Mechanical intelligent fault diagnosis algorithms based on deep learning have achieved considerable success in recent years. However, degradation of the diagnosis accuracy and operation speed has been even pronounced due to the horrible working condition and increasing network depth. An improved ResNets is proposed in this paper to address the issues. The advantages of the proposed network are presented as follows. Firstly, a multi--scale feature fusion block (MSFFB) is designed to extract multi--scale fault feature information. Secondly, an improved residual block (RB) based on the depthwise separable convolution (DSC) is used to improve the operation speed and alleviate the computation burden of the network. The effectiveness of the proposed network is validated by discriminating diverse health states in the gearbox under normal and noisy environments. Experimental results show that the proposed network model has higher classification accuracy than the classical CNNs, LeNet-5, AlexNet and ResNets and has faster calculation speed than the classical deep neural networks (DNNs). Furthermore, a visual study of different stages in the network model is conducted to comprehend the operation process of the proposed model effectively.

中文翻译：

---

一种改进的多尺度特征融合深度残差网络用于旋转机械故障诊断

近年来，基于深度学习的机械智能故障诊断算法取得了相当大的成功。然而，由于恶劣的工作条件和不断增加的网络深度，诊断精度和操作速度的下降甚至很明显。本文提出了一种改进的 ResNets 来解决这些问题。提出的网络的优点如下。首先，设计了多尺度特征融合块（MSFFB）来提取多尺度故障特征信息。其次，使用基于深度可分离卷积（DSC）的改进残差块（RB）来提高运算速度并减轻网络的计算负担。通过在正常和嘈杂环境下区分变速箱中的不同健康状态，验证了所提出网络的有效性。实验结果表明，所提出的网络模型比经典的 CNNs、LeNet-5、AlexNet 和 ResNets 具有更高的分类精度，并且比经典的深度神经网络 (DNNs) 具有更快的计算速度。此外，对网络模型中的不同阶段进行了可视化研究，以有效地理解所提出模型的运行过程。