机电工程技术2025，Vol.54Issue(11)：158-163,6.DOI:10.3969/j.issn.1009-9492.2025.11.030
多头注意力机制结合生成对抗神经网络的轴承故障诊断方法
Bearing Fault Diagnosis Method Using Multi-head Attention Mechanism Combined with Generative Adversarial Neural Network
摘要
Abstract
To address the difficulties of acquiring high-quality labeled bearing fault data in industrial scenarios resulting in limited diagnosis precision in traditional supervised learning method,a fault diagnosis method based on a semi-supervised generative adversarial network(SGAN)integrated with a multi-head attention mechanism(MSA-SGAN)is proposed.A 32-dimensional feature vector is extracted via 5-layer time-frequency decomposition of bearing vibration signals using wavelet packet transform.The proposed method incorporates a multi-head self-attention mechanism into the discriminator of the semi-supervised generative adversarial network,constructing an optimized framework with a generator and an enhanced discriminator.Experimental results on the Case Western Reserve University(CWRU)dataset demonstrates that the proposed method achieves a fault diagnosis accuracy of 98.9%with only 20 labeled samples,outperforming the standard SGAN and semi-supervised support vector machine(S3VM)by 3.1%and 17.5%,respectively.Comparative analysis reveals superior diagnostic accuracy over traditional supervised learning models,including extreme learning machine(ELM),backpropagation neural network(BPNN),and support vector machine(SVM),with improvements ranging from 43.1%to 66.3%.These results validate the effectiveness and superiority of the proposed model in industrial small-sample scenarios.The research provides a novel approach to mitigate the limitations of traditional fault diagnosis methods that heavily depend on labeled data,offering significant engineering application value.关键词
轴承故障诊断/半监督学习/生成对抗神经网络/多头注意力机制Key words
bearing fault diagnosis/semi-supervised learning/generative adversarial netral network/multi-head attention mechanism分类
机械工程引用本文复制引用
杨炳珍,林元浩,季丽花,常凯,夏宇栋..多头注意力机制结合生成对抗神经网络的轴承故障诊断方法[J].机电工程技术,2025,54(11):158-163,6.基金项目
浙江省"尖兵""领雁"研发攻关计划"项目(2024C01229) （2024C01229）
