An Optimization Method for Acoustic Sensor Array Based on Neural Network Enhancement of NSGA-II
By introducing a neural network module into NSGA-II, a closed-loop optimization method is formed, which solves the problems of low computational efficiency and insufficient adaptability of traditional acoustic sensor array optimization in dynamic marine environments. This enables efficient and fast array position optimization and reconfiguration, resulting in a high-quality solution set.
Patent Information
- Application Number
- CN202611229182.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-13
- Publication Date
- 2026-11-17
AI Technical Summary
Traditional acoustic sensor array optimization methods are computationally expensive and lack environmental adaptability when facing dynamic marine environments. They are difficult to find the globally optimal array position, and the decision-making process is cumbersome and cannot respond quickly to environmental changes.
Neural networks are introduced as intelligent enhancement modules for NSGA-II. Through agent evaluation networks, solution generation networks, and policy generation networks, a closed-loop optimization method combining data-driven and rule-based evolution is formed to perform intelligent population initialization, mixed precision evaluation, and directed genetic evolution, and to update the model in real time to adapt to environmental changes.
It significantly improves computational efficiency, enables real-time array optimization and reconfiguration, enhances the algorithm's search capability in complex constrained spaces, obtains a better final solution set, and possesses strong generalization ability and robustness.
Smart Images

Figure CN123116789A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of underwater acoustic engineering and underwater acoustic detection technology, specifically to an optimization method for acoustic sensor arrays based on neural network-enhanced NSGA-II. Background Technology
[0002] Acoustic sensors are a crucial component of underwater monitoring systems, and their detection performance is highly dependent on the selection of deployment sites. Traditional site selection methods rely primarily on historical experience, nautical chart data, and limited on-site hydrological surveys, which presents the following problems: First, once a site is selected, it remains fixed and cannot adapt to the spatiotemporal variations of the marine environment. This leads to significant fluctuations in acoustic sensor performance across different seasons, regions, and even different times of day, resulting in performance blind spots at certain times. Second, reliance on expert experience lacks comprehensive and quantitative acoustic performance evaluation across a wide range of candidate sea areas, making it difficult to find the globally optimal site and even more difficult to optimize arrays composed of multiple acoustic sensors. Third, the decision-making process is cumbersome, requiring manual analysis of large amounts of fragmented environmental data, and cannot quickly respond to new mission requirements or sudden environmental changes.
[0003] Currently, research on underwater acoustic sensor array planning is often limited to theoretical calculations under specific conditions. Based on pre-calculated data under different combinations of parameters such as sea area, season, and array type, the monitoring performance of sensor networks under certain specific conditions is statically analyzed, which lacks adaptability to the complex marine environment. The traditional NSGA-II has excellent Pareto front search capabilities and good convergence, but each iteration requires accurate performance evaluation of a large number of individuals in the population, such as calling complex underwater acoustic propagation models or simulations, which is computationally intensive. Moreover, the standard algorithm has limited detection and development capabilities under complex constraints and is prone to getting stuck in local Pareto fronts. Neural networks have powerful nonlinear fitting and pattern recognition capabilities, but it is difficult to directly generate high-quality solution sets that satisfy constraints in multi-objective optimization.
[0004] In summary, the traditional NSGA-II suffers from high computational cost, insufficient environmental adaptability, and limited search efficiency when optimizing acoustic sensor arrays for dynamic marine environments and complex engineering constraints. To address this, this invention introduces a neural network as an intelligent enhancement module for NSGA-II, which synergistically enhances the traditional evolutionary process from three stages: initialization, evaluation, and genetic operations, forming a closed-loop optimization method that combines data-driven and rule-based evolution. Summary of the Invention
[0005] To achieve the above objectives, the present invention provides the following technical solution: An optimization method for acoustic sensor arrays based on neural network-enhanced NSGA-II includes the following steps: Acquire multi-source marine environmental data and sensor array parameters of the target sea area within a specific time period in the future. Perform unified coordinate, unified grid, unified time step and normalization processing on the data to establish an environmental field corresponding to the optimization task. Based on the environmental field, train a surrogate evaluation network for rapid evaluation of array position performance, a solution generation network for generating high-quality initial array positions, and a strategy generation network for guiding crossover and mutation operations. A solution generator network is used to perform intelligent population initialization. In each generation iteration, a hybrid precision parallel evaluation strategy combining a proxy evaluation network and a high-fidelity simulator is used to obtain the individual objective function value. The strategy generator network outputs the crossover mask probability, mutation adjustment vector and adaptive variance coefficient to achieve directed genetic evolution. After multiple generations of non-dominated sorting, crowding calculation and elite selection, the convergence condition or the maximum number of iterations is met. Real-time collection of position-precise performance data generated by the high-fidelity simulator during the iteration process. When the amount of new data reaches a preset threshold, or when the prediction error or prediction uncertainty of the agent evaluation network exceeds a preset range, the new data is used to perform lightweight fine-tuning on the agent evaluation network and the policy generation network. The optimal non-dominated solution set after the output iteration terminates is selected. Based on real-time task requirements such as coverage, average detection probability, array deployment availability, communication assurance, and engineering constraints, a suitable array deployment scheme is selected from the optimal non-dominated solution set and the engineering deployment is completed.
[0006] Preferably, the multi-source marine environmental data includes marine hydrological forecast data, environmental noise data, seabed geographic information data, and sensor array parameters. The marine hydrological forecast data includes three-dimensional gridded data covering seawater temperature, salinity, sound speed, and depth, collected through HYCOM and ROMS numerical forecast models. This data is unified into a standardized three-dimensional grid and a time step is set using Kriging interpolation or bilinear interpolation. The environmental noise data sources include AIS data, historical marine environmental noise databases, and wind-generated noise models. The calculation of sea surface wind speed and wind-generated noise level is combined, and the total noise spectrum level is obtained using a superposition formula. The seabed geographic information data includes the integration of topographic data from the Global Ocean Depth Sounding Data Center and seabed sediment classification maps, and a lookup table is constructed for sediment type and seabed reflection loss, sedimentary sound speed, and attenuation coefficient. The three-dimensional spatial coordinates and accessory attributes of the sensors are constructed into an array matrix, and the coordinates are normalized to intervals.
