Sound source separation system and method based on orthogonalized source components and spatial entropy constraint

Abstract

The present application relates to the technical field of acoustic signal processing and deep learning, in particular to a sound source separation system and method based on orthogonalized source components and spatial entropy constraint, comprising collecting acoustic measurement data, performing feature decomposition on the acoustic measurement data to generate orthogonalized virtual source components; outputting a predicted sound source spatial energy distribution map through a deep neural network; calculating a spatial entropy loss, weighting and summing the spatial entropy loss and a reconstruction error loss to obtain a total loss, updating network parameters of the deep neural network using the total loss to obtain a trained deep neural network, and outputting an updated predicted sound source spatial energy distribution map; and performing channel energy statistics on the predicted sound source spatial energy distribution map to determine the number of effective sound sources. The present application effectively solves the key problems of poor physical interpretability, preset source number, insufficient post-processing robustness, and limited model generalization ability in sound source separation.

Description

Technical Field

[0001] This invention relates to the fields of acoustic signal processing and deep learning technology, and more specifically, to a sound source separation system and method based on orthogonalized source components and spatial entropy constraints. Background Technology

[0002] Current acoustic imaging source separation techniques primarily rely on eigenvalue decomposition (such as PCA and EVD) of the cross-spectral matrix of microphone array signals. While these methods can generate a set of orthogonal "virtual" sound source components, their inherent limitations restrict practical applications: First, the virtual sources obtained through decomposition are products of mathematical abstraction and cannot form a clear, spatially interpretable energy distribution map that directly corresponds to the real physical sound sources, making the results difficult to use for engineering diagnostics. Second, these methods typically require pre-setting the number of sound sources, resulting in poor applicability in real-world multi-source dynamic scenarios. Furthermore, traditional post-processing methods (such as optimization algorithms based on minimum spatial entropy or spatial orthogonality) for recovering physical sources from virtual sources are computationally complex, parameter-sensitive, and lack robustness, exhibiting a sharp performance degradation under noise interference or complex acoustic environments, and exhibiting limited generalization ability.

[0003] In recent years, deep learning-based sound source separation methods have provided new insights into the aforementioned problems, but significant limitations remain. These methods typically use virtual sources as input and learn the mapping to physical sources through neural networks. However, their performance is highly dependent on the realism and diversity of the simulation training data, making them prone to overfitting to simulation settings and resulting in weak generalization ability on real measurement data (such as wind tunnel experiments and industrial sites). Furthermore, existing deep learning frameworks lack clearly defined physical constraints on the output results, failing to fundamentally address the ambiguity of sound source spatial representation and making it difficult to guarantee that the separation results conform to the physical laws of sound propagation, thus limiting their application in zero-sample, high-reliability scenarios. Therefore, a unified deep post-processing framework is needed to map orthogonalized representations to physical sound sources, possessing interpretability, generalization ability, and adaptive source number capabilities. Therefore, this paper proposes a sound source separation system and method based on orthogonalized source components and spatial entropy constraints. Summary of the Invention

[0004] The purpose of this invention is to provide a sound source separation system and method based on orthogonalized source components and spatial entropy constraints, so as to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, one of the objectives of this invention is to provide a sound source separation system based on orthogonalized source components and spatial entropy constraints, including an orthogonalized source generation module, a deep post-processing module, a priori constraint construction module, a source arrangement resolution module, and a source number estimation module.

[0006] The orthogonalization source generation module is used to collect acoustic measurement data, perform feature decomposition on the acoustic measurement data, and generate orthogonalization virtual source components.

[0007] The deep post-processing module is used to receive orthogonalized virtual source components, perform nonlinear mapping through a deep neural network, and output a predicted spatial energy distribution map of the sound source.

[0008] The prior constraint construction module is used to calculate the spatial entropy loss based on the predicted sound source spatial energy distribution map, perform arrangement matching between the predicted sound source spatial energy distribution map and the preset real source target amplitude tensor and calculate the reconstruction error loss, and obtain the total loss by weighted summation of the spatial entropy loss and the reconstruction error loss. The total loss is used to update the network parameters of the deep neural network to obtain the trained deep neural network and output the updated predicted sound source spatial energy distribution map.

[0009] The source arrangement resolution module is used to receive the predicted sound source spatial energy distribution map output by the trained deep neural network and the preset real source target amplitude tensor, perform arrangement matching, and generate the sound source estimation result with the order aligned.

