Underwater Acoustic Target Recognition Based on Multi-Dimensional Feature Fusion

The accuracy of underwater acoustic target recognition depends on signal features from different dimensions. These multi-dimensional signal representations include time-domain features, spectral statistical features, time-frequency domain features, and even visual features. The integration of such diverse signal representations is beneficial for improving final target recognition performance. Against this backdrop, this paper proposes a target recognition method based on multi-dimensional feature fusion. It selects five types of underwater acoustic signal features with high discriminative power and employs Support Vector Machine (SVM) as the classifier. These five types of features include Mel-frequency cepstral coefficients (MFCCs), line spectrum features, time-domain and frequency-domain statistical features, spectrogram visual features, and pre-trained features from one-dimensional time-domain signals. Experiments demonstrate that each individual feature set contributes to the performance of the proposed underwater acoustic target recognition model, and the combination of different feature sets can further enhance model performance. The fusion of all five feature types achieves the highest recognition accuracy, reaching 95.91% on the ShipsEar dataset and 88.54% on the DeepShip dataset, outperforming any combination of four feature types.