[0007] Preferably, the total noise spectrum level obtained by using the energy superposition formula includes: The shipping noise level is obtained by analyzing the ship density, type, and speed information provided by AIS data. The calculation method is as follows: ; In the formula, Here, f represents the shipping noise level, R is the frequency, R is the reference distance from the ship or channel to the receiving point, and Den is the ship density per unit area or unit channel length. , Both K and K are empirical coefficients determined based on statistical data of ship type, speed, sea state, and sea area. Based on sea surface wind speed information, the wind-induced noise level can be obtained, and the calculation method is as follows: ; In the formula, The wind-induced noise level is defined as f, where f is the frequency and W is the sea surface wind speed. The wind-induced noise spectrum level is used to characterize the contribution of sea surface scattering and breaking wave noise caused by wind and waves to the background noise at the receiver. Based on each noise component, the total noise spectrum level is obtained by superimposing the energy corresponding to the sound pressure spectral level. ; In the formula, The total noise spectrum level, excluding shipping noise. Wind creates noise In addition, rainfall noise, marine biological noise, platform self-noise, or other measurable background noise components can be added according to the target sea area conditions, and the total noise spectrum level can be obtained by superimposing the same energy.
[0008] Preferably, the proxy evaluation network that enables rapid evaluation of array position performance based on separate training of environmental fields includes: Training of the surrogate evaluation network is performed based on a convolutional neural network architecture. The surrogate evaluation network is denoted as... Where X is the position matrix and E is the local environmental field characteristic corresponding to the position. To evaluate the network parameters so that they approximate the array position performance mapping given by a high-fidelity underwater acoustic simulator, i.e. ; Define loss function The calculation method is as follows: ; In the formula, To evaluate the training loss of the network, To evaluate the network parameters for the agent, including weights and biases, The number of samples in the training set; Let be the array position matrix of the i-th sample; The environmental field features corresponding to the i-th sample; The true performance vector of the i-th sample is obtained through high-fidelity simulation. The true performance vector includes at least one of the following: effective coverage area, regional coverage, average detection probability, redundancy coverage, or engineering deployment index. The vector representing the network's prediction performance for the i-th sample is used to evaluate the proxy. It is the square of the L2 norm, i.e., the square of the Euclidean distance, used to measure the prediction error; The regularization coefficient is used. This is a regularization term.
[0009] Preferably, the solution generation network that is trained based on the environmental field to generate high-quality initial matrix positions includes: The signal-to-noise ratio is calculated using the sonar equation, and the effective coverage area / regional coverage, average detection probability, and detection range contour lines are defined as performance evaluation indicators for the array. A dataset is created by collecting high-quality array configuration schemes and their performance evaluation results generated from high-fidelity simulations or historical projects. Construct a condition vector for each sample , the position coordinate matrix After normalization, with the condition vector Pairing them up as training samples; The conditional generative model is trained using a conditional variational autoencoder (CVAE), which consists of an encoder, a decoder, and a latent space. The conditional generative model is denoted as... Based on the given performance preference or constraint c, generate potential high-quality matrix positions that meet the conditions. CVAE learns a latent variable The conditional distribution generates diverse and controllable outputs. Encoder Real formation and conditions Encoding as latent variables The posterior distribution, decoder From latent variables and conditions Reconstruction of positions ; When selecting a training target, the loss function is first defined using the evidence lower bound loss. The calculation method is as follows: ; In the formula, The total loss of the conditional variational autoencoder; These are decoder parameters; For encoder parameters; This is the actual sensor array position matrix; The vector represents performance preferences or engineering constraints; z represents latent variables. This represents the posterior distribution of the encoder output; To conditionally reconstruct the distribution of the decoder output; For the prior distribution of latent variables, To control the hyperparameters of regularization strength, Here, KL divergence is used to measure the difference between two distributions; overall, the first term is the reconstruction loss, used to ensure that the generated matrix positions are within a given range. and Approximately high-quality formations under certain conditions The second term is the KL divergence term, which is used to constrain the latent variable distribution to approximate the prior distribution; Randomly initialize the parameters and use gradient descent to minimize the total loss on dataset D until convergence.
[0010] Preferably, the strategy generation network that generates guided intelligent genetic operations based on environmental field training includes: constructing a demonstration dataset; designing an interactive simulation environment with position schemes, population statistics, and environmental context as states, and crossover masks and mutation perturbations as actions; designing an encoder, a state fusion unit, and a two-branch policy head; performing supervised pre-training using the demonstration dataset; connecting the pre-trained network to the interactive simulation environment, maximizing the cumulative reward using a pruning objective function optimized by the proximal policy, and adding imitation loss as a regularization term to achieve phased hybrid training.
[0011] Preferably, a hybrid precision parallel evaluation strategy is adopted after the intelligent population is initialized by a solution generation network. Specifically, during each generation of evaluation, a dynamic scheduling mechanism is used, in which most individuals are evaluated quickly by a proxy evaluation network; at the same time, a small number of individuals and randomly sampled individuals are selected and submitted to a high-fidelity simulator for evaluation.
[0012] Preferred, Generation G population Each individual in The probability of it being assigned to use high-fidelity simulation The following rules determine the dynamic nature of the situation: ; In the formula, Let i be the probability that the i-th individual is assigned to be evaluated using a high-fidelity simulator. Let be the indicator function, representing the non-dominated ranking rank of the i-th individual based on the performance of the previous generation. Better than or equal to the former If the name is specified, use 1; otherwise, use 0. Where N represents the proportion of elites and N is the population size. Let be the crowding degree of the i-th individual in the decision space or the minimum distance to other individuals, and normalized to . Intervals are used to encourage exploration of sparse regions; To assess the average prediction uncertainty of the network near the current Gth generation population; These are weighting coefficients used to control the relative importance of three factors: elite priority, diversity exploration, and uncertainty correction; according to The evaluation path for each individual is determined by roulette wheel or direct threshold method. After the evaluation is completed, all individuals obtain their objective function value vector.
[0013] Preferably, when the population is selected through multiple generations of iterative screening until convergence or the termination condition is met, an environment selection step is required. The environment selection step operates on the population that has undergone mixed evaluation. and its offspring The main process is as follows: merge: , This refers to the candidate population after merging; Non-dominated sorting: merge candidate populations All solutions are hierarchically arranged according to Pareto dominance relations; Crowding Calculation: Within the same undominated layer, calculate the crowding distance of each solution in the target space to measure the distribution density of the solutions; Elite Selection: Select the top-performing solutions according to the undominated layers from best to worst, while also considering the crowding density within each layer from highest to lowest. Individuals form a new generation population, where N is the preset population size.
