Digital equivalent point source sound field reconstruction method based on weighted truncation threshold

Abstract

The invention discloses a digital equivalent point source sound field reconstruction method based on a weighted truncation threshold, and the method comprises the steps: S1, arranging a measurement plane at one side of a target sound source, distributing M measurement points, and collecting sound pressure; s2, introducing a reinforcement learning mechanism into a target sound source surface to determine equivalent point source distribution, and generating N global equivalent point sources by adopting a low-density grid to construct a sound field model; s3, obtaining source intensity statistical characteristics based on the optimized self-attention mechanism, adjusting a sparse solution in the sound field model based on the source intensity statistical characteristics to obtain optimal source intensity statistical characteristics, improving the sound field model according to the optimal source intensity statistical characteristics, and performing optimal coarse grid global modeling according to the improved sound field model; and S4, obtaining optimal local encryption region details by adopting a multi-scale equivalent point source layout, and realizing high-precision reconstruction of a complex sound field by combining optimal coarse grid global modeling and local encryption region detail capture.

Description

Technical Field

[0001] The present invention relates to the technical field of sound field reconstruction, and particularly relates to a digital equivalent point source sound field reconstruction method based on weighted truncation threshold. Background Art

[0002] With the increasing demand for high-precision sound field reconstruction in fields such as industrial noise diagnosis and automotive NVH development, the core challenge lies in the contradiction between resolution and the number of sampling points in traditional sound field reconstruction. After the introduction of compressive sensing theory, researchers have improved the reconstruction accuracy under limited sampling through sparsity constraints. However, existing methods have low reconstruction efficiency when lacking prior information, and the computational complexity surges in large model scenarios, making it difficult to meet industrial requirements in terms of convergence and real-time performance, which has become a bottleneck for popularization in high-complexity scenarios.

[0003] Currently, in the prior art, traditional sound field reconstruction methods usually rely on regular grids (such as uniform rectangular grids), and cannot adapt to complex characteristics such as asymmetry and strong directivity of the target sound source, which will lead to unreasonable equivalent point source densities in key regions (such as the edge of the sound source, interference peaks and valleys), easily missing important sound field features or introducing redundant calculations, and only describing the source strength distribution through simple statistics (maximum value, standard deviation), ignoring the spatial correlation between equivalent point sources and the influence of position information on the sound field, resulting in inaccurate sparse solution optimization, incomplete removal of weak sources or incorrect deletion of strong sources, affecting the reconstruction accuracy, and lacking hierarchical processing of the global contour and local details in complex sound fields. Either the global fine grid calculation amount explodes and it is difficult to reconstruct in real time, or the high-frequency components (such as diffraction, scattering) are lost due to local simplification. Therefore, a digital equivalent point source sound field reconstruction method based on weighted truncation threshold is proposed here to solve the above problems. Summary of the Invention

[0004] In order to overcome the above defects of the prior art and achieve the above object, the present invention proposes the following technical solutions: A digital equivalent point source sound field reconstruction method based on weighted truncation threshold, comprising: S1: Arrange a measurement plane on one side of the target sound source, distribute M measurement points and collect sound pressure; S2: Introduce a reinforcement learning mechanism on the target sound source surface to determine the equivalent point source distribution, and use a low-density grid to generate N global equivalent point sources to construct a sound field model; S3: Obtain source strength statistical features based on an optimized self-attention mechanism, adjust the sparse solution in the sound field model based on the source strength statistical features to obtain the optimal source strength statistical features, improve the sound field model according to the optimal source strength statistical features, and perform optimal coarse grid global modeling according to the improved sound field model; S4: Use a multi-scale equivalent point source layout to obtain the best local encrypted area details, and combine the optimal coarse grid global modeling with the local encrypted area detail capture to achieve the reconstruction of complex sound fields. The optimized self-attention mechanism is implemented by introducing position encoding on the basis of the original self-attention mechanism.

[0005] The process of introducing the reinforcement learning mechanism to determine the equivalent point source distribution is as follows: Define the agent as the equivalent point source generation policy; Take the current equivalent point source distribution and the sound field simulation error on the equivalent point source surface as the state space; Take adding, removing, or adjusting the equivalent point source positions on the equivalent point source surface as the action space; Take the degree of reduction of the sound field simulation error as the reward function; The agent interacts with the environment for learning, updates the generation policy using the policy gradient algorithm, and determines the final equivalent point source distribution through multiple iterative updates.

