An Investigation on Attention Mechanism–Based Long Short‐Term Memory for Gearbox Fault Diagnosis

The gearbox is widely used in various machines, and their fault diagnosis can reduce unexpected downtime and maintenance costs. Although deep learning networks (DLNs) can automatically extract features from original series data and perform diagnostic analysis, their ability to distinguish complex or imbalanced fault patterns may be unsatisfactory. This study introduces a fault diagnosis technique that integrates wavelet packet decomposition (WPD) with a long short‐term memory (LSTM) enhanced by an attention mechanism. The fault information at different scales is extracted using WPD, and the wavelet packet energy is employed as input data for the LSTM network. Subsequently, the LSTM network searches for the long‐term dependence between the fault feature information to represent the fault state. An attention mechanism is introduced, allowing the network to selectively emphasize important time steps in the input data by applying varied levels of importance, which enhances the representation of critical features. Finally, the effectiveness of the proposed approach is validated on a standardized dataset, with experimental outcomes exceeding 98.2% average precision. This indicates that the proposed approach effectively utilizes both feature extraction based on expert knowledge and DLN.