[0014] Preferably, the directed genetic evolution achieved by using a strategy generation network includes: inputting parent gene fragments, local environmental context, global optimization state, and evolutionary stage information into the strategy generation network; the crossover strategy head in the dual-branch strategy head outputs a crossover mask probability vector, and for each sensor dimension, performs standard binary crossover or directly inherits the parent value according to probability; the mutation strategy head outputs a mutation adjustment vector and adaptive variance coefficient, and performs biased mutation operations.
[0015] This invention provides an acoustic sensor array optimization method based on neural network enhancement of NSGA-II, which has the following beneficial effects: This invention significantly improves the overall computational efficiency of a single optimization task by replacing a large number of high-fidelity simulation calls with a proxy evaluation network, achieving an order-of-magnitude reduction in computational cost and enabling real-time, online array optimization and reconfiguration. Intelligent initialization provides a high starting point, accelerates early convergence, and allows the algorithm to approach the global Pareto front more quickly. Neural network-guided genetic operators increase the probability of generating feasible and high-quality solutions, enhance the algorithm's search capability in complex constraint spaces, and thus obtain a final solution set with better distribution and convergence.
[0016] The unique online learning mechanism of this invention transforms the system from a static optimization tool into an intelligent agent capable of learning from each optimization task and continuously improving itself. When applied to different sea areas or new sensors, it can adapt quickly, demonstrating strong generalization ability and robustness.
[0017] This invention forms a complete closed-loop solution from data preparation and model training to optimized deployment. It is compatible with existing high-fidelity simulation tools, does not require replacement of the underlying physical model, is easy to integrate into existing underwater system design processes, and has high potential for engineering implementation.
[0018] This invention allows the knowledge learned by the generator network and policy network during the optimization process to be fed back to the designers, enhancing their understanding of the complex optimization problem itself and providing deeper insights for solution decisions. Attached Figure Description
[0019] Figure 1 This is a structural diagram of the initial population of the solution generation network in this invention; Figure 2 This is a structural diagram of the neural network-enhanced NSGA-II in this invention; Figure 3 This is a structural diagram of the hybrid precision parallel evaluation strategy in this invention. Detailed Implementation
[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0021] Example 1: like Figure 1-3 As shown, this invention relates to the fields of underwater acoustic engineering and underwater acoustic detection technology, and in particular to an acoustic sensor array optimization method based on neural network-enhanced NSGA-II, which is applicable to the array positioning planning of shore-based sonar, buoy and other detection systems.
[0022] The present invention aims to overcome the shortcomings of the prior art and provide an underwater sensor array optimization method that deeply integrates neural networks into the NSGA-II evolution process. The core idea of the present invention is to use neural networks as intelligent enhancement modules of NSGA-II, which are applied to the three key stages of the optimization process: evaluation, initialization and evolution, forming a dynamic closed-loop framework driven by learning and evolution.
[0023] To achieve the above objectives, this invention employs several key steps: constructing and training a basic neural network, initiating the NSGA-II optimized main loop for neural network enhancement, dynamic data collection and online model updates, and optimizing result output and application.
[0024] (1) Constructing and training the basic neural network (1.1) Establishment of the environmental field; Acquire multi-source marine environmental data for the target sea area over a specific future time period, including marine hydrological forecast data, environmental noise data, and seabed geographic information data, and perform standardization and spatiotemporal matching processing to establish a unified environmental field.
[0025] (1.1.1) Marine hydrological data: Marine environmental data sources include numerical forecasting models such as HYCOM and ROMS, covering elements such as seawater temperature, salinity, sound speed, and depth. The data is three-dimensional gridded, and three-dimensional spatiotemporal interpolation, such as Kriging interpolation or bilinear interpolation, is performed on the target sea area to unify data from different sources and resolutions onto a standardized three-dimensional grid. The time step is [missing information]. ,in, This represents the time interval between two adjacent environmental field update times, which can be determined based on the forecast data resolution and the task replanning cycle.
[0026] (1.1.2) Environmental noise data: The sources of environmental noise data include AIS data, historical marine environmental noise databases and wind-induced noise models.
[0027] Based on the ship density, type, and speed information provided by AIS data, the shipping noise level can be obtained, and the calculation method is as follows: ; In the formula, Here, f represents the shipping noise spectrum level, R is the reference distance from the ship or channel to the receiving point, and Den is the ship density per unit area or unit channel length. , K is an empirical coefficient determined based on statistical data of ship type, speed, sea state, and sea area.
[0028] Based on sea surface wind speed information, the wind-induced noise level can be obtained, and the calculation method is as follows: In the formula, The noise spectrum is represented by f, where f is the frequency and W is the sea surface wind speed. This noise component reflects the contribution of wind and wave conditions to the background noise spectrum.
[0029] Based on each noise component, the total noise spectrum level can be obtained by superimposing the energy corresponding to the sound pressure spectral level. When there are other background noise components such as rainfall noise, marine biological noise, and platform self-noise in the target sea area, they are added as additional terms to participate in the energy superposition.
[0030] ; (1.1.3) Seafloor topography and sediment data: including GEBCO seafloor topography data and seafloor sediment classification map, and establish lookup tables for sediment types (such as silt, clay, etc.) and seafloor reflection loss, sedimentary sound velocity and attenuation coefficient, which serve as key inputs for the underwater acoustic propagation model.
[0031] (1.1.4) Sensor array parameters: For an array consisting of M sensors, its array positions can be expressed as follows: ; In the formula, This is the array position matrix of the sensor array. Represents the three-dimensional spatial coordinates (horizontal coordinate, vertical coordinate, and depth coordinate) of the m-th sensor. The vector represents the additional attributes of the m-th sensor, including sensor type, operating frequency band, array orientation, array form or deployment constraints, etc.; M is the number of sensors; to facilitate neural network processing, the coordinates and continuous attributes are normalized to the [0,1] interval, and the discrete attributes are encoded as one-hot vectors or embedded vectors.
