To address the issues of incomplete feature extraction, poor stability, and limited generalization in traditional fault diagnosis models, a model based on a multi-scale convolutional neural networks (MCNN), bidirectional gated recurrent units (BiGRU), and multi-head self-attention mechanism (MSA) was proposed. The model was designed to achieve comprehensive feature extraction from both spatial and temporal perspectives. It took raw vibration signals as input, and multi-scale features were extracted through convolution kernels of different sizes. A multi-head self-attention mechanism was used to dynamically adjust output weights, disregarding redundant information and weighting the extracted features for fusion. Then the fused features were input into a BiGRU network, which utilized a bidirectional information fusion mechanism to explore information from both past and future directions, capturing dependencies between different parts of the input sequence. Finally, Softmax was employed for classification. Experimental validation was conducted using three bearing fault datasets, and the results show that the proposed model has excellent performance metrics on different datasets and showcases good generalization and feasibility.