[0010] The source quantity estimation module is used to perform channel energy statistics on the predicted sound source spatial energy distribution map output by the trained deep neural network to determine the number of effective sound sources.

[0011] Furthermore, the orthogonalization source generation module is used to acquire acoustic measurement data, perform feature decomposition on the acoustic measurement data, and generate orthogonalized virtual source components, including:

[0012] Acquire the sound pressure time-domain signals of each channel of the microphone array, apply a window function to the sound pressure time-domain signals and perform a short-time Fourier transform to obtain the complex sound pressure spectrum matrix at each frequency point;

[0013] Multiply the complex acoustic compression spectrum matrix by its conjugate transpose and average it over all time frames to obtain the cross-spectral matrix;

[0014] Perform eigenvalue decomposition on the cross-spectral matrix to obtain an orthogonal matrix composed of eigenvectors and a diagonal matrix composed of eigenvalues;

[0015] Based on the geometric position of the microphone array and the discrete grid of the source plane, a Green's function matrix is ​​constructed in conjunction with the sound propagation model, and the pseudo-inverse of the Green's function matrix is ​​calculated as the acoustic backpropagation operator.

[0016] Multiplying the eigenvector matrix by the eigenvalue diagonal matrix yields the weighted eigenvector matrix;

[0017] The weighted eigenvector matrix is ​​multiplied by the acoustic backpropagation operator and mapped onto the discrete grid of the source plane to generate a spatial distribution map of the orthogonalized virtual source components.

[0018] Furthermore, the deep post-processing module is used to receive orthogonalized virtual source components, perform nonlinear mapping through a deep neural network, and output a predicted sound source spatial energy distribution map, including:

[0019] Identify the maximum value in the amplitude tensor of the orthogonalized virtual source component of the input, divide the amplitude tensor by the maximum value, and generate the normalized input tensor;

[0020] The normalized input tensor is input into the Transformer-enhanced U-Net neural network, and convolution and spatial downsampling are performed on each source channel of the normalized input tensor to generate encoder feature maps.

[0021] At each jump connection in the encoder-decoder path, the encoder feature map of the current level is input into the Transformer module. The correlation weights between positions in the sequence are calculated through a self-attention mechanism based on scaling dot products, and then weighted summation is performed. The enhanced feature sequence is generated through feedforward neural network processing, resulting in a feature map processed by Transformer.

[0022] At each level of the decoder path, the feature map of the previous level is upsampled, and the upsampled result is concatenated with the feature map of the corresponding level processed by Transformer in the channel dimension to generate a fused feature map.

[0023] A two-dimensional convolution operation is performed on the fused feature map output from the last layer of the decoder path to map the number of channels to the preset number of output channels K, generating a normalized preliminary predicted spatial energy distribution map of the sound source.

[0024] The normalized preliminary predicted spatial energy distribution map of the sound source is multiplied by the maximum value of the corresponding input case stored, and then rescaled to generate the predicted spatial energy distribution map of the sound source.

[0025] Furthermore, the prior constraint construction module is used to calculate the spatial entropy loss based on the predicted sound source spatial energy distribution map, perform permutation matching between the predicted sound source spatial energy distribution map and the preset real source target amplitude tensor, calculate the reconstruction error loss, weighted sum the spatial entropy loss and the reconstruction error loss to obtain the total loss, update the network parameters of the deep neural network using the total loss, obtain the trained deep neural network, and output the updated predicted sound source spatial energy distribution map, including:

[0026] The amplitude of each output source channel in the predicted sound source spatial energy distribution map is squared to obtain the pixel energy value of each output source channel; an arithmetic addition operation is performed on all pixel energy values ​​of each output source channel to obtain the total energy of each output source channel.

[0027] Divide the energy value of each pixel in each output source channel by the total energy of that channel to obtain the normalized spatial energy distribution of each output source channel.

[0028] Multiply the normalized spatial energy distribution of each output source channel by its natural logarithm to obtain the entropy component of each channel.

[0029] Perform arithmetic addition on all entropy components of each output source channel and take the inverse to obtain the spatial entropy value of each output source channel;

[0030] Perform arithmetic addition on the spatial entropy values ​​of all output source channels to obtain the spatial entropy loss term;

[0031] Based on the predicted spatial energy distribution map of the sound source and the amplitude tensor of the real source target, the Hungarian matching algorithm is applied to find the permutation matrix that minimizes the L1 norm distance between the two.