[0006] The environment is defined as the sound field radiated by the target sound source, and the equivalent point source generation policy is defined as the equivalent point source distribution.

[0007] The process of constructing the sound field model is as follows: Based on the final equivalent point source distribution, select a low-density grid and determine the grid spacings in two directions direction and direction to construct the equivalent point source surface; Obtain global equivalent point sources based on the equivalent point source surface, and calculate the sound pressure at each point in the space based on the global equivalent point sources to construct the sound field model.

[0008] The process of obtaining the source strength statistical features based on the optimized self-attention mechanism is as follows: Define a position encoding vector, add the position encoding to the source strength vector, and then construct improved query vectors, key vectors, and value vectors based on the source strength vector; Based on the attention score matrix obtained from the improved key vectors, obtain the attention degree between different source strength elements, and obtain the enhanced source strength representation through the attention score matrix and the improved value vectors; Obtain the source strength statistical features based on the enhanced source strength representation.

[0009] The process of obtaining the optimal source strength statistical features is as follows: Take the source strength statistical features as points in the high-dimensional space, perform manifold learning using isometric mapping to obtain the geodesic distance between two source strength statistical features; Based on the geodesic distance, use multidimensional scaling to map the high-dimensional source strength data to the low-dimensional manifold space to obtain the local density; Set an adjustment rule including a truncation threshold based on the local density, and optimize the sparse solution based on the adjustment rule to obtain the optimal source strength statistical features.

[0010] The process of improving the sound field model according to the optimal source strength statistical characteristics is as follows: Based on the source strength in the sound field model, obtain the source strength statistical characteristics by optimizing the self-attention mechanism; Adjust the source strength statistical characteristics based on the adjustment rule to obtain the optimal source strength statistical characteristics; Based on the optimal source strength statistical characteristics, recalculate the sound pressure on the measurement plane and perform multiple iterations to obtain the improved sound field model.

[0011] The process of performing optimal coarse grid global modeling according to the improved sound field model is as follows: Perform coarse grid division based on the improved sound field model; For each coarse grid unit in the divided coarse grid, traverse the equivalent point source position information in the improved sound field model to obtain the number of equivalent point sources located within the grid; Obtain the optimal source strength within the coarse grid based on the number of equivalent point sources; Equivalent each coarse grid unit to an equivalent point source and take the center of the coarse grid unit as the optimal position; Perform global modeling based on the optimal position and the optimal source strength and obtain the sound pressure of the optimal coarse grid global modeling.

[0012] The process of obtaining the details of the optimal local encryption area is as follows: Define the local encryption area, arrange relatively sparse equivalent point sources in the local encryption area, and gradually encrypt the relatively sparse equivalent point sources according to the complexity of the regional sound field to obtain the local encryption area.

[0013] The process of realizing the reconstruction of the complex sound field is as follows: For the equivalent point sources in the local encryption area, establish the transfer relationship with the measurement points on the measurement plane and obtain the corresponding sound pressure Green's function according to the acoustic theory; Obtain the sound pressure generated by the equivalent point sources in the local encryption area based on the corresponding sound pressure Green's function; Then add the sound pressure generated by the equivalent point sources in the local encryption area to the sound pressure obtained by the optimal coarse grid global modeling to obtain the final reconstructed sound pressure and complete the sound field reconstruction.

[0014] The present invention has the following beneficial effects: In the present invention, firstly, through the interactive learning between the intelligent agent and the sound field radiated by the target sound source, the position of the equivalent point source is dynamically adjusted (such as aggregating towards the high-energy radiation area), which solves the layout deficiency of the traditional fixed grid for asymmetric and strongly directional sound sources; Secondly, position encoding is introduced into the self-attention mechanism. By embedding the spatial order of point sources through sine and cosine functions, the model can capture the synergistic effect between source strength and position (such as the enhancement effect of adjacent point sources on interference peaks). Compared with traditional calculation methods (only calculating simple statistics), the enhanced source strength representation fuses correlation information, and the maximum value and standard deviation extracted can more realistically reflect the source strength distribution. Moreover, the local density piecewise linear threshold based on manifold learning can dynamically adjust the truncation strategy according to the source strength data distribution, set sensitive slopes in complex density intervals, avoid feature loss caused by a global unified threshold, and achieve a balance between preserving details in key regions and improving efficiency in sparse regions. Finally, the improved sound field model is divided into uniform and non-uniform coarse grids. Each unit is equivalent to the optimal source strength of the central source strength, greatly reducing the computational scale, quickly constructing the overall contour of the sound field, effectively capturing low-frequency and large-scale propagation characteristics. At the same time, for regions with large sound pressure gradients, new equivalent point sources are inserted through multi-scale, retaining high-frequency components (the retention rate of >800Hz is increased from 65% to 92%), making the peak position error of the reconstructed sound field <0.01m and the interference valley depth error <5%, realizing accurate simulation of the details of the sound field in complex scenarios. Description of the Drawings

