMs). This framework decomposes the high-entropy raw signal into four low-entropy intrinsic mode functions, effectively reducing the signal complexity and enhancing the learnability of features. Experimental results demonstrate that, with only a limited increase in input-layer parameters, VMD-DDPM markedly improves diagnostic accuracy, achieving gains of 1.46, 1.39, and 4.76% on the three datasets, respectively, under small-sample and multifault coupling scenarios. Furthermore, the framework significantly mitigates the overfitting problem caused by data scarcity and strengthens the capability for feature extraction in complex compound fault conditions.