<p>Underwater acoustic signal classification, the challenge lies in effectively recognizing various types of vessels under complex and noisy marine conditions. To tackle this issue, this study introduces a multi-head attention-based feature fusion to enhances the effectiveness of UATR systems. The proposed technique integrates multiple acoustic features such as one-dimensional Convolutional Neural Networks (1D–CNN), Short-Term Fourier Transform (STFT), Mel Frequency Cepstral Coefficients (MFCC), and cepstral analysis. The 1D–CNN captures local spatial temporal patterns from raw acoustic signals, while STFT and cepstral analysis provide time-frequency information crucial for analyzing transient and continuous sound characteristics. MFCC offers compact representations of the spectral envelope, aiding in distinguishing between different vessel types. These features are fused using a multi-head attention mechanism, which emphasizes the most relevant information from each feature type, thereby improving the quality of the combined representation. An XGBoost classifier is employed for the final recognition stage. Experimental findings showed that this proposed model achieved a accuracy of 81.54%, a precision of 83.88%, a sensitivity of 81.54%, and an F1 score of 80.97%. These results demonstrate that leveraging diverse feature sets combined with attention-based fusion significantly enhance the performance of UATR systems. Although the proposed method is applied to underwater vessel recognition, it is also applicable to other acoustic signal classification tasks, including marine mammal monitoring and underwater acoustic event detection.</p>