[0032] (1.2) Agency evaluation network; One of the key steps in optimizing the acoustic sensor array is to evaluate the acoustic propagation characteristics at a specified location. This involves dividing the target sea area into a regular latitude and longitude grid, using the center of each grid point as the virtual array position. Using ray acoustic models (such as Bellhop) and normal mode models (such as Kraken), and combining information such as sound velocity profiles, seabed parameters (acoustic properties corresponding to depth and sediment type), and sea surface parameters within a region with the analysis distance as the radius for each virtual array position, the system outputs information from the sound source to the receiving point. While propagation loss can be mitigated, sound field calculations based on physical models are typically time-consuming due to the wide geographical area, complex marine environment, and numerous computational parameters.
[0033] Therefore, a deep regression network is established to quickly predict the performance indicators of a given array position, replacing most of the time-consuming simulation calculations. In this invention, the simulation time for acoustic field calculations of regional grids in complex marine environments over vast sea areas is affected by many factors such as sea area, grid accuracy, number of azimuths, number of frequency points, and algorithm model. To make the simulation calculation results more accurate, the simulation calculation time will increase exponentially. Therefore, using a proxy evaluation network to replace acoustic field simulation calculations to generate pre-training datasets will significantly improve computational efficiency.
[0034] A convolutional neural network architecture is chosen for training the surrogate evaluation network, which is denoted as surrogate evaluation network. Where X is the position matrix and E is the local environmental field characteristic corresponding to the position. To evaluate the network parameters so that they approximate the array position performance mapping given by the high-fidelity underwater acoustic simulator, i.e. ; Define loss function The calculation method is as follows: ; In the formula, To evaluate the training loss of the network, Evaluate the network parameters (weights and biases) for the agent. The number of samples in the training set; Let be the array position matrix of the i-th sample; The environmental field features corresponding to the i-th sample; This is the true performance vector obtained from high-fidelity simulation of the i-th sample; The vector representing the network's prediction performance for the i-th sample is used to evaluate the proxy. It is the square of the L2 norm, i.e., the square of the Euclidean distance, used to measure the prediction error; The regularization coefficient is . The first term is the regularization term (e.g., L2 regularization); overall, the first term is the mean squared error (MSE), which ensures the accuracy of the prediction, and the second term is the regularization term, which is used to prevent overfitting.
[0035] (1.3) Solve the generator network; When solving the multi-objective optimization problem of acoustic sensor deployment, it is first necessary to define the evaluation index. The optimization objectives of multi-array cooperative deployment usually include maximizing the total network coverage area and minimizing redundant coverage. Before giving the definition of the evaluation index, we will first introduce the related concept of signal-to-noise ratio: Assume the spectral level of the target radiated noise is At the target radiating point (xi,yi) and the receiving point The signal-to-noise ratio (SNR) at frequency f is: ; In the formula, SL(f) is the radiated noise spectral level of the target at frequency f; The target radiating point at frequency f to the receiving point The loss of transmission, denoted as the background noise spectral level at frequency f; SNR(f,xi,yi,xr,yr) represents the signal-to-noise ratio under the corresponding frequency and spatial location conditions.
[0036] Broadband energy detection model is used in the specified frequency band Perform integration processing: ; In the formula, For the specified frequency band The broadband signal-to-noise ratio, , respectively, represent the lower and upper frequency limits of the integration band; SL(f), TL(f,xi,yi,xr,yr) and NL(f) represent the target radiated noise spectral level, propagation loss and background noise spectral level at frequency f, respectively.
[0037] Detection probability and detection threshold calculation: Define the detection threshold as DT, when At that time, it is assumed that the target can be detected at the receiving point; for each virtual array position, a polar coordinate grid search is performed within its 360-degree azimuth and maximum detection range, and points that satisfy SNR≥DT are marked as detectable points.
[0038] (1.3.1) Effective coverage area (Absolute quantity) / Regional coverage (Relative quantity); ; In the formula, For formation positions The effective coverage area at time t; Maximum detection range; The polar coordinate grid for the search; This is an indicator function (1 if the condition is met, 0 otherwise); This represents the signal-to-noise ratio at the polar coordinate sampling point, where DT is the detection threshold; This indicates that the detection conditions are met; dA represents a small area element.
[0039] ; In the formula, For regional coverage, Let Q be the total area of the monitoring area, and Q be the total number of grid cells in the area. For the first q The area weight corresponding to each grid point; For the first q The center position of each grid point; SNR(X,Pq) is the signal-to-noise ratio of the array scheme X to the grid point Pq; I(·) is the indicator function.
[0040] It intuitively reflects the absolute monitoring and control range of a single array position or array position scheme. This reflects the proportion of the mission area that is effectively detected and covered, and is used to characterize the integrated surveillance and control capability.
[0041] (1.3.2) Average detection probability ; ; In the formula, This represents the average detection probability. The signal-to-noise ratio is a function that maps to the detection probability. It can be established based on the sonar equations and statistical detection theory, such as using the Sigmoid function or a detection model based on the false alarm probability; Q is the total number of grid cells in the region, which reflects the robustness of the array detection.
[0042] (1.3.3) Contour lines of the detection range; The contour lines of the detection range, i.e. the boundary lines of SNR=DT, represent the instantaneous detection range map of this array position.
[0043] (1.3.4) Solve the generator network; The solution generator network is primarily responsible for intelligently generating high-quality initial candidate array site schemes. Essentially, it is a conditional generative model. This network learns and internalizes implicit knowledge extracted from massive historical or simulation data regarding the spatial structural characteristics that a high-quality sensor array site should possess. This elevates the starting point of the evolutionary algorithm from a completely random population to an experience-guided pool of high-quality solutions, shortening early convergence time and saving ineffective iterations. The technical roadmap of the solution generator network is as follows: Figure 1 As shown.
[0044] Step 1: Collect high-quality array positioning schemes and their performance evaluation results generated from high-fidelity simulations or historical projects to form a dataset. Construct a condition vector for each sample , the position coordinate matrix After normalization, with the condition vector Pair them up as training samples.
[0045] Step 2: Select a Conditional Variational Autoencoder (CVAE) to train the conditional generative model. A CVAE consists of an encoder, a decoder, and a solution space. The conditional generative model is denoted as... This enables it to generate potential high-quality matrix positions that meet given performance preferences or constraints c. CVAE learns a latent variable The conditional distribution can produce diverse and controllable outputs; the encoder Real formation and conditions Encoding as latent variables The distribution (assuming) ); decoder From latent variables and conditions Reconstruct (or generate) array positions .
