摘要
为缓解轴承故障样本稀缺对智能故障诊断模型性能的制约,课题组提出一种融合注意力机制的改进辅助分类生成对抗网络(Auxiliary Classifier Generative Adversarial Network,ACGAN)模型,用于高质量故障图像生成与扩展。首先,将采集的一维振动信号经短时傅里叶变换(Short-Time Fourier Transform,STFT)转换为时频图像,增强其时频特征表达能力;其次,用最小二乘损失代替传统ACGAN二元交叉熵损失以解决训练过程中存在梯度消失的缺点,提升训练稳定性;最后,在判别器结构中嵌入通道与空间注意力模块(Spatial Attention Module,SAM),增强模型对关键故障特征的响应能力。实验结果表明:该方法能够有效生成高质量故障样本,提高故障诊断的精度。
In order to alleviate the limitations of intelligent fault diagnosis models caused by the scarcity of labeled bearing fault samples,the research group proposed an enhanced Auxiliary Classifier Generative Adversarial Network(ACGAN)model with attention mechanisms to enable high-quality fault image generation and data augmentation.Firstly,the collected one-dimensional vibration signals were transformed into time frequency representations using the Short-Time Fourier Transform(STFT),enhancing the expressiveness of features in both temporal and spectral domains.Secondly,the Least Squares Loss was used to replace the traditional ACGAN binary cross entropy loss to solve the shortcomings of gradient disappearance in the training process and improve the training stability.Finally,a channel and Spatial Attention Module(SAM)were integrated into the discriminator structure to increase its sensitivity of the model to critical fault-related features.Experimental results demonstrate that the proposed approach can effectively synthesize high-quality fault samples,thereby significantly improve diagnostic accuracy in small-sample scenarios.
作者
柯斌
张守京
董彬彬
许涛
KE Bin;ZHANG Shoujing;DONG Binbin;XU Tao(School of Mechanical and Electrical Engineering,Xi′an Polytechnic University,Xi′an 710600,China)
出处
《轻工机械》
2025年第6期61-68,共8页
Light Industry Machinery
基金
陕西省自然科学基金(2025JC-YBMS-406)。
关键词
轴承
故障诊断
辅助分类生成对抗网络
短时傅里叶变换
最小二乘损失
注意力机制
bearing
fault diagnosis
ACGAN(Auxiliary Classifier Generative Adversarial Network)
STFT(Short-Time Fourier Transform)
Least Squares Loss
attention mechanism