[0032] The permutation matrix is ​​used to rearrange the predicted spatial energy distribution map of the sound sources;

[0033] The L1 norm distance between the rearranged predicted sound source spatial energy distribution map and the pre-acquired real source target amplitude tensor is calculated to obtain the Hungarian L1 loss term.

[0034] The total loss function is obtained by performing arithmetic addition on the results of multiplying the Hungarian L1 loss term and the spatial entropy loss term by the preset regularization weight coefficient.

[0035] The gradient is calculated using the total loss function, and the network parameters of the deep neural network in the deep post-processing module are updated. After training, the predicted spatial energy distribution map of the sound source generated by the trained deep neural network is output.

[0036] Furthermore, the preset regularization weight coefficients are obtained through the following steps:

[0037] Multiple network models are initialized and trained using different candidate regularized weight coefficients from the set {0,1,2,5,8,10,20,30,40,50,100}.

[0038] On the experimental validation dataset with coherent source configuration, the median spatial correlation of the output results of each trained network model is calculated;

[0039] The candidate regularization weight coefficients that maximize the median spatial correlation are selected as the preset regularization weight coefficients.

[0040] Furthermore, the source arrangement resolution module is used to receive the predicted sound source spatial energy distribution map output by the trained deep neural network and the preset real source target amplitude tensor, perform arrangement matching, and generate a sequentially aligned sound source estimation result, including:

[0041] A channel distribution from the predicted sound source spatial energy distribution map is selected sequentially and paired with a channel distribution from the pre-acquired real source target amplitude tensor to form a calculation pair;

[0042] For each calculation pair, perform element-by-element subtraction on the two sets of distribution data to obtain the difference distribution data;

[0043] Perform the absolute value operation on all elements in each difference distribution data to obtain the absolute difference distribution data;

[0044] Sum all elements in each absolute difference distribution data point to obtain a scalar distance value;

[0045] Arrange all the calculated scalar distance values ​​in the order of the selected predicted sound source channels as rows and the order of the selected real source target channels as columns to generate the first distance matrix;

[0046] Input the first distance matrix into the Hungarian matching algorithm, perform the optimal allocation solution, and generate the second allocation matrix;

[0047] Based on the second allocation matrix, a permutation operation is performed on all channels of the predicted sound source spatial energy distribution map to generate sequentially aligned sound source estimation results.

[0048] Furthermore, element-wise subtraction is performed on the two sets of distribution data in each calculation pair to obtain the difference distribution data, including:

[0049] Extract the two-dimensional distribution data of the currently selected channel from the predicted spatial energy distribution map of the sound source;

[0050] Extract the two-dimensional distribution data of the currently selected channel from the amplitude tensor of the real source target;

[0051] The values ​​at the same spatial pixel positions of the two extracted two-dimensional distribution data are subtracted to obtain difference distribution data with the same dimension.

[0052] Furthermore, the source quantity estimation module is used to perform channel energy statistics on the predicted sound source spatial energy distribution map output by the trained deep neural network to determine the number of effective sound sources, including:

[0053] The sum of squared amplitudes of all pixels in each channel of the predicted sound source spatial energy distribution map is calculated to obtain the total energy value of each channel;

[0054] Based on the total energy value of all channels, the maximum value is identified as the energy reference benchmark;

[0055] The relative energy ratio of each channel is obtained by comparing the total energy value of each channel with the energy reference benchmark.

[0056] The relative energy ratio is compared with a preset energy threshold to filter out channels whose relative energy ratio is greater than the threshold.

[0057] The number of filtered channels is counted, and the count is output as an estimate of the number of effective sound sources.

[0058] The second objective of this invention is to provide a sound source separation method based on orthogonalized source components and spatial entropy constraints, and a sound source separation system based on orthogonalized source components and spatial entropy constraints as described in any one of the above, comprising:

[0059] S1. Acquire acoustic measurement data, perform feature decomposition on the acoustic measurement data, and generate orthogonal virtual source components;

[0060] S2. Receive the orthogonalized virtual source components, perform nonlinear mapping through a deep neural network, and output a predicted spatial energy distribution map of the sound source.

[0061] S3. Calculate the spatial entropy loss based on the predicted sound source spatial energy distribution map, perform arrangement matching between the predicted sound source spatial energy distribution map and the preset real source target amplitude tensor and calculate the reconstruction error loss, and obtain the total loss by weighted summation of the spatial entropy loss and the reconstruction error loss. Use the total loss to update the network parameters of the deep neural network to obtain the trained deep neural network and output the updated predicted sound source spatial energy distribution map.