[0046] Step 3: Select the training objective, such as maximizing the probability of data occurrence under given conditions. First, define the loss function using the lower bound of evidence loss. The calculation method is as follows: ; In the formula, The total loss of the conditional variational autoencoder. For decoder parameters, For encoder parameters, This is the actual sensor array position matrix. Let z be a condition vector representing performance preferences or engineering constraints, and z be a latent variable. To control the hyperparameters of regularization strength, for KL Divergence is used to measure the difference between two distributions.
[0047] Overall, the first item is the reconstruction loss, ensuring the forced generation of array positions. In a given and The next position is close to the true quality of the formation. The second term is the KL divergence, which constrains the distribution of latent variables in the encoder output. Approximate prior distribution To regulate potential space and enhance generative diversity.
[0048] Randomly initialize the parameters and use gradient descent to minimize the total loss on dataset D until convergence.
[0049] (1.4) Policy generation network; The policy network is primarily responsible for learning and predicting how to modify the parent genes to efficiently generate high-quality offspring. Its core objective is to learn a policy that can directly output the optimal genetic operation parameters based on the current optimization state and the parent genes, thereby maximizing the probability that offspring individuals will become elites (enter the Pareto frontier) in the future population.
[0050] (1.4.1) Data and simulation foundation construction; Step 1: Construct a demonstration dataset, run the basic NSGA-II algorithm, optimize and record in different scenarios such as flat seabed and complex sea area, and use filters to retain only the records of offspring genetic operations (crossover / mutation) that produce at least one offspring entering the next generation Pareto front. Each record should include state information such as parent genes, population statistics and environmental context, as well as expert action information. For crossover, it is the exact crossover point mask matrix that produces elite offspring. For mutation, it is the optimal perturbation vector that makes the parent become an elite offspring.
[0051] Step 2: Design an interactive simulation environment to encapsulate the optimization problem of matrix bit encoding to performance evaluation into a standard reinforcement learning environment. Design a reward function and define the immediate reward for each genetic operation. ; ; In the formula, The immediate reward obtained from the t-th genetic operation; This represents the increase in the population oversize index after the introduction of offspring. It is a reward and punishment factor for offspring evolution. If the offspring enters a new frontier or is preserved to the next generation, it is a positive reward; if it is eliminated, it is a negative reward or a minor punishment. To encourage exploration of sparse regions, the goal is to minimize the distance or diversity contribution of offspring to the existing population in the decision-making space. This is the weighting factor.
[0052] (1.4.2) Model design; Step 1: Encoder. The sensor array is modeled as a heterogeneous graph, where nodes represent sensors and include latitude and longitude coordinates and type features, while edges represent physical connections or signal interaction relationships and include distance and coupling strength features. A 2-3 layer heterogeneous graph attention network is used to aggregate differentiated information from different types of nodes and edges, outputting the context-aware code for each node. .
[0053] Step 2: State fusion, encoding nodes Perform max pooling and average pooling to obtain graph-level features g, and then combine g with the global evolutionary state. (Frontier indicators, population diversity) and environmental encoding env are concatenated and fused through a multilayer perceptron to form a comprehensive state representation vector z.
[0054] Step 3: The dual-branch strategy head and crossover strategy head take the graph features of the two parents as input, calculate the original association matrix through bilinear attention, and normalize the rows and columns alternately to make it a doubly random matrix. This can be directly interpreted as an executable, soft crossover probability mask, satisfying the biological constraint that each gene inherits approximately only one parent. The mutation strategy head directly inputs the individual graph features and finally outputs a parameter of conditional Gaussian analysis, the mean. Characterizing the direction of variation, diagonal covariance Characterize the magnitude of variation.
[0055] (1.4.3) Phased mixed training; Step 1: Supervised imitation learning. Pre-train the network using a demonstration dataset to allow it to quickly learn and replicate successful experiences. Define the loss function. ; ; In the formula, The loss is the imitation learning loss; D is the demonstration dataset; s is the state sample. For expert action samples; Let be the cross-loss entropy of the cross-mask, where This is the crossover probability matrix predicted by the network. For cross-disciplinary actions by experts; The mean square error of the direction of variation, where The mean of the predicted mutation directions by the network. For expert variation direction; The difference between the network-predicted Gaussian action distribution and the expert action distribution is measured, where... To predict covariance for the network, For expert action covariance; 、 、 This is the importance coefficient used to balance the three losses.
[0056] Step 2: Online Proximity Policy Optimization. The pre-trained network is connected to an interactive simulation environment. By interacting with the environment, the cumulative reward is maximized. A proximity policy optimization algorithm is used, with its pruning objective function... for: ; In the formula, The pruning objective function optimized for the near-end strategy; The ratio of the probability of the new and old strategies on the same state-action sample. For the estimation of the advantage function, Limit the probability ratio r to Within the interval, ε is the pruning factor to ensure that the policy update step size is not too large and to improve training stability; during training, it is advisable to consider... Adding imitation loss As a regularization term, it prevents the strategy from forgetting pre-trained knowledge.
[0057] (2) Start the NSGA-II optimization main loop with neural network enhancement; The NSGA-II standard process introduces three enhancement modules: intelligent population initialization, a mixed-precision parallel evaluation strategy, and neural network-guided genetic evolution. The technical approach is as follows: Figure 2 As shown: (2.1) Intelligent population initialization; Using intelligent population initialization instead of random generation, and utilizing the solution generation network trained in (1.3), an initial population that meets basic physical constraints and has excellent performance is generated in batches according to the initial preferences set for the current task, providing a high-starting-point "seed bank" for evolution.
