In practical industrial scenarios, fault samples of high-power motor insulated bearings are rare, making it difficult to obtain a sufficient amount of real fault data for model training. Bearing vibration data exhibits temporal characteristics, and the electrical erosion damage of bearings is often confused with traditional fatigue failure features, significantly increasing the complexity of fault diagnosis. Moreover, traditional neural network models are unable to effectively extract features from long time-series data, leading to insufficient feature extraction. To address these issues, this paper establishes a motor insulated bearing test platform and proposes a fault diagnosis method that integrates Bi-GRU and Transformer. The bidirectional gated recurrent unit (Bi-GRU) is used to retain the local temporal correlations of fault features in both forward and backward directions, while the self-attention mechanism of the Transformer effectively captures global correlations of fault features and suppresses interference for accurate fault diagnosis. Experimental results show that the proposed method achieves a diagnosis accuracy of up to 97.08%, with training speed and diagnostic performance surpassing those of BiGRU-Attention, Transformer, BiGRU, and GRU models.