CNNs have promoted the development of intelligent fault diagnosis and improved the diagnostic ability of models. However, due to the widespread existence of complex noise in actual industrial production environments, these noises may make it impossible to fully extract fault information by relying solely on CNN. Therefore, in order to meet this challenge, this paper proposes a new mechanical fault diagnosis framework that combines multi-scale separable convolution and dilated attention mechanisms (MSCDA). This method combines the advantages of CNN in local feature extraction and the ability of Transformer to capture global information, improving the diagnostic ability of the model under noise interference. The model first uses multi-scale separable convolution (MSC) to extract detailed local features from vibration signals. Secondly, the multi-scale dilated attention (MSDA) mechanism is used to capture more and more extensive feature information. Experimental results show that, compared with existing diagnosis framework based on CNN and transformer, this methodology has higher accuracy and anti-noise capability, and can be better applied in practice.