[0058] (2.2) Hybrid precision parallel evaluation strategy; The hybrid precision parallel evaluation strategy aims to construct an intelligent scheduling system that dynamically determines the evaluation path for each candidate solution, achieving an optimal balance between accuracy and speed. In each generation of evaluation, a dynamic scheduling mechanism is employed, with most individuals being rapidly evaluated by a proxy evaluation network. Simultaneously, a small subset of key individuals (such as elite solutions and diverse solutions) and randomly sampled individuals are submitted to a high-fidelity simulator for precise evaluation. This strategy ensures overall search speed while anchoring the accuracy of solutions through high-fidelity evaluation and monitoring the errors of the proxy model. The technical roadmap of the hybrid precision parallel evaluation strategy is as follows: Figure 3 As shown: For the Gth generation population Each individual in The probability of it being assigned to use high-fidelity simulation The following rules determine the dynamic nature of the situation: ; In the formula, For the first i The probability of each individual being assigned to be evaluated using a high-fidelity simulator. For indicator functions, if the first... i Non-dominant ranking of individuals based on the performance of the previous generation Better than or equal to the former If the name is specified, the value is 1; otherwise, it is 0. Where N represents the proportion of elites and N is the population size. For the first i The crowding degree of an individual in the decision space, or the minimum distance from other individuals, is normalized to The interval encourages exploration of sparse regions; The average prediction uncertainty of the surrogate model near the current population (e.g., the prediction variance of the ensemble model) is updated as optimization proceeds; These are weighting coefficients, used to control the relative importance of three factors: elite priority, diversity exploration, and uncertainty correction.
[0059] Based on this probability The evaluation path for each individual is determined by roulette wheel or direct thresholding. For individuals entering the high-fidelity path, an underwater acoustic propagation model or simulator is invoked to obtain an accurate performance vector. For individuals entering the surrogate path, a surrogate evaluation network is invoked to obtain a predicted performance vector. After evaluation, each individual obtains an objective function value vector for non-dominated ranking.
[0060] The environmental selection process involves populations that have undergone mixed evaluation. and its offspring The main process is as follows: Step 1: Merge: ; in, This represents the parent population of generation G. This represents the offspring population produced by crossover and mutation of the parent population in generation G. This represents the candidate population after merging.
[0061] Step 2: Non-dominated sorting: Merge candidate populations All solutions are hierarchically ordered according to Pareto dominance (front sort); Step 3: Crowding Calculation: In the same non-dominated layer, calculate the crowding distance of each solution in the target space to measure the distribution density of the solutions; Step 4: Elite Selection: Based on the non-dominated layer from best to worst, and also considering the crowding level within the layer from highest to lowest, select the top performers. Individuals form a new generation of population , where N is the preset population size.
[0062] (2.3) Neural networks guide genetic evolution; This section mainly explains how to replace or enhance the blind crossover and mutation operators in traditional genetic algorithms using neural networks to achieve directed and intelligent evolution. The core of this approach is to train a policy network to learn to predict, that is, what kind of "modification" to be made to a given parent gene is most likely to produce high-quality offspring. In crossover and mutation operations, the output of this network is used to guide the execution of genetic operators, making the generation of offspring more likely to conform to the physical laws of the underwater environment, thereby improving evolutionary efficiency.
[0063] The inputs to the policy network include: parent gene fragments, i.e., the sensor coordinate matrix of the parent individual to be operated on; local environmental context, i.e., the feature encoding of the operation focus area such as terrain and sound velocity profile; global optimization state, i.e., the Pareto front approximation performance of the current population; and evolutionary stage information, i.e., the ratio of the current generation to the total number of generations, used to indicate the bias of exploration and development. The outputs include: the proposed crossover probability distribution and the mutation adjustment vector.
[0064] Traditional analog binary crossover (SBX) generates offspring according to a fixed probability distribution, without explicitly utilizing the environmental field and population state. Enhanced SBX, incorporating neural networks, first inputs gene fragments from the parent pair (e.g., coordinates of a sensor row) and local environmental context (surrounding terrain, sound velocity profile, and constraint encoding) into the policy network. The policy generation network outputs a crossover mask probability vector. For the first... Each sensor dimension, in The probability of performing a standard SBX operation is given. The probability is directly inherited from a parent value, which makes crossover more likely to occur on dimensions that the "strategy" deems beneficial.
[0065] Traditional multinomial mutation (PM) randomly perturbs genes with fixed probabilities and distributions. Neural network-guided multinomial mutation first inputs the individual X to be mutated into the policy network, and the network outputs a mutation adjustment vector. and an adaptive variance coefficient Finally, the mutation operation is performed; ; In the formula, Let be the original value of the j-th decision variable in the i-th individual; The mutated variable value; The mutation indicator variable is executed according to adaptive probability, and its importance is determined by the prediction of this dimension by the policy generation network. These are adaptive variance coefficients; The biased mutation adjustment for the output of the policy generation network; The perturbation is a random disturbance from a traditional multinomial distribution; this operation allows the mutation to explore in a small, precise direction in the favorable direction predicted by the network, while retaining the ability to explore randomly.
[0066] (3) Dynamic data collection and online model updates; The new data pairs (array position-precision performance) generated by the high-fidelity simulation evaluation in step (2.2) are stored in the historical database in real time. When the amount of new data accumulates to a certain threshold, or when the prediction error of the surrogate model on the sampled high-fidelity samples exceeds the set range, the online learning mechanism is triggered to fine-tune the surrogate evaluation network and the policy generation network with the new data, so that the optimized network can continuously adapt to new sea areas, seasons, tasks or sensor configuration conditions.
[0067] (4) Optimize result output and application; Once the optimization iteration meets the termination condition (such as reaching the maximum number of iterations or the convergence criterion), the final non-dominated solution set (Pareto optimal front) is output. Decision-makers can flexibly select the final array deployment scheme from this front according to the real-time task requirements.
[0068] The essence of this invention is not to simply use neural networks and NSGA-II side by side, but to use neural networks as functional modules to deeply embed and modify the three core processes of NSGA-II: initialization, evaluation and evolution, forming a collaboratively enhanced intelligent optimization network that is interdependent and provides bidirectional feedback.
[0069] This invention proposes a core feature: enhancing the initialization, individual evaluation, and genetic operations of evolutionary algorithms using a trained neural network model. This invention presents an intelligent population initialization method based on generative networks, a dynamic hybrid evaluation strategy based on surrogate networks and high-fidelity simulation, a method for guiding the generation of genetic operators using policy networks, and a closed-loop learning mechanism for online updating of neural network modules using new data generated during the optimization process.
[0070] Example 2 For the technical solution of this invention, namely, an acoustic sensor array optimization method based on neural network enhancement NSGA-II, there are other alternative solutions that can also achieve the purpose of underwater sensor array configuration evaluation and optimization, including but not limited to: Neural network model alternatives include: First, the agent evaluation network can use graph neural networks to better handle the topological relationships between sensor nodes, or use the Transformer architecture to capture the mutual influence between array elements over long distances; second, the solution generation network can use generative adversarial networks or their variants to generate more realistic array position samples; and third, the policy network can be implemented using the Actor network in reinforcement learning, treating genetic operations as a sequential decision-making process.
