To address the issue that most models are incapable of adequately capturing multi-scale features in the original vibration signal, as well as feature extraction and fault diagnosis of machinery in a high-noise environment, a neural network based on the fusion of global and local joint multi-scale CNN and a hybrid transformer is proposed. The feature information of different scales is obtained first by dilated convolution, then the feature extractor of hybrid transformer architecture is used to extract strongly robust global features while maintaining local features, the spatial features of different scales are fused, and the fused features are finally input to the newly designed structure for adaptive learning. The experimental results show that the convolutional neural network's fault diagnosis capability can be greatly improved.