[0062] S4. Receive the predicted sound source spatial energy distribution map output by the trained deep neural network and the preset real source target amplitude tensor, perform arrangement matching, and generate the sound source estimation results with the order aligned.

[0063] S5. Perform channel energy statistics on the predicted sound source spatial energy distribution map output by the trained deep neural network to determine the number of effective sound sources.

[0064] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0065] 1. By introducing spatial entropy as a physical prior constraint, the spatial interpretability and structural clarity of the sound source separation results are significantly improved, effectively compressing the uncertainty of sound source energy distribution and enabling the output sound source to have a clear physical location. This method constructs a unified deep post-processing framework, supports adaptation to various network structures, and utilizes orthogonalized source components as array geometry-independent intermediate representations, enhancing the system's adaptability to different microphone arrays and frequency ranges.

[0066] 2. This method achieves adaptive estimation of the number of sound sources without requiring a preset number of sound sources, significantly improving the system's practicality and intelligence. It also possesses strong zero-shot generalization capability, allowing direct application to complex acoustic environments such as real wind tunnels and industrial testing after training only on simulation data, effectively overcoming the dependence of traditional methods on the realism of simulation data. This framework can also be seamlessly integrated with existing beamforming systems, further improving the quality of traditional acoustic imaging results through transfer learning. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of the overall structure of the present invention;

[0068] Figure 2 This is a flowchart illustrating the orthogonalization source generation module of the present invention.

[0069] Figure 3 This is a flowchart illustrating the deep post-processing module of the present invention.

[0070] Figure 4 This is a flowchart illustrating the prior constraint construction module of the present invention.

[0071] Figure 5 This is a flowchart illustrating the source arrangement elimination module of the present invention;

[0072] Figure 6 This is a flowchart illustrating the source quantity estimation module of the present invention. Detailed Implementation

[0073] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0074] like Figures 1-6 As shown, one of the objectives of this invention is to provide a sound source separation system based on orthogonalized source components and spatial entropy constraints, comprising the following steps:

[0075] The orthogonalization source generation module is used to collect acoustic measurement data, perform feature decomposition on the acoustic measurement data, and generate orthogonalization virtual source components;

[0076] The orthogonalization source generation module is used to collect acoustic measurement data, perform feature decomposition on the acoustic measurement data, and generate orthogonalized virtual source components, including:

[0077] The sound pressure time-domain signals of each channel of the microphone array are acquired, a window function is applied to the sound pressure time-domain signals and a short-time Fourier transform is performed to obtain the complex sound pressure spectrum matrix at each frequency point.

[0078] Sound signals are synchronously acquired through each channel of the microphone array to obtain the time-domain sound pressure level (SPL) signal for each channel. A window function is applied to the time-domain signal of each channel to reduce spectral leakage, and then a short-time Fourier transform is performed to convert the time-domain signal into a frequency-domain signal. At each frequency point, the complex spectral values ​​of all channels at that frequency are arranged to obtain a complex SPL matrix, where the rows of the matrix correspond to the microphone channels and the columns correspond to the time frames.

[0079] Multiply the complex acoustic compression matrix by its conjugate transpose and average it over all time frames to obtain the cross-spectral matrix.

[0080] For each frequency point, the complex acoustic compression spectrum matrix is ​​multiplied by its conjugate transpose to obtain a preliminary matrix. Then, this preliminary matrix is ​​arithmetically averaged over all time frames to obtain the cross-spectral matrix corresponding to that frequency point. This matrix reflects the spatial correlation between signals received by different microphone channels.

[0081] Perform eigenvalue decomposition on the cross-spectral matrix to obtain an orthogonal matrix composed of eigenvectors and a diagonal matrix composed of eigenvalues.

[0082] Eigenvalue decomposition is performed on the cross-spectral matrix. This decomposition operation breaks down the cross-spectral matrix into the product of three matrices: an orthogonal matrix composed of eigenvectors, a diagonal matrix composed of eigenvalues, and the conjugate transpose of the orthogonal matrix. The eigenvectors span the subspace of the signal, and each eigenvector is considered an orthogonal "virtual" sound source component; the eigenvalues ​​represent the energy magnitude of the corresponding virtual source.

[0083] Based on the geometric position of the microphone array and the discrete grid of the source plane, a Green's function matrix is ​​constructed in conjunction with the sound propagation model, and the pseudo-inverse of the Green's function matrix is ​​calculated as the acoustic backpropagation operator.