[0015] Figure 1 It is a method step diagram of a digital equivalent point source sound field reconstruction method based on weighted truncation threshold proposed by the present invention. Detailed Embodiments

[0016] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0017] Embodiment: As Figure 1 shown, a digital equivalent point source sound field reconstruction method based on weighted truncation threshold proposed by the present invention includes: S1: Arrange a measurement plane on one side of the target sound source, distribute M measurement points and collect sound pressure; Arrange a measurement plane on one side of the target sound source. Assume the measurement plane is a two-dimensional plane, on which measurement points are distributed, and microphones are arranged at corresponding positions on each measurement point to obtain the sound pressure at each measurement point on the measurement plane ; Specifically, the distribution method of the measurement points is irregularly arranged according to the characteristics of the sound source such as symmetry or directivity. For example, for a sound source with axial symmetry characteristics, the measurement points are symmetrically distributed on both sides of the axis of symmetry, and the sound pressure is collected It is the basic input for subsequent sound field reconstruction. These data contain information such as the spatial distribution, frequency characteristics, and intensity of the sound source, and can reflect the distance relationship between the measurement points and the sound source as well as the intensity of the sound source.

[0018] S2: Introduce a reinforcement learning mechanism on the target sound source surface to determine the equivalent point source distribution, and use a low-density grid to generate N global equivalent point sources to construct a sound field model; Arrange an equivalent point source surface on the target sound source surface, and arrange equivalent point sources on the equivalent point source surface to simulate the sound field radiated by the target sound source. The sound pressure on the measurement plane is the superposition of the sound fields generated by the global equivalent point sources distributed on the equivalent point source surface. The equivalent point source distribution is obtained by introducing reinforcement learning; The process of introducing a reinforcement learning mechanism to determine the equivalent point source distribution is as follows: Define the intelligent agent as the equivalent point source generation strategy (equivalent point source distribution); Take the current equivalent point source distribution and the sound field simulation error on the equivalent point source surface as the state space; Take adding, removing, or adjusting the positions of equivalent point sources on the equivalent point source surface as the action space; Take the degree of reduction of the sound field simulation error as the reward function; The intelligent agent interacts and learns with the environment (the sound field radiated by the target sound source), and uses the policy gradient algorithm to update the generation strategy (equivalent point source distribution), and adaptively generates equivalent point sources. The equivalent point source position update formula is:

[0019] where is the position of the th equivalent point source at the th step, is the learning rate, is the (the position of the th equivalent point source) cumulative reward function based on the policy; By continuously iterating and updating the positions of the equivalent point sources, the distribution of the equivalent point sources is gradually optimized to better simulate the sound field radiated by the target sound source. After multiple iterations, the distribution of the equivalent point sources tends to be stable, and the equivalent point source distribution at this time is the finally determined equivalent point source distribution; Specifically, by dynamically optimizing the equivalent point source distribution through reinforcement learning, the problem that the traditional fixed grid layout cannot adapt to the complex characteristics of the sound source is solved; The process of the sound field model is as follows: The process of generating N global equivalent points using a low-density grid based on the final equivalent point source distribution is as follows: Based on the final equivalent point-source distribution, a low-density grid is selected, and the grid spacings in two directions in the grid are determined Direction and the grid spacing in the direction. The grid spacing is taken as one-tenth of the characteristic length of the sound source to construct an equivalent point-source surface; Starting from the vertex of the equivalent point-source surface, equivalent point sources are arranged. In the direction, an equivalent point source is arranged every . In the direction, a row of equivalent point sources is arranged every . Suppose there are point sources arranged in the direction and point sources arranged in the direction. Then the total number of global equivalent point sources is ; Specifically, in combination with the generation of the low-density grid, while ensuring the calculation efficiency, accurate simulation of the radiation sound field of the target sound source is achieved. The low-density grid reduces the number of equivalent point sources and the computational complexity. The spacing of one-tenth of the characteristic length can ensure a certain sparsity of the point-source distribution and retain sufficient spatial resolution in the key area (near the sound source), avoiding the loss of sound-field characteristics due to too sparse a grid. Finally, by obtaining the global equivalent point sources, it is ensured that the equivalent point-source surface is covered by the system, avoiding simulation blind spots; Based on global equivalent point sources, the sound pressure at each point in space is calculated to construct a sound-field model. Suppose there are measurement points on the measurement plane, the position of the equivalent point source is , and the source strength of the th equivalent point source is . Then the sound pressure at the measurement point on the measurement plane is expressed as:

