Haoran Ji

This study investigates how ocean fronts affect underwater sound propagation in the shallow waters surrounding Hainan Island. To characterize the regional environment, HYCOM oceanographic fields and ETOPO1 global bathymetry are transformed from geographic coordinates into Lambert conformal conic (LCC) coordinates to construct a regional environmental database. Building on this database, a Vertical Depth Fine Interpolation Method (VD-FIM) is developed to refine sound-speed profiles and to generate an ocean-acoustic dataset for sound-field simulations. Using this dataset, transmission-loss calculations indicate that seasonal variability, regional circulation, and atmospheric forcing drive strong interactions between cold and warm water masses. These interactions form pronounced fronts that add about 10–20 dB of transmission loss and significantly modify the shallow-water waveguide. In autumn and winter, weaker vertical sound-speed gradients are more favorable for long-range propagation, and sources placed in cold-water masses exhibit improved detection performance at long ranges. Acoustic measurements collected in the South China Sea in November 2023 are consistent with the modeled transmission loss. This consistency indicates that the proposed environmental-modeling and data-processing framework can quantify front-induced acoustic effects and guide the selection of detection periods and source locations. Code for this study is available at GitHub: https://github.com/jxk2026/Hainan-Shallow-Water-Acoustics.

Although beamforming design enables port-starboard discrimination to some extent, it performs poorly under low signal-to-noise ratio (SNR) conditions and exhibits low port-starboard suppression, leading to potential misclassification. To address the challenge of port-starboard discrimination in underwater acoustic detection using dual linear arrays, this paper proposes a novel discrimination algorithm based on dual linear array processing. The proposed algorithm first utilizes the Linearly Constrained Minimum Variance (LCMV) criterion to construct a beamforming vector and calculate the spatial spectrum of the received data, thereby obtaining a coarse estimation of the target's azimuth. Building on the beamforming process, we construct a dictionary matrix to establish the relationship between the spatial spectrum and the spatial power distribution vector. Furthermore, to mitigate the impact of sidelobes in the spatial spectrum, we design a set of projection operators to process the spatial spectrum and dictionary matrix, thereby obtaining a cleaner spatial relationship. In addition, to enhance numerical stability, we incorporate a regularization term into the objective function using the noise subspace. Simulations and experimental demonstrate that the proposed algorithm achieves an exceptionally high port-starboard suppression ratio and superior azimuth resolution capability.

In order to achieve efficient power transmission between the transducer and the power amplifier, and to solve the problem of acoustic wave distortion in the transmission, this paper proposes a tunable matrix capacitive impedance matching network (TMCIMN) for the Electroacoustic Transduction System (ETS). Firstly, the circuit model of the giant magnetostrictive transducer (GMT) was analyzed, and a method for fitting its load impedance was proposed. Subsequently, the TMCIMN topology, working principle, and matching range were introduced. This network implements controlled switching of matrix capacitors across operational frequency bands, enabling wideband efficient power transmission. To solve the amplitude-frequency response distortion in transducer sound waves, a pre-distortion method based on finite impulse response (FIR) filter frequency compensation was proposed. This method estimates the system impulse response using an equalization algorithm. The equalizer pre-processes the reference signal to compensate for missing frequency components and suppress excessive ones, thereby improving the sound output linearity. Simulation and experimental results demonstrated that, compared to traditional static impedance matching networks, the proposed method combines dynamic bandwidth matching switching with acoustic distortion compensation capability, achieving over 8-fold bandwidth expansion.

Deconvolution algorithms exhibit significant limitations in complex environments. To enhance the accuracy of direction-of-arrival (DOA) estimation, this article proposes an estimation algorithm that integrates deep learning with deconvolution processing. The algorithm preprocesses the received data using a deconvolution approach and employs quadratic programming and compressed sensing algorithms to obtain the spatial power distribution vector and the input data for the neural network. To improve the accuracy of covariance matrix estimation, a stacked convolutional autoencoder model is constructed, which suppresses noise in the covariance matrix through model pretraining. In addition, we design a data selection operation to enhance the network’s focus on target locations. Furthermore, an attention mechanism tailored for beamforming results is introduced, utilizing multiple attention taps and bias taps to process data, enhancing the accuracy of error estimation. By integrating the attention mechanism with a multimodal network, the proposed algorithm achieves high-precision DOA estimation. Simulation and experimental results validate that the algorithm demonstrates superior estimation accuracy and resolution.

Deconvolution algorithms often rely on conventional beamforming methods to obtain beamforming vectors, which limit their resolution. To enhance parameter estimation resolution, this article introduces the Parameter Estimation by Alternating Reconstruction and Sensation (PEARS) algorithm. In the proposed algorithm, direction estimation leverages a linearly constrained quadratic programming method and weighted L1-norm to solve the objective function, achieving higher resolution in the direction spectrum under fixed weighted vector conditions. The algorithm utilizes the gradient descent method to update the weighted vector, and the relationship among the dictionary matrix, direction spectrum, and weighted vector is computed using the chain rule. This process improves direction estimation results, particularly in scenarios with low signal-to-noise ratios. By alternating between target parameter estimation and weight vector calculation, the PEARS algorithm achieves highly accurate target azimuth estimation. Simulation results validate the algorithm's ability to accurately estimate target azimuth angles. In addition, lake and sea experimental results demonstrate the algorithm's effectiveness in correctly estimating direction in complex environments.

This article proposes a new active sonar detector based on a beamformed deep neural network (BDNN) in three steps. The process involves a preprocessing step, a deep neural network (DNN) application step, and a subsequent postprocessing step. In the preprocessing step, partial spectra are extracted from multiple directions through frequency-domain beamforming. These partial spectra from different directions serve as DNN input, yielding estimated target probabilities as output in the DNN application step. In the postprocessing step, a multiframe probability multiplication technique is proposed, and the number of frames is determined adaptively. The proposed BDNN generates a gridded azimuth-distance graph, where each grid cell represents the probability of a target's presence at a specific azimuth and distance. To guarantee real-time application, we also propose a graphics processing unit based parallel acceleration method, which increases the computation speed of the beamforming process by nearly two orders of magnitude compared to the CPU. The proposed BDNN is verified through sea and lake trials. The results demonstrate that the proposed BDNN achieves better detection performance compared to the conventional matched filter method and exhibits remarkable generalization capabilities.

In the field of signal processing, traditional time of array(TOA) algorithms, as well as those based on compressed sensing, excel in determining time estimates for data. However, these algorithms typically operate within linear models and may not fully extract data information, thereby affecting estimation accuracy and resolution. Capitalizing on their robust nonlinear processing capabilities, deep learning algorithms demonstrate proficiency in extracting features from data. Consequently, we propose a time delay estimation algorithm that initially preprocesses data through compressed sensing. Subsequently, the data undergo processing through a constructed multimodal model, where various features are fused, leading to the final time delay estimation result. By comprehensively processing various features, the algorithm extracts information more thoroughly, resulting in heightened resolution. Experimental assessments confirm the commendable performance of the algorithm.