A new underwater acoustic signal recognition method: Fusion of cepstral feature and multi-path parallel joint neural network

With the development of underwater target stealth technology, feature extraction under low signal-to-noise ratio faces huge challenge. Aiming at feature extraction of underwater acoustic signal (UAS), An innovative UAS recognition framework integrating adaptive preprocessing, novel cepstral feature fusion, and a multi-path parallel joint neural network is proposed in this paper. To address the non-adaptive decomposition level selection issue in multivariate variational mode decomposition (MVMD), an adaptive MVMD algorithm is developed based on mode energy ratio analysis. This approach dynamically optimizes the decomposition parameter by quantifying energy distribution across intrinsic mode functions (IMFs). In feature extraction stage, triangular filter bank is replaced by Gaussian filter bank in Mel frequency cepstral coefficient (MFCC) extraction process., and MFCC with Mel-domain transformation filter (TMFCC) is proposed. For feature recognition, a multi-path parallel CNN-Transformer network (MPCT-Net) is designed. This architecture integrates the local feature learning capability of CNN with the global dependency modeling of Transformer through parallel processing branches. MPCT-Net extracts hierarchical features across multiple scales while maintaining computational efficiency. Feature verification and network verification experiments are carried out respectively on the proposed feature extraction method in two kinds of measured UAS. The results show that the recognition rates of ship-radiated noise and marine environmental noise are 98.75% and 97.75% respectively, which are all better than those of verification models. The proposed feature extraction method provides new idea and method for feature extraction of UAS under low signal-to-noise ratio.