Abstract: In order to improve the performance of the ship radiated noise classification system and further improve the recognition accuracy, a method based on wavelet packet decomposition combined with multi-feature extraction in the long short-term memory (LSTM) network is proposed in this paper. This method first uses wavelet packet decomposition to extract multiple features of ship radiated noise in different frequency bands, and uses principal component analysis (PCA) for data reduction of the extracted features. By the LSTM network added with the attention mechanism algorithm the learning efficiency and recognition accuracy for radiated noise classification are improved. In order to extract the features precisely, the features in time-frequency domain and the features of wavelet transform and Mel-frequency cepstral coefficients (MFCC) of ship radiated noise are extracted in different frequency bands. Then, the performances of the algorithm for features with and without frequency band partition, multi-features and single feature, and different signal to noise ratios are compared. The experimental results show that the model based on wavelet packet decomposition and PCA-Attention-LSTM can effectively improve the performance of ship radiated noise classification and it is a feasible classification method.