[0071] Evolutionary Algorithm Replacement: NSGA-II can be used as the basic framework, but its core multi-objective processing mechanism can be replaced by a reference point-based high-dimensional objective optimization algorithm (NSGA-III), a decomposition-based multi-objective evolutionary algorithm (such as MOEA / D), or an index selection multi-objective evolutionary algorithm (such as IBEA). The enhancement modules for initialization, evaluation, and operators in this invention can be migrated to these algorithms in parallel.
[0072] The hybrid evaluation of "surrogate model + high-fidelity model" can adopt a more refined adaptive trust threshold mechanism, which dynamically decides whether to enable high-fidelity verification based on the historical prediction error of the surrogate model near the individual; the online update strategy can adopt incremental learning or continuous learning methods to better alleviate the catastrophic forgetting problem of the model caused by changes in data distribution.
[0073] Problem generalization alternative: The method described in this invention can be directly applied to any field involving multi-objective optimization of spatial resources, such as surface ship sonar array optimization, aerial UAV monitoring network deployment, and land wireless sensor layout, simply by replacing the corresponding performance evaluation simulator and constraints.
[0074] The following is the software flow illustrating this method, describing the implementation steps of the dynamic sensor array positioning planning method based on acoustic performance prediction in the evaluation and optimization of sensor array schemes: Step 1: Build a historical / simulation database.
[0075] Step 2: Offline training of agents to evaluate the neural network and generate the neural network.
[0076] Step 3: Start the enhanced NSGA-II optimization loop; initialize the population using the generated NN; in the loop, some individuals are evaluated by the agent NN, and others are evaluated by simulation; selection, crossover (assisted by the policy network), mutation; merge the populations, and perform non-dominated sorting and crowding calculation.
[0077] Step 4: After the termination condition is met, output the Pareto optimal solution set for the decision-maker to choose from.
[0078] Step 5: Add the high-quality data generated from this optimization to the database and update the neural network model.
[0079] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.
[0080] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.
[0081] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.
Claims
1. An acoustic sensor array optimization method based on neural network enhanced NSGA-II, characterized in that, Includes the following steps: Acquire multi-source marine environmental data and sensor array parameters of the target sea area within a specific time period in the future. Perform unified coordinate, unified grid, unified time step and normalization processing on the data to establish an environmental field corresponding to the optimization task. Based on the environmental field, train a surrogate evaluation network for rapid evaluation of array position performance, a solution generation network for generating high-quality initial array positions, and a strategy generation network for guiding crossover and mutation operations. A solution generator network is used to perform intelligent population initialization. In each generation iteration, a hybrid precision parallel evaluation strategy combining a proxy evaluation network and a high-fidelity simulator is used to obtain the individual objective function value. The strategy generator network outputs the crossover mask probability, mutation adjustment vector and adaptive variance coefficient to achieve directed genetic evolution. After multiple generations of non-dominated sorting, crowding calculation and elite selection, the convergence condition or the maximum number of iterations is met. Real-time collection of position-precise performance data generated by the high-fidelity simulator during the iteration process. When the amount of new data reaches a preset threshold, or when the prediction error or prediction uncertainty of the agent evaluation network exceeds a preset range, the new data is used to perform lightweight fine-tuning on the agent evaluation network and the policy generation network. The optimal non-dominated solution set after the output iteration terminates is selected. Based on real-time task requirements such as coverage, average detection probability, array deployment availability, communication assurance, and engineering constraints, a suitable array deployment scheme is selected from the optimal non-dominated solution set and the engineering deployment is completed.
2. The acoustic sensor array optimization method based on neural network-enhanced NSGA-II according to claim 1, characterized in that: Multi-source marine environmental data includes marine hydrological forecast data, environmental noise data, seabed geographic information data, and sensor array parameters. Marine hydrological forecast data includes three-dimensional gridded data covering seawater temperature, salinity, sound speed, and depth collected through HYCOM and ROMS numerical forecast models. The data is unified into a standardized three-dimensional grid and a time step is set through Kriging interpolation or bilinear interpolation. The sources of environmental noise data include AIS data, historical marine environmental noise databases, and wind-generated noise models. The calculation of sea surface wind speed and wind-generated noise level is combined, and the total noise spectrum level is obtained by using the superposition formula. The seabed geographic information data integrates topographic data from the Global Ocean Depth Sounding Data Center and seabed sediment classification maps, and constructs lookup tables for sediment types, seabed reflection loss, sedimentary sound velocity, and attenuation coefficients; it constructs a matrix of sensor three-dimensional spatial coordinates and accessory attributes, and normalizes the coordinates to intervals.
3. The acoustic sensor array optimization method based on neural network-enhanced NSGA-II according to claim 2, characterized in that: Using the energy superposition formula, the total noise spectrum level is derived as follows: The shipping noise level is obtained by analyzing the ship density, type, and speed information provided by AIS data. The calculation method is as follows: ; In the formula, For shipping noise levels, f Where R is the frequency, and R is the reference distance from the ship or waterway to the receiving point. Den The density of ships per unit area or per unit length of waterway. , , K All of these are empirical coefficients calibrated based on statistical data of ship type, speed, sea state, and sea area; Based on sea surface wind speed information, the wind-induced noise level can be obtained, and the calculation method is as follows: ; In the formula, The noise level is determined by wind. f Where is the frequency and W is the sea surface wind speed; the wind-induced noise spectrum level is used to characterize the contribution of sea surface scattering and breaking wave noise caused by wind and waves to the background noise at the receiver. Based on each noise component, the total noise spectrum level is obtained by superimposing the energy corresponding to the sound pressure spectral level. ; In the formula, The total noise spectrum level, excluding shipping noise. Wind creates noise In addition, rainfall noise, marine biological noise, platform self-noise, or other measurable background noise components can be added according to the target sea area conditions, and the total noise spectrum level can be obtained by superimposing the same energy.
