Abstract: Under low frequency monitoring conditions such as low speed and heavy load,traditional vibration signals are difficult to effectively capture early micro damage characteristics due to their weak energy,susceptibility to mechanical reso nance and background noise interference.A self feature fusion fault diagnosis model based on acoustic emission signals and Spatial Group wise Enhancement Deep Convolutional Neural Network (SGE-DCNN) is proposed to address the problems of complex structure,large parameter quantity,and high computational cost in existing intelligent diagnostic models.Using the Gram Angular Field (GAF),one-dimensional non-stationary acoustic emission signals were converted into two-dimensional an gular relationship images.The dynamic features and global correlations of the temporal signals were encoded into high-dimen sional visual patterns,which constructed an image dataset rich in fault information for subsequent deep feature extraction.Fur thermore,a lightweight Spatial Grouping Enhancement (SGE) attention module is introduced into the Deep Convolutional Neu ral Network (DCNN) to adaptively enhance key fault features in both spatial and channel dimensions through grouping en hancement mechanism,while suppressing irrelevant noise interference,achieving automatic extraction and fusion of deep fea tures without manual intervention.The experimental results show that:① The proposed method achieves an average recognition accuracy of 96.54% on a test set containing multiple fault modes such as outer ring pitting and inner ring cracking.In strong noise environments (with signal-to-noise ratios as low as 2.5 dB),the model can still maintain high recognition accuracy and demonstrate strong robustness.② Due to the lightweight design of the SGE module,the model maintains high accuracy while its training time is comparable to the benchmark DCNN model,demonstrating a good balance between diagnostic accuracy and computational efficiency.The proposed method provides an effective technical approach to solve the problems of difficult signal extraction and model deployment in early bearing fault diagnosis.

Key words: rolling bearing,fault diagnosis,Gramian angular field,acoustic emission,deep convolutional neural network