[0020] where is the sound pressure at the measurement point on the measurement plane, is the Green's function at the th measurement point, representing the sound propagation characteristics from the equivalent point-source position to the spatial point ; For the free-space sound field, corresponding to the measurement points on the measurement plane, the sound pressure at all measurement points is obtained to achieve a preliminary simulation of the radiation sound field of the target sound source and obtain a sound-field model; Specifically, traditional equivalent point source distributions usually adopt a regular grid layout, lacking the ability to adaptively adjust to the actual characteristics of sound sources. By introducing a reinforcement learning mechanism into the process of determining equivalent point source distributions, interactive learning between the agent and the radiation sound field of the target sound source is achieved, breaking the static and fixed equivalent point source arrangement mode, and better reflecting the radiation characteristics of the target sound source to construct a sound field model.

[0021] S3: Based on the optimized self-attention mechanism, obtain the source strength statistical features. Based on the source strength statistical features, adjust the sparse solution in the sound field model to obtain the optimal source strength statistical features. According to the optimal source strength statistical features, improve the sound field model, and perform optimal coarse grid global modeling based on the improved sound field model; The process of obtaining the source strength statistical features based on the optimized self-attention mechanism is as follows: After obtaining the final equivalent point source distribution through reinforcement learning optimization and constructing the sound field model, analyze the source strength of the equivalent point sources. For each equivalent point source , mine and obtain the source strength statistical features based on an optimized self-attention mechanism; The process of mining the source strength statistical features based on the optimized self-attention mechanism is as follows: When mining the source strength statistical features, introduce the self-attention mechanism. Traditional mining methods usually only calculate simple statistics such as the maximum value and standard deviation of the source strength, while the self-attention mechanism enables the mining method to automatically focus on the correlations between different elements in the source strength vector. However, in the self-attention mechanism, the elements in the source strength vector originally lack position information, which will lead to ignoring the influence of the spatial position correlations of the equivalent point sources on the source strength when mining features. Therefore, introduce position encoding on the basis of the original self-attention mechanism; Define a position encoding vector , the th ( ) position encoding vector is expressed as: , ; Among them, is the dimension of the source strength vector, represents the dimension index; Add the position encoding to the source strength vector: , and then construct improved query vectors, key vectors, and value vectors based on the source strength vector ; Among them, the improved query vector , is the weight matrix of the initial query vector, the improved key vector , is the weight matrix of the initial key vector, the improved value vector , is the weight matrix of the initial value query vector. Based on the improved query vector, key vector, and value vector, when calculating the attention score, the self-attention mechanism can consider the position information of the equivalent point sources and extract more spatially correlated source strength statistical features; Then calculate the attention score matrix , where is the improved key vector is the dimension of is the key vector is the transpose of The obtained attention score matrix represents the degree of attention between different source strength elements. Then, through the attention score matrix and the improved value vector are combined with each other to obtain the strengthened source strength representation The strengthened source strength representation represents the fusion of the correlation information and position information between source strength elements. Based on the strengthened source strength representation the source strength statistical features are obtained; Specifically, based on the strengthened source strength representation the source strength statistical features are obtained. First, find the element with the largest absolute value in , which reflects the maximum magnitude in the source strength. Then, obtain the dispersion degree of the source strength relative to the average value. Combine these two statistical features, the maximum element and the dispersion degree, to comprehensively reflect the distribution characteristics of the source strength. At the same time, because the strengthened source strength representation fuses the correlation and position information, compared with simply calculating simple statistics for the source strength vector in the traditional way, the obtained features can better reflect the spatial correlation between equivalent point sources and the actual distribution of the source strength; The process of obtaining the optimal source strength statistical features is as follows: First, perform a manifold mapping on the source strength vector. Take the obtained source strength statistical features as a point in a high-dimensional space, and use the Isometric Mapping (ISOMAP) for manifold learning. For two source strength statistical features and , obtain the geodesic distance between source strength vectors; Based on the geodesic distance , use the multi-dimensional scaling method to map the high-dimensional source strength data to a low-dimensional manifold space. In the low-dimensional manifold space, define the local density function to measure the local density of the source strength data as ; Specifically, let a point in the low-dimensional space be , and its local density is obtained by calculating the reciprocal of the average distance of neighboring points around it, which is expressed as:

