Abstract
The study of the low-frequency analysis and recording spectrum (LOFARgram) of ship-radiated noise is essential for extracting critical information, such as target motion trajectories. However, the quality of LOFARgrams often degrades due to the inherent stochasticity of ship noise and the interference of environmental noise. We significantly enhance the clarity and quality of LOFARgrams by employing the U-Net++ neural network model for preprocessing. Effective training of neural network models usually requires large datasets, but the available actual LOFARgrams are often limited and costly to collect. To ensure an adequate dataset for neural network training, this paper introduces an innovative forward model that simulates LOFARgrams from stochastic noise sources. This model uses explosive decaying cosine pulses as basic units to simulate ship noise sources and employs the KRAKEN normal mode model to simulate the underwater acoustic channel’s transfer function, thereby efficiently creating high-fidelity ship noise LOFARgrams. The forward model supplies sufficient data to train the U-Net++ neural network, enabling it to demonstrate effective recovery of LOFARgrams. Additionally, we introduce a new algorithm that utilizes data prior to the Closest Point of Approach (CPA) to predict the CPA parameters, applied to both the original LOFARgrams and those processed with U-Net++. Results indicate that predictions based on U-Net++ enhanced LOFARgrams are more accurate. Our work demonstrate the effectiveness of the forward model and U-Net++ enhanced LOFARgrams for ship-radiated noise analysis and precise prediction of target motion.