4. The acoustic sensor array optimization method based on neural network-enhanced NSGA-II according to claim 1, characterized in that: The proxy evaluation network that enables rapid performance evaluation of array positions based on separate training in the environmental field includes: Training of the surrogate evaluation network is performed based on a convolutional neural network architecture. The surrogate evaluation network is denoted as... Where X is the position matrix and E is the local environmental field characteristic corresponding to the position. To evaluate the network parameters so that they approximate the array position performance mapping given by a high-fidelity underwater acoustic simulator, i.e. ; Define loss function The calculation method is as follows: ; In the formula, To evaluate the training loss of the network, To evaluate the network parameters for the agent, including weights and biases, The number of samples in the training set; Let be the array position matrix of the i-th sample; The environmental field features corresponding to the i-th sample; For the first i The real performance vector of a sample obtained by high-fidelity simulation includes at least one of the following: effective coverage area, regional coverage, average detection probability, redundancy coverage, or engineering deployment index. For the agent to evaluate the network on the first i The prediction performance vector for each sample; It is the square of the L2 norm, i.e., the square of the Euclidean distance, used to measure the prediction error; The regularization coefficient is used. This is a regularization term.
5. The acoustic sensor array optimization method based on neural network-enhanced NSGA-II according to claim 4, characterized in that: The solution generation network, trained separately based on the environmental field, generates high-quality initial matrix positions and includes: The signal-to-noise ratio is calculated using the sonar equation, and the effective coverage area / regional coverage, average detection probability, and detection range contour lines are defined as performance evaluation indicators for the array. A dataset is created by collecting high-quality array configuration schemes and their performance evaluation results generated from high-fidelity simulations or historical projects. Construct a condition vector for each sample , the position coordinate matrix After normalization, with the condition vector Pairing them up as training samples; The conditional generative model is trained using a conditional variational autoencoder (CVAE), which consists of an encoder, a decoder, and a latent space. The conditional generative model is denoted as... Based on the given performance preference or constraint c, generate potential high-quality matrix positions that meet the conditions. CVAE learns a latent variable The conditional distribution generates diverse and controllable outputs; encoder Real formation and conditions Encoding as latent variables The posterior distribution, decoder From latent variables and conditions Reconstruction of positions ; When selecting a training target, the loss function is first defined using the lower bound of evidence loss. The calculation method is as follows: ; In the formula, The total loss of the conditional variational autoencoder; These are decoder parameters; For encoder parameters; This is the actual sensor array position matrix; Z represents the performance preference or engineering constraint vector; z represents the latent variable. This represents the posterior distribution of the encoder output; To conditionally reconstruct the distribution of the decoder output; For the prior distribution of latent variables, To control the hyperparameters of regularization strength, for KL Divergence measures the difference between two distributions; overall, the first term is the reconstruction loss, used to ensure that the generated matrix positions are within a given range. and Approximately true high-quality formations under certain conditions The second item is KL The divergence term is used to constrain the latent variable distribution to approximate the prior distribution. Randomly initialize the parameters and use gradient descent to minimize the total loss on dataset D until convergence.
6. The acoustic sensor array optimization method based on neural network-enhanced NSGA-II according to claim 5, characterized in that: The strategy generation network for guiding intelligent genetic operations, trained separately based on the environmental field, includes: constructing a demonstration dataset; designing an interactive simulation environment with position schemes, population statistics, and environmental context as states, and crossover masks and mutation perturbations as actions; designing an encoder, state fusion unit, and a two-branch policy head; performing supervised pre-training using the demonstration dataset; connecting the pre-trained network to the interactive simulation environment, maximizing the cumulative reward using a pruning objective function optimized by the proximal policy, and adding imitation loss as a regularization term to achieve phased hybrid training.
7. The acoustic sensor array optimization method based on neural network-enhanced NSGA-II according to claim 6, characterized in that: A hybrid precision parallel evaluation strategy is adopted after the intelligent population is initialized by a solution generator network. Specifically, in each generation of evaluation, a dynamic scheduling mechanism is used, in which most individuals are evaluated quickly by a proxy evaluation network; at the same time, a small number of individuals and randomly sampled individuals are selected and submitted to a high-fidelity simulator for evaluation.
8. The acoustic sensor array optimization method based on neural network-enhanced NSGA-II according to claim 7, characterized in that: Generation G population Each individual in The probability of it being assigned to use high-fidelity simulation The following rules determine the dynamic nature of the situation: ; In the formula, For the first i The probability of each individual being assigned to be evaluated using a high-fidelity simulator. For indicator functions, when the first i Non-dominated ranking of individuals based on the performance of the previous generation Better than or equal to the former If the name is specified, use 1; otherwise, use 0. Where N represents the proportion of elites and N is the population size. For the first i The crowding degree of each individual in the decision space, or the minimum distance from other individuals, and normalized to Intervals are used to encourage exploration of sparse regions; To assess the average prediction uncertainty of the network near the current Gth generation population; These are weighting coefficients used to control the relative importance of three factors: elite priority, diversity exploration, and uncertainty correction; according to The evaluation path for each individual is determined by roulette wheel or direct threshold method. After the evaluation is completed, all individuals obtain their objective function value vector.
9. The acoustic sensor array optimization method based on neural network-enhanced NSGA-II according to claim 8, characterized in that: When iteratively selecting a population across multiple generations until convergence or the termination condition is met, an environment selection phase is required. The environment selection phase operates on the population that has undergone mixed evaluation. and its offspring The main process is as follows: merge: , This refers to the candidate population after merging; Non-dominated sorting: merge candidate populations All solutions are hierarchically arranged according to Pareto dominance relations; Crowding Calculation: Within the same undominated layer, calculate the crowding distance of each solution in the target space to measure the distribution density of the solutions; Elite Selection: Select the top-performing solutions according to the undominated layers from best to worst, while also considering the crowding density within each layer from highest to lowest. Individuals form a new generation population, where N is the preset population size.
10. The acoustic sensor array optimization method based on neural network-enhanced NSGA-II according to claim 9, characterized in that: The implementation of directed genetic evolution using a policy generation network includes: inputting parent gene fragments, local environmental context, global optimization state, and evolutionary stage information into the policy generation network; the crossover policy head in the dual-branch policy head outputs a crossover mask probability vector, and for each sensor dimension, performs standard binary crossover or directly inherits parent values according to probability; the mutation policy head outputs a mutation adjustment vector and adaptive variance coefficients, and performs biased mutation operations.
