Abstract: To address the substantial differences in feature distribution under varying operating conditions and the limited domain adaptation capabilities in the fault diagnosis of rolling bearings, this paper proposes a fault diagnosis method that integrates a Hybrid Convolutional Neural Network (Hybrid Convolutional Neural Network, HCNN) with a Deep Subdomain Adaptation Network (Deep Subdomain Adaptation Network, DSAN) to effectively adapt to diverse operating conditions. Initially, vibration signals from bearings under different load conditions are collected and categorized into source and target domains. Subsequently, a hybrid convolutional neural network architecture, incorporating multi-scale convolution and attention mechanisms, is developed to extract fault characteristics. Feature distribution alignment between the source and target domains is achieved through a Gaussian Mixture Model (Gaussian Mixture Model, GMM) combined with a Local Maximum Mean Discrepancy (Local Maximum Mean Discrepancy, LMMD) loss function. Furthermore, an innovative weighting factor is introduced to enhance the convergence efficiency of the DSAN and optimize the diagnostic process. The proposed methodology is validated using the rolling bearing dataset from Huazhong University of Science and Technology. The experimental findings indicate that this methodology significantly improves the accuracy of fault diagnosis, demonstrating impressive performance.

Key words: rolling bearing, fault diagnosis, multi-scale convolution, attention mechanism, adaptive network, local maximum mean discrepancy

摘要： 针对滚动轴承故障诊断中不同工况下特征分布差异显著及域适应能力不足的问题，本文提出了一种结合混合卷积神经网络（Hybrid Convolutional Neural Network，HCNN）和深度子领域自适应网络（Deep Subdomain Adaptation Network，DSAN）的滚动轴承故障诊断方法，以适应多种工况。首先，对不同负载情形下的轴承振动信号加以采集，将采集到的信号分为源域与目标域；然后，构建了一个基于多尺度卷积及注意力机制的混合卷积神经网络，进行故障特征提取；再通过高斯混合模型（Gaussian Mixture Model，GMM）结合局部最大均值差异（Local Maximum Mean Discrepancy，LMMD）损失函数实现源域和目标域特征分布的对齐；最后设计了一个新的权衡因子以提高DSAN收敛性能并优化进程。实验使用华中科技大学滚动轴承数据集进行验证。实验结果表明，该方法能有效提高故障诊断的精确性，呈现出优良性能。

关键词: 滚动轴承, 故障诊断, 多尺度卷积, 注意力机制, 自适应网络, 局部最大均值差异