A K-SVD-Guided BiLSTM Network for Intelligent Classification of Rail Acoustic Emission Signals

In order to enhance the recognition accuracy of rail damage signals and address data imbalance in acoustic emission (AE) classification, this paper proposes a k-singular value decomposition (K-SVD) guided bidirectional long short-term memory (BiLSTM) network for intelligent classification of rail AE signals. A multi-layer wavelet packet transform with a kurtosis–energy metric is applied for effective noise suppression. The improved K-SVD then learns shared and stage-specific atoms to extract discriminative sparse coefficients, which are fed into a BiLSTM network optimized by a combined cross-entropy and focal loss (BiLSTM-CEFL). Experimental studies show that the proposed method achieves high reconstruction fidelity and superior classification performance, with the accuracy of 93.7% and AUC of 0.9619. The proposed method effectively distinguishes AE signals from different stages under the unbalanced dataset and interference, providing a robust and interpretable solution for rail structural.