Research on fault diagnosis of three-phase induction motor bearings based on diffusion-convolutional neural network
LIU Chao
LIU Qinming
YE Chunming
WANG Yujie
Abstract：[Objective]Addressing the issue of data scarcity in bearing fault diagnosis of three-phase induction motors within industrial settings,where insufficient actual fault samples hinder the effective training of neural network models,a novel diffusion-convolutional neural network(DCNN)model was proposed.The DCNN model integrates the advantages of the denoising diffusion probabilistic model(DDPM)and convolutional neural network(CNN),thereby overcoming the limitations of conventional deep learning approaches in handling small-sample datasets.[Methods]Firstly,the DCNN model employed the Gramian angular difference field(GADF)to transform raw vibration signals into information-rich two-dimensional time-frequency images,enhancing the representational capacity of data features.Secondly,the DDPM generator network simulated the distribution of actual fault data to generate physically meaningful,high-quality synthetic samples,thus augmenting the training dataset.Furthermore,the DCNN incorporated an improved U-Net architecture as the core denoising module;through temporal encoding and conditional embedding techniques,the model's capability to recognize complex fault characteristics was strengthened.Finally,the Wasserstein distance was utilized to minimize the discrepancy between generated and real data to optimize model training,while spectral normalization was applied to enhance model stability.The CNN classifier,trained systematically thereafter,was employed for final fault diagnosis.[Results]Results demonstrate that the proposed DCNN model exhibits superior performance surpassing traditional generative models,achieving a diagnostic accuracy of 99.95%,representing a significant improvement over conventional methods.These findings validate the efficacy and excellence of the proposed model in addressing small-sample fault diagnosis challenges.
Keywords：Fault diagnosisDiffusion-convolutional neural networkDenoising diffusion probabilistic modelVirtual sample generationThree-phase induction motor bearing
Publication Date：2026-05-31
Online Publishing Date：2026-09-12(First online date of this platform, not the publication date of the document)
Pages：14( 126-139 )