Recently, deep learning methods such as convolutional neural networks (CNN) have been widely used in fault diagnosis of rotating machinery. However, most methods are not designed to consider the influence of the working conditions and limited to classifying several fault types. In this paper, we propose two CNN-based multitask models, i.e., sigmoid multitask fault diagnosis model and multisoftmax multitask classification model, which can classify the fault types and the working conditions of the signal simultaneously. All samples are jointly learned at the shared network layers, and different types of learning tasks are completed at different subnetwork layers. Results show that the proposed sigmoid multitask fault diagnosis model achieves the overall classification accuracy of 96.8% when testing the CWRU bearing dataset. The proposed multisoftmax multitask classification model is used to classify different working conditions and fault types of planetary gearbox. With frequency signals as inputs, the accuracy rate is able to reach 96.2%, and the classification accuracy of gear and speed conditions reaches more than 99%. Additionally, the gradient-weighted class activation mapping (Grad-CAM) method is used to visualize the weight vectors of different convolutional layers and locate the signal segment of interest to the model.

近年来，诸如卷积神经网络（CNN）之类的深度学习方法已广泛用于旋转机械的故障诊断中。但是，大多数方法并非旨在考虑工作条件的影响，而仅限于对几种故障类型进行分类。本文提出了两种基于CNN的多任务模型，即S型多任务故障诊断模型和multisoftmax多任务分类模型，它们可以同时对信号的故障类型和工作条件进行分类。在共享网络层共同学习所有样本，并且在不同子网层完成不同类型的学习任务。结果表明，在测试CWRU轴承数据集时，所提出的S形多任务故障诊断模型可达到96.8％的总体分类精度。所提出的multisoftmax多任务分类模型用于对行星齿轮箱的不同工作条件和故障类型进行分类。使用频率信号作为输入，准确率可以达到96.2％，齿轮和速度条件的分类准确度可以达到99％以上。此外，梯度加权类别激活映射（Grad-CAM）方法用于可视化不同卷积层的加权向量，并为模型定位感兴趣的信号段。