Deep residual networks (DRNs) are a state-of-the-art deep learning model used in the data-driven fault diagnosis field. Their especially deep architectures give them sufficient capacity to deal with very complex diagnosis issues. However, a neural network with excellent performance usually requires hundreds of thousands of parameters, which is unaffordable for use in current industrial machines due to their limited computational resources. To enable practical applications for fault diagnosis, developing deep learning methods that can perform powerfully and have an economical computational burden is necessary. This study proposes a novel bearing fault diagnosis method based on the wavelet packet transform (WPT) and a lightweight variant of DRN called a multi-branch deep residual network (MB-DRN) in order to resolve the above issues. WPT is utilized to map raw signals into the time-frequency domain, from which the MB-DRN can extract a set of robust features more easily. Additionally, MB-DRN builds several small-sized convolutional layer branches in each building block to increase the network non-linearity, the construction of layer branches can be achieved freely and this design strategy largely saves the parameter usage while approaching a stronger model’s capacity. Two rolling bearing datasets with variable operating conditions were conducted on the proposed method to validate performance. The results verify the necessity of the WPT-based data processing method and show that MB-DRN can outperform the accuracies of standard DRN with only one quarter of the parameter amount, revealing the significant potential of the proposed method for realistic industrial fault diagnosis applications.

深度残差网络 (DRN) 是用于数据驱动故障诊断领域的最先进的深度学习模型。他们特别深的架构使他们有足够的能力来处理非常复杂的诊断问题。然而，一个性能优异的神经网络通常需要数十万个参数，由于计算资源有限，在当前的工业机器中使用是无法承受的。为了实现故障诊断的实际应用，开发性能强大且计算负担经济的深度学习方法是必要的。本研究提出了一种基于小波包变换 (WPT) 和称为多分支深度残差网络 (MB-DRN) 的轻量级 DRN 变体的新型轴承故障诊断方法，以解决上述问题。WPT 用于将原始信号映射到时频域，MB-DRN 可以从中更轻松地提取一组稳健的特征。此外，MB-DRN 在每个构建块中构建了几个小尺寸的卷积层分支以增加网络的非线性，层分支的构建可以自由实现，这种设计策略在接近更强模型容量的同时大大节省了参数使用。使用具有可变操作条件的两个滚动轴承数据集对所提出的方法进行了性能验证。结果验证了基于 WPT 的数据处理方法的必要性，并表明 MB-DRN 只需四分之一的参数量就可以超越标准 DRN 的精度，