Transformer-Based Adaptive Line Enhancer for Passive Sonar Detection

The low-frequency narrow-band tonal components in the radiated noise of underwater targets are crucial features for passive sonar detection. Traditional adaptive line enhancer (ALE) exhibit limited performance at low signal-to-noise ratios (SNR). This paper proposes a Transformer-based adaptive line enhancer (TALE) to address this limitation. The proposed method leverages Transformer networks to enhance radiated noise signals from hydroacoustic targets in the time domain. The attention mechanism of the Transformer neural network enables the model to effectively learn both time-domain signal information and signal correlations. Simulation results demonstrate that the TALE algorithm offers significant spectral enhancement. Compared to traditional ALE and a deep-learning-based line enhancer (DLE), this algorithm can effectively improve the SNR of ship-radiated noise signals by 14 dB and 11 dB, respectively, under very low SNR conditions of -30 dB.