[0022] Among them, represents the local density of a point in the low-dimensional space ; represents the neighbor set of the -th low-dimensional space point, and is the number of neighbor set points; Based on the local density, an adjustment rule is set, and a truncation threshold is set. In each acquisition of the source strength statistical feature, for the source strength of the -th equivalent point source in the -th iteration, if , then set the source strength of this equivalent point source to 0. If , then retain the source strength of this equivalent point source to obtain the optimal source strength statistical feature ; Specifically, the truncation threshold is obtained based on the local density. The value range of the local density is divided into segments, denoted as . For the local density of the -th segment, the calculation formula of the truncation threshold

[0023] is: where is the initial threshold corresponding to the -th segment, and is the slope of the -th segment. By means of piecewise linearity, different threshold adjustment strategies are set for different local density intervals. In the density interval where the source strength data changes more complexly, a more sensitive threshold adjustment slope can be set; Through the above adjustment rule, the weak sources in the source strength vector are gradually removed, and the strong sources that contribute more to the sound field are retained, so as to optimize the sparse solution and obtain the optimal source strength statistical feature ; The process of improving the sound field model according to the optimal source strength statistical feature is as follows: After the sound field model is constructed, start iteratively adjusting the sparse solution. Based on the source strength in the sound field model, obtain the source strength statistical feature through optimizing the self-attention mechanism; Based on the adjustment rule, adjust the source strength statistical feature in the sound field model to obtain the optimal source strength statistical feature ; Recalculate the sound pressure on the measurement plane. Through multiple iterations, the weak sources in the source strength vector are gradually removed, and the strong sources that contribute more to the sound field are retained, obtaining an improved sound field model; Specifically, by iteratively adjusting the sparse solution, the weak sources in the source strength vector are gradually removed using a truncation threshold, and the strong sources that contribute more to the sound field are retained. Each iteration is based on the source strength statistical features obtained by optimizing the self-attention mechanism and recalculate the sound pressure on the measurement plane, making the model gradually focus on the key source strengths, reducing the interference of weak sources, and more accurately reflecting the radiation characteristics of the target sound source. At the same time, for the weak source pairs that contribute little to the sound field but still consume resources during calculation, the weak sources are removed through the truncation threshold, which can simplify the source strength vector and reduce the subsequent calculation amount. In multiple iterations, only the strong sources are retained for calculation, avoiding a large amount of invalid calculations. While ensuring the accuracy, the calculation efficiency of the improved sound field model is greatly improved, making it suitable for the fast reconstruction scenario of complex sound fields in the later stage; The process of performing optimal coarse grid global modeling based on the improved sound field model is as follows: Based on the improved sound field model, perform coarse grid division, and determine the size and shape of the coarse grid according to the shape, size of the target sound source, and the distribution characteristics of the sound field; For example, for a sound source with a relatively regular shape, use a uniform rectangular coarse grid, and for a sound source with a complex shape, use an adaptive non-uniform coarse grid; Let the coarse grid unit of the divided coarse grid be . For each coarse grid unit , traverse the equivalent point source position information in the improved sound field model to obtain the number of equivalent point sources located within this grid , add up the optimal source strength statistical features within each coarse grid and divide by the number of equivalent point sources, and then obtain the optimal source strength within each coarse grid unit ; Equivalent each coarse grid unit to an equivalent point source, and take the center of each coarse grid unit as the optimal position ; Based on the optimal position and optimal source strength of the equivalent point sources, establish a global model of the entire sound field. When calculating the sound pressure at a certain point on the measurement plane in the subsequent calculation, it is obtained by superimposing the sound pressures generated by all equivalent point sources, that is:

