基于CWT-ECCN的滚动轴承复合故障诊断

Compound Fault Diagnosis of Rolling Bearings Based on CWT-ECCN

摘要: 针对滚动轴承出现多种损伤并存的复合故障时，其故障特征相互耦合，导致单一时域或频域分析方法表征能力不足的问题，提出了一种基于连续小波变换与改进卷积胶囊网络的复合故障诊断框架(CWT-ECCN)。该方法首先采用连续小波变换将一维振动信号转换为二维时频图；其次，构建了一个融合轻量化卷积与双重注意力机制的高效特征提取前端；然后，设计了具备长程依赖建模能力的胶囊结构，并基于自注意力路由算法计算胶囊层之间的相关度，从而实现故障特征的分类。在HUST与SEU两个公开数据集上，该方法于噪声干扰与变工况下的平均诊断准确率分别达到98.77%和99.23%。结果表明，该方法能够有效识别复合故障，并展现出良好的鲁棒性与泛化性。

Abstract: To address the problem that fault features become coupled when rolling bearings suffer from compound faults with multiple coexisting damages, making single time-domain or frequency-domain analysis insufficient for effective representation, a compound fault diagnosis framework based on Continuous Wavelet Transform and an Enhanced Convolutional Capsule Network (CWT-ECCN) is proposed. First, the one-dimensional vibration signals are transformed into two-dimensional time–frequency representations using CWT. Then, an efficient feature extraction front-end is constructed by integrating lightweight convolutions with dual attention mechanisms. Furthermore, a capsule structure capable of modeling long-range dependencies is designed, where the correlations between capsule layers are computed through a self-attention routing algorithm to achieve fault feature classification. Experiments on two public datasets, HUST and SEU, show that the proposed method achieves average diagnostic accuracies of 98.77% and 99.23% under noise interference and varying working conditions, respectively. The results demonstrate that the method can effectively identify compound faults and demonstrate superior robustness and generalization.