An Active Echo Target Detection Method Combining Transformer and Multi-dimensional Feature Fusion

To address the challenges of unified modeling for heterogeneous features and unstable detection performance under strict false-alarm constraints in active sonar target detection, this paper proposes a window-level method that integrates multidimensional feature extraction with Transformer-based fusion. Seven complementary features—namely, raw waveform (RAW), wavelet-modulated frequency (Wavelet-MF), CFAR output, short-time Fourier transform (STFT), smoothed pseudo-Wigner–Ville distribution (SPWVD), discrete wavelet decomposition (DWT), and mel-frequency cepstral coefficients (MFCC)—are extracted from beamformed echoes and temporally aligned via one-dimensional convolution and linear projection prior to joint temporal–feature modeling by the Transformer. Experiments on sea-trial data (10 kHz sampling rate, 1.2 s analysis window) show that Transformer-Small and Transformer-Base achieve F1 scores of 0.9575 and 0.9606, respectively. When the window-level false alarm rate is constrained to below 0.001, their recall values remain at 0.9392 and 0.9453, and their expected calibration errors (ECE) are 0.0055 and 0.0056, respectively. Ablation studies further demonstrate that the proposed fusion scheme effectively aggregates the most informative modalities—particularly MFCC, DWT, SPWVD, and RAW—whereas non-Transformer deep baselines consistently underperform the fused Transformer model. Thus, the proposed method provides a robust and reliable solution for active sonar target detection in complex acoustic backgrounds.