[0024] where, is the sound pressure obtained by optimal coarse grid global modeling, is the optimal source strength of the equivalent point source within the i-th coarse grid unit, is the number of coarse grids, is the optimal position of the source strength of the i-th equivalent point source, For the Green's function of sound propagation in the optimal coarse grid global modeling, complete the global modeling of the optimal coarse grid; Specifically, by traversing the equivalent point source position information in the improved sound field model, obtain the number of equivalent point sources and the optimal source strength in each coarse grid cell, simplify multiple equivalent point sources into one equivalent point source, reduce the computational complexity, and at the same time retain the main information of the contribution of the equivalent point sources in this area to the sound field (reflected by the optimal source strength). Since the sound pressure contributions of each equivalent point source to the measurement point can be linearly superposed, the coarse grid cell is equivalent to an equivalent point source, and the sound pressure is calculated using this principle, which can not only ensure the physical accuracy of the calculation, but also significantly improve the calculation efficiency due to the reduction in the number of equivalent point sources (compared with the original equivalent point sources), and the sound pressure on the measurement plane can be obtained quickly.

[0025] S4: Use a multi-scale equivalent point source layout to obtain the details of the best local refinement area, and realize the reconstruction of the complex sound field by combining the optimal coarse grid global modeling and the capture of local refinement area details; Determine the areas with large sound pressure gradients and complex sound field changes as local refinement areas; Specifically, because these local refinement areas contribute more to the sound field details and require a more refined equivalent point source layout. For example, at the edges of sound sources, near reflecting surfaces, etc., the sound pressure changes violently. If no refinement treatment is carried out, it will lead to the loss of sound field details and the inability to accurately simulate the real sound field. By focusing on such areas, it is ensured that the subsequent equivalent point source layout can capture the key sound field characteristics; In the local refinement area, arrange relatively sparse equivalent point sources. Let the number of equivalent point sources of the relatively sparse equivalent point sources be , the position be , and the source strength be , and gradually refine the relatively sparse equivalent point sources according to the complexity of the regional sound field to obtain the local refinement area; Specifically, the implementation method of refining the relatively sparse equivalent point sources is to insert new equivalent point sources between the relatively sparse equivalent point sources. The positions and source strengths of the new equivalent point sources are determined by analyzing the distribution of the original equivalent point sources and the characteristics of the regional sound field. Let the number of inserted equivalent point sources be , the position be , and the source strength be ; The process of realizing the reconstruction of the complex sound field is as follows: For the equivalent point sources in the local refinement area, establish the transfer relationship with the measurement points on the measurement plane; The transfer relationship is: Let the position of the th equivalent point source in the local refinement area be , and the position of the measurement point on the measurement plane be , and then obtain the corresponding sound pressure Green's function according to the acoustic theory; Measuring the sound pressure generated by the equivalent point sources in the locally refined region on the corresponding sound pressure Green's function measurement plane:

[0026] where is the total source strength in the locally refined region, is the Green's function in the locally refined region; Then, add the sound pressure generated by the locally refined region to the sound pressure obtained from the optimal coarse grid global modeling to obtain the final reconstructed sound pressure, expressed as:

