Traditional deep learning methods for bearing fault diagnosis rely on time-frequency transformations and exhibit inefficient feature extraction from temporal signals. An end-to-end bearing fault diagnosis method using a Local Feature Enhancement LSTM network (LFE-LSTM) is proposed to address these limitations. Specifically, the process employs a convolutional feature extraction module with dual-layer small-kernel convolutions to capture nonlinear mappings in vibration signals, while batch normalization ensures training stability. Subsequently, an LSTM network captures long-term temporal dependencies in fault information, whereas dropout layers improve generalization and robustness by preventing overfitting. Furthermore, experimental validation on ten bearing fault categories achieves an average accuracy of 99.57%. Consequently, comparative results demonstrate superior performance over 1DCNN (96.43%), standalone LSTM (98.71%), and WDCNN-LSTM (98.93%), thereby confirming the method’s effectiveness for practical fault diagnosis applications.