Traditional multi-channel automatic identification system (AIS) blind signal separation algorithms often suffer from unstable performance under low signal-to-noise ratio (SNR) conditions. Additionally, in real-world communication scenarios, the number of source signals is frequently unknown. To address these challenges, this paper proposes an improved joint approximate diagonalization of eigenmatrices (JADE) algorithm with preprocessing steps aimed at enhancing robustness and stability.

First, the received signal is preprocessed using the singular spectrum analysis (SSA) algorithm to reduce noise. This is done by decomposing and reconstructing the trajectory matrix of the time series, effectively suppressing Gaussian white noise while preserving key features of the signal. Next, the minimum description length (MDL) algorithm is employed to estimate the number of source signals in the processed mixed matrix. By analyzing the eigenvalue distribution of the covariance matrix, we overcome the limitation of traditional algorithms that assume the number of source signals is known, thus enabling the adaptive identification of unknown source numbers. Finally, to address the issue of imprecise diagonalization measurement in the objective function of the traditional JADE algorithm, we propose an improved JADE optimization algorithm for separating AIS mixed signals. This method reconstructs an optimized objective function with scaling invariance and incorporates a column-cyclic iteration strategy, enhancing the accuracy of the joint diagonalization process while reducing computational redundancy.

By simulating the improved JADE algorithm, its performance was compared with that of the traditional JADE algorithm, the fast independent component analysis (FastICA) algorithm, the robust independent component analysis (RobustICA) algorithm, and the Information Maximization (Informax) algorithm. The results show that, when separating 2-, 3-, and 4-channel AIS observation signals, the improved JADE algorithm achieves the highest correlation coefficient, indicating the strongest correlation between the separated signal and the source signal. This coefficient exceeds 0.8015 across all SNR ranges from 0 to 20 dB. Additionally, in terms of crosstalk interference suppression, this method demonstrates superior performance, with suppression values consistently below 0.164 across all SNR ranges from −10 to 20 dB. Furthermore, the algorithm's running time is reduced by 30.38% compared to the traditional JADE, meeting the real-time requirements of satellite-based systems.

The improved JADE optimization algorithm offers significant advantages in terms of separation accuracy, algorithm reliability, and computational stability. The research findings provide a theoretical basis and technical guidance for enhancing the separation performance and real-time capabilities of AIS receivers, as well as optimizing the design of maritime communication systems in practical engineering applications.

In high-density traffic areas, ship collision automatic identification system (AIS) signals often occur, so high requirements are put forward for the separation performance and real-time performance of the receiver.

For mixed signals with different signal-to-noise ratios (SNR), a separation algorithm S-RICA based on singular spectrum analysis (SSA) and robust independent component analysis (RobustICA) is proposed. The Hankel matrix of the single-channel AIS signal is processed by singular value decomposition and reconstructed by time series respectively, SSA is used to replace whitening pre-processing in traditional independent component analysis (ICA), and the optimal step size of each iteration of the separation matrix is calculated using the kurtosis contrast function to quickly obtain the optimal separation matrix.

The simulation results show that the signal mean squared error (SMSE) value of S-RICA is stable at about 1.5 when the signal length changes, while the SMSE of fast independent component analysis (FastICA) is very unstable. S-RICA has a bit error rate of 0.97×10⁻² to 1.97×10⁻² at a SNR of 0 dB to 9 dB, an order of magnitude improvement over RobustICA and FastICA, and an improvement of 4 dB to 6 dB over S-FICA at an SNR of 0 dB to 7 dB. The average calculation time and number of iterations of S-RICA are about 18.5 ms and 13.6 times respectively, showing obvious advantages.

When the sample size and SNR change, S-RICA shows better separation performance. The results of this study have certain reference value for the application of S-RICA in future satellite-based AIS systems.