[0027] where is the sound pressure obtained from the optimal coarse grid global modeling, is the sound pressure generated by the locally refined region; Specifically, the optimal coarse grid global modeling provides the overall framework of the sound field and the low-frequency, large-scale characteristics, and the sound pressure of the locally refined region supplements the complex details and high-frequency characteristics. Through the combination of the two, high-precision reconstruction of the complex sound field in the full frequency band and the full space is achieved; Example: In a scenario, the target sound source is two symmetrically distributed monopole speakers (frequency 1000 Hz, spacing 0.5 m), located in free space. The measurement plane is 0.3 m away from the sound source plane, and 32 measurement points are arranged (symmetrically distributed on both sides of the symmetry axis, a total of 16 pairs of measurement points). The characteristic length of the sound source L = 0.5 m (speaker spacing); Through S1: The measurement points are symmetrically distributed on both sides of the symmetry axis (x-axis), with a spacing of 0.05 m, a total of 32 points. The microphone collects the sound pressure at each point, including the direct wave and interference wave information; where: The correlation coefficient between the measured data and the theoretical sound pressure reaches 0.92, proving that the data effectively captures the sound field characteristics; Through S2: 100 equivalent point sources are initially randomly distributed by the agent. Using the sound field simulation error (L2 norm of the measured and simulated sound pressures) as the reward function, after 50 iterations of the policy gradient algorithm, the equivalent point sources gather towards the speaker positions; The low-density grid spacing is taken as L / 10 = 0.05 m, and 11×11 = 121 global equivalent point sources are generated to construct the sound field model; where: The root mean square error between the simulated sound pressure and the measured sound pressure is reduced from the initial 0.35 Pa to 0.12 Pa, proving that reinforcement learning effectively optimizes the point source distribution. The low-density grid retains the key sound field characteristics while reducing the computational amount (the computational time is reduced by 36% compared to the traditional uniform grid); Through S3: The position encoding embeds the point source position sequence numbers through sine and cosine functions to construct query, key, and value vectors. The attention score matrix shows that the attention weights of adjacent point sources are higher (the weights of point sources near the speaker are greater than 0.8, and those of far-field point sources are less than 0.2). The enhanced source strength representation fuses the position correlation, and the maximum value and standard deviation are extracted as features; Local density segmentation, high-density area (near the sound source) threshold , low-density area (far field) , and 37 strong sources are retained after pruning (accounting for 30%); Among them: The model parameters are reduced by 70% after pruning, but the RMSE further drops to 0.08 Pa, proving that the dynamic threshold effectively removes weak sources, retains key strong sources, and the self-attention mechanism captures the spatial correlation (simulation of adjacent source strength synergistically enhancing interference peaks); Through S4: Coarse grid division: A 10×10 uniform rectangular coarse grid (grid size 0.05 m×0.05 m) is adopted for the regular sound source area. Each coarse grid cell is equivalent to an equivalent point source at the center position, and the source strength takes the optimal source strength within the cell; Local encryption: The interference peak area is encrypted to a spacing of 0.025 m, and 50 new sources are inserted. The source strength is determined by interpolating the adjacent source strengths; Sound pressure superposition: The global coarse grid sound pressure and the local encrypted sound pressure are added to finally reconstruct the sound pressure; The peak position error of the reconstructed sound field is <0.01 m, and the interference valley depth error is <5%. The calculation efficiency is increased by 3 times compared with the full fine grid (the encrypted area accounts for 20% of the calculation amount); Performance comparison table of the sound field reconstruction scheme:

[0028] Reconstruction RMSE (Pa): RMSE is the root mean square error, which is used to measure the deviation between the reconstructed sound field and the real sound field. The smaller the value, the higher the reconstruction accuracy; Calculation time (seconds): It reflects the duration required to complete the sound field reconstruction calculation. The shorter the time, the higher the efficiency; Number of equivalent point sources (pieces): Equivalent point sources are virtual point sources used to simulate the radiation sound field of real sound sources, and the number affects the calculation complexity; High-frequency component retention (>800HZ): It refers to the retention ratio of high-frequency acoustic wave information above 800 Hz in the reconstructed sound field. The higher the ratio, the better the restoration of high-frequency sound details (such as sharp noises, complex acoustic interferences, etc.);

[0029] The finally fused sound field realizes the fine simulation of local details and complex acoustic phenomena, achieving the high-precision reconstruction of complex sound fields.

[0030] In the application, several formulas involved are all calculated by taking their numerical values after dimensionless treatment. The establishment of the formula is obtained by collecting a large amount of data for software simulation to get a formula closest to the actual situation. Some coefficients or weights in the formula are set by those skilled in the art according to the actual situation, so no more details will be given here.

[0031] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution.

[0032] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A digital equivalent point source sound field reconstruction method based on weighted truncation threshold, characterized in that Including: S1: Arrange a measurement plane on one side of the target sound source, distribute M measurement points and collect sound pressure; S2: Introduce a reinforcement learning mechanism on the target sound source surface to determine the equivalent point source distribution, and use a low-density grid to generate N global equivalent point sources to construct a sound field model; S3: Obtain source strength statistical features based on the optimized self-attention mechanism, adjust the sparse solution in the sound field model based on the source strength statistical features to obtain the optimal source strength statistical features, improve the sound field model according to the optimal source strength statistical features, and perform optimal coarse grid global modeling according to the improved sound field model; S4: Use a multi-scale equivalent point source layout to obtain the details of the best local encryption area, and combine the optimal coarse grid global modeling with the capture of local encryption area details to realize the reconstruction of a complex sound field; The optimized self-attention mechanism is realized by introducing position encoding on the basis of the original self-attention mechanism.

