Enhancing underwater acoustic signal prediction with adaptive double-stage decomposition and a hybrid model

Underwater acoustic signal prediction is a key underwater acoustic signal processing task. However, the strong non-linear and non-smooth nature of underwater acoustic signals leads to their high unpredictability. In this study, we propose an underwater acoustic signal prediction model based on complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), generalized link dispersion entropy (GLDE), beluga whale optimization (BWO), successive variational mode decomposition (SVMD) double-stage decomposition, and a convolutional neural network (CNN)-bidirectional long and short-term memory with temporal attention mechanism (BiLSTM-Attention). First, CEEMDAN decomposes the underwater acoustic signal into one segment to obtain several modes. Then, these modes are clustered into three classes of Co-IMFs, characterized by high and low complexity, based on the calculation of the GLDE for each mode. The Co-IMF with high complexity is decomposed in a secondary-stage decomposition using the BWO-optimized SVMD. Finally, CNN-BiLSTM-Attention is used to predict all decomposed components. The proposed model is validated using underwater acoustic signal data sourced from the National Park Service, and the experimental results demonstrate that the proposed double-stage decomposition prediction model outperforms the existing model, achieving the highest prediction accuracy and providing the best fit.