Hybrid CNN-LSTM Architecture with Adaptive Kalman Filtering for Enhanced Underwater Acoustic Vector Sensor Signal Processing

• Novel Hybrid Architecture: Integrates CNN, LSTM, and Adaptive Kalman Filtering (AKF) specifically for MEMS-based underwater acoustic vector sensors. • Superior Signal Enhancement: Achieves a 35% improvement in SNR and a 28% reduction in Mean Squared Error (MSE) compared to traditional Kalman filtering. • High Classification Accuracy: Reaches 92.5% accuracy in target detection, outperforming standalone deep learning models like CNN or LSTM. • Dynamic Adaptation: The AKF component provides robust state estimation by dynamically adjusting noise covariance matrices under varying underwater conditions. • Real-Time Feasibility: Comprehensive processing time analysis validates that the framework is suitable for deployment on modern embedded systems. Underwater acoustic vector sensors (AVS) face significant challenges in signal processing due to high noise levels, multipath propagation, and non-stationary signal characteristics in marine environments. This paper presents a novel hybrid signal processing architecture combining Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and an Adaptive Kalman Filter (AKF) to enhance the performance of MEMS-based underwater AVS. The proposed CNN-LSTM-AKF framework leverages CNN for robust feature extraction from noisy acoustic signals, LSTM for temporal sequence modeling, and AKF for dynamic state estimation under varying noise conditions. The CNN employs multi-scale receptive fields to capture both transient and tonal features simultaneously, the LSTM’s gating mechanism exploits periodicity patterns over 256-step temporal contexts, and the AKF provides innovation-based confidence-weighted smoothing that adapts to rapid environmental changes. Extensive simulations on three acoustic signal classes (narrowband tonal, broadband transient, and frequency-modulated chirp) with Wenz-model ambient noise and Bellhop multipath channels demonstrate that the proposed method achieves 35% improvement in signal-to-noise ratio (SNR), 28% reduction in mean squared error (MSE), and 92.5% classification accuracy compared to conventional filtering techniques. The system operates effectively in the frequency range of 20–200 kHz, making it suitable for underwater communication, target detection, and marine surveillance applications.