2. The digital equivalent point source sound field reconstruction method based on weighted truncation threshold according to claim 1, characterized in that The process of introducing a reinforcement learning mechanism to determine the equivalent point source distribution is as follows: Define the intelligent agent as the equivalent point source generation strategy; Take the current equivalent point source distribution and sound field simulation error on the equivalent point source surface as the state space; Take adding, removing or adjusting the position of the equivalent point source on the equivalent point source surface as the action space; Take the reduction degree of the sound field simulation error as the reward function; The intelligent agent interacts with the environment to learn, updates the generation strategy using the policy gradient algorithm, and determines the final equivalent point source distribution through multiple iterative updates.

3. A digital equivalent point source sound field reconstruction method based on a weighted truncation threshold according to claim 2, characterized in that, The environment is defined as the sound field radiated by the target sound source, and the equivalent point source generation strategy is defined as the equivalent point source distribution.

4. A digital equivalent point source sound field reconstruction method based on a weighted truncation threshold, characterized in that The construction process of the sound field model is as follows: Based on the final equivalent point source distribution, a low-density grid is selected, and the grid spacings in two directions in the grid are determined direction and direction to construct an equivalent point source surface; Obtained based on the equivalent point source surface global equivalent point sources, and based on the global equivalent point sources, calculate the sound pressure at each point in space to construct an acoustic field model.

5. A digital equivalent point source sound field reconstruction method based on a weighted truncation threshold, characterized in that The process of obtaining source strength statistical features based on the optimized self-attention mechanism is as follows: Define a position encoding vector, add the position encoding to the source strength vector, and then construct improved query vectors, key vectors and value vectors based on the source strength vector; Based on the attention degree between different source strength elements in the attention score matrix obtained from the improved key vector, obtain the enhanced source strength representation through the attention score matrix and the improved value vector; Obtain source strength statistical features based on the enhanced source strength representation.

6. A digital equivalent point source sound field reconstruction method based on a weighted truncation threshold according to claim 5, characterized in that, The process of obtaining the optimal source strength statistical features is as follows: Take the source strength statistical features as points in a high-dimensional space, perform manifold learning using isometric mapping to obtain the geodesic distance between two source strength statistical features; Based on the geodesic distance, use the multi-dimensional scaling method to map the high-dimensional source strength data to a low-dimensional manifold space to obtain the local density; Set an adjustment rule including a truncation threshold based on the local density, optimize the sparse solution based on the adjustment rule, and obtain the optimal source strength statistical features.

7. A method for reconstructing a digital equivalent point source sound field based on a weighted truncation threshold, characterized in that The process of improving the sound field model according to the optimal source strength statistical features is as follows: Obtain source strength statistical features based on the source strength in the sound field model through the optimized self-attention mechanism; Adjust the source strength statistical features based on the adjustment rule to obtain the optimal source strength statistical features; Recalculate the sound pressure on the measurement plane based on the optimal source strength statistical features and perform multiple iterations to obtain the improved sound field model.

8. A digital equivalent point source sound field reconstruction method based on a weighted truncation threshold according to claim 1, characterized in that The process of performing optimal coarse grid global modeling according to the improved sound field model is as follows: Perform coarse grid division based on the improved sound field model; For each coarse grid unit in the divided coarse grid, traverse the equivalent point source position information in the improved sound field model to obtain the number of equivalent point sources located in this grid; Obtain the optimal source strength within the coarse grid based on the number of equivalent point sources; Equivalent each coarse grid cell to an equivalent point source and take the center of the coarse grid cell as the optimal position; Conduct global modeling based on the optimal position and optimal source strength and obtain the sound pressure of the optimal coarse grid global modeling.

9. A digital equivalent point source sound field reconstruction method based on a weighted truncation threshold, characterized in that The process of obtaining the details of the optimal local refinement area is as follows: Define the local refinement area, arrange relatively sparse equivalent point sources in the local refinement area, and gradually refine the relatively sparse equivalent point sources according to the complexity of the regional sound field to obtain the local refinement area.

10. A digital equivalent point source sound field reconstruction method based on a weighted truncation threshold according to claim 1, characterized in that, The process of realizing the reconstruction of the complex sound field is as follows: For the equivalent point sources within the local refinement area, establish the transfer relationship with the measurement points on the measurement plane and obtain the corresponding sound pressure Green's function according to the acoustic theory; Obtain the sound pressure generated by the equivalent point sources in the local refinement area based on the corresponding sound pressure Green's function; Then add the sound pressure generated by the equivalent point sources in the local refinement area to the sound pressure obtained from the optimal coarse grid global modeling to obtain the final reconstructed sound pressure and complete the sound field reconstruction.

Images