摘要： 滚动轴承是机械设备中的常用部件，传统方法难以对多噪声环境下具有众多复杂特征的信号进行分类，并且常在一维数据通过经典深度学习模型进行故障诊断，无法对复杂特征进行充分提取。因此，提出一种基于ACNN-LFSwin Transformer双通道故障诊断方法分别在一维数据和二维图像上进行故障诊断。首先，将原始信号分别进行基于完全自适应指数模型分解(CEEMDAN)与STFT处理，获取模态分量(IMF)与二维图像；接着，在通道一中将CEEMDAN分解后的IMF放入基于注意力机制的卷积神经网络(ACNN)中进行特征提取；然后，在通道二中将轴承数据构成的二维图像作为局部特征提取的Swin Transformer网络(LFSwin Transformer)的输入，进行图像特征提取。最后，将两通道特征进行串联融合，进行故障诊断。ACNN运用注意力机制对信号特征进行自动权重分配，以强调关键特征；LFSwin Transformer模型在传统Swin Transformer基础上进行向量转换，将输入向量转换为图像并对其进行卷积操作，使模型在故障局部特征提取方面更具优势。所提方法分别采用CWRU数据集和帕德博恩数据集进行实验验证，实验结果表明，故障诊断准确率达97%以上。说明所提出方法不仅能对多种故障进行精确诊断，还可以有效避免复杂噪声环境的干扰。

Abstract: Rolling bearings are commonly used components in mechanical equipment, and fault diagnosis of them plays an important role in formulating a reasonable maintenance plan and ensuring the safe operation of the equipment. Traditional methods are difficult to classify signals with many complex features in multi-noise environments, and often fail to adequately extract complex features in one-dimensional data through classical deep learning models for fault diagnosis. Therefore, a dual-channel fault diagnosis method based on ACNN-LFSwin Transformer is proposed for multidimensional fault diagnosis. Firstly, the original signals are processed based on fully adaptive exponential model decomposition (CEEMDAN) and STFT to obtain the modal components (IMF) and 2D images, respectively; then, the IMF after CEEMDAN decomposition is put into the convolutional neural network (ACNN) based on the attentional mechanism for feature extraction in channel one; then, the 2D image composed of bearing data is used as the SFSwin Transformer for local feature extraction in channel two. image as the input of Swin Transformer Network for Local Feature Extraction (LFSwin Transformer) for image feature extraction. Finally, the two-channel features are fused in series for fault diagnosis.The ACNN applies the attention mechanism to automatically assign weights to signal features to emphasise key features; the LFSwin Transformer model performs vector transformation on top of the traditional Swin Transformer by converting the input vectors into an image and performing a convolutional operation on it, which makes the model more advantageous in fault local feature extraction. The proposed method is validated experimentally using the CWRU dataset and the Paderborn dataset, respectively, and the experimental results show that the accuracy of fault diagnosis reaches more than 97%. It shows that the proposed method can not only accurately diagnose multiple faults, but also effectively avoid the interference of complex noise environment.