A Hybrid LSTM–Transformer Framework for Rolling Bearing Fault Diagnosis with Time-Frequency Feature Learning

Abstract

Rolling bearing fault diagnosis is essential for ensuring the reliability and safety of rotating machinery in nuclear industry applications. To improve diagnostic accuracy and robustness in noisy diagnostic environments, this paper proposes a hybrid deep learning model combining Long Short-Term Memory (LSTM) and Transformer networks. Continuous Wavelet Transform (CWT) is first employed to convert raw vibration signals into time-frequency representations for feature extraction. To reduce potential information leakage caused by overlapping sliding windows, the original continuous vibration records are divided into training, validation, and testing subsets before window segmentation. The generated CWT images are then reshaped into sequential representations and fed into the LSTM module to capture temporal dependencies and fault evolution characteristics, while the Transformer module further enhances global feature learning through a self-attention mechanism. Experiments are conducted on the Case Western Reserve University (CWRU) bearing dataset under different noise conditions. The proposed model achieves competitive diagnostic performance compared with conventional CNN, ResNet, LSTM, and Transformer models under the adopted experimental setting. The results demonstrate that the proposed approach provides high accuracy, robustness, and computational efficiency for rolling bearing fault diagnosis.