矿用主扇风机滚动轴承的状态监测与故障诊断研究对煤矿生产安全具有重要意义。现有的滚动轴承故障诊断方法在实际工况中进行直接应用时存在训练不足、故障诊断准确率不足的问题，且矿用主扇风机滚动轴承长期处于正常运行状态，正常样本的数量远多于故障样本，即存在样本不平衡问题。因此，提出一种基于迁移学习的矿用主扇风机滚动轴承故障诊断方法（TLCNN+加权交叉熵损失）。该方法将常规滚动轴承数据作为源域数据，将矿用主扇风机滚动轴承数据作为目标域数据。首先利用对称极坐标（SDP）方法将振动信号转换为SDP图像；然后利用充足的源域图像样本对常规滚动轴承故障诊断模型进行训练，训练完成后将诊断模型的参数迁移至矿用主扇风机滚动轴承故障诊断模型中；其次迁移过程中对低层网络进行锁定并通过目标域图像样本对模型的高层网络进行微调，便可得到参数权重优化后的矿用主扇风机滚动轴承故障诊断模型。最后，为了解决样本不平衡问题，在模型中添加了加权交叉熵损失函数进行训练，使诊断模型对作为少数类的故障样本赋予更高的权重并在诊断过程中更加关注故障样本，从而提高诊断准确率。为了验证提出方法的有效性，通过常规滚动轴承故障试验台与实际工况中的矿用主扇风机滚动轴承数据进行了试验验证。结果表明所提方法可以对矿用主扇风机滚动轴承的运行状态进行准确识别分类，准确率达99.28%。

The condition monitoring and fault diagnosis of the rolling bearings of the main fan in the mine are significant to the safety of coal mine production. The existing fault diagnosis methods of rolling bearing have the problems of insufficient training and accuracy when applied directly in actual working conditions. Moreover, the rolling bearings of the mine main fan are in normal operation for a long time, and the number of normal samples is much more than the faulty samples, so there is a sample imbalance problem. Therefore, this paper proposes a fault diagnosis method for rolling bearings of mine main fan based on transfer learning. The method takes the conventional rolling bearing data as the source domain data and the mine main fan rolling bearing data as the target domain data. Firstly, the one-dimensional vibration signal is converted into two-dimensional SDP images using the SDP method, and then the conventional rolling bearing fault diagnosis model is trained using sufficient source domain image samples. After training, the parameters of the diagnostic model are transferred to the mine main fan rolling bearing fault diagnosis model, and the lower layer network is locked and the higher layer network of the model is fine-tuned by the target domain image samples during the transfer process, and finally the mine main fan rolling bearing fault diagnosis model with optimized parameter weights is obtained. Meanwhile, in order to solve the sample imbalance problem, a weighted cross-entropy loss function is added to the model for training, so that the diagnosis model gives higher weights to the fault samples as a minority class and pays more attention to the fault samples in the diagnosis process, thus improving the diagnosis accuracy. In order to verify the effectiveness of the proposed method, this paper uses a conventional rolling bearing fault test bench and the rolling bearing data of the mine main fan fan in actual working conditions for experimental verification. The results show that the proposed method can accurately identify and classify the operating status of the mine main fan rolling bearings, and the accuracy rate is 99.28%.