[0084] First, based on the actual geometric coordinates of the microphone array in space and the pre-defined discrete grid coordinates of the source plane, combined with the physical model of sound wave propagation, the acoustic transfer function from each source grid point to each microphone is calculated. These transfer functions are then arranged to form the Green's function matrix. Next, a Moore-Penrose pseudo-inverse operation is performed on this Green's function matrix to obtain the acoustic backpropagation operator. This operator is used to reverse-engineer the signal observed at the microphone back to the possible sound source location on the source plane.

[0085] Multiplying the eigenvector matrix by the eigenvalue diagonal matrix yields the weighted eigenvector matrix.

[0086] Perform matrix multiplication on the eigenvector matrix and the eigenvalue diagonal matrix. This operation weights each eigenvector (the spatial pattern of the virtual source) with its corresponding eigenvalue (energy magnitude), resulting in an energy-weighted eigenvector matrix, which provides energy-information-containing input for subsequent mapping to the source plane.

[0087] The weighted eigenvector matrix is ​​multiplied by the acoustic backpropagation operator and mapped onto the discrete grid of the source plane to generate a spatial distribution map of the orthogonalized virtual source components.

[0088] The weighted eigenvector matrix is ​​multiplied by the acoustic backpropagation operator. This operation projects each weighted virtual source eigenvector (dimension corresponding to microphone channels) onto a discrete spatial grid on the source plane through the backpropagation operator, thereby generating a set of spatial distribution maps. Each map corresponds to the estimated energy distribution of an orthogonalized virtual source component on the source plane, serving as the initial source map for subsequent deep learning network inputs.

[0089] The deep post-processing module is used to receive orthogonalized virtual source components, perform nonlinear mapping through a deep neural network, and output a predicted spatial energy distribution map of the sound source.

[0090] The deep post-processing module receives orthogonalized virtual source components, performs nonlinear mapping through a deep neural network, and outputs a predicted spatial energy distribution map of the sound source, including:

[0091] Identify the maximum value in the amplitude tensor of the orthogonalized virtual source component of the input, divide the amplitude tensor by the maximum value, and generate the normalized input tensor;

[0092] The system receives the amplitude tensor of the orthogonalized virtual source components from the orthogonalization source generation module. This tensor contains multiple channels, each a two-dimensional image representing the energy distribution of the virtual source. First, it iterates through all values ​​in the tensor, performing a maximum value lookup operation to find the global maximum amplitude. Then, it divides every value in the entire amplitude tensor by this found maximum value, performing element-wise division. This operation linearly scales all values ​​to the range of zero to one, resulting in a normalized input tensor. The purpose of this step is to eliminate differences in absolute energy magnitude between different samples, providing a scale-uniform and stable input basis for subsequent neural network processing.

[0093] The normalized input tensor is input into the Transformer-enhanced U-Net neural network, and convolution and spatial downsampling are performed on each source channel of the normalized input tensor to generate encoder feature maps.

[0094] The normalized input tensor obtained in the previous step is fed into the encoder part of a pre-defined Transformer-enhanced U-Net neural network model. The encoder consists of multiple layers connected sequentially. In each layer, the network independently performs a 2D convolution operation on each source channel of the input tensor (i.e., each 2D image), using a small square convolution kernel sliding across the image to extract local spatial features. Next, spatial downsampling is performed, for example by increasing the convolution stride or using pooling methods, to reduce the spatial dimensions (height and width) of the feature image while increasing the number of feature channels. This process proceeds progressively from high resolution to low resolution, generating a series of encoder feature maps with different spatial scales, each capturing the spatial patterns of the input data at different levels of abstraction.

[0095] At each hop connection in the encoder-decoder path, the encoder feature map of the current level is input into the Transformer module. The correlation weights between positions in the sequence are calculated through a self-attention mechanism based on scaling dot products, and then weighted and summed. The enhanced feature sequence is generated through a feedforward neural network, resulting in a feature map processed by the Transformer.

[0096] In the U-Net network architecture, the feature map generated by each stage of the encoder is passed to the corresponding layer of the decoder through a shortcut (called a skip connection). Here, the encoder feature map from the current skip connection is fed into a Transformer module for processing. Specifically, the feature map is first unfolded in the spatial dimension, transforming it into a feature sequence, where each element corresponds to a spatial location in the original image and its features across all channels. The core of the Transformer module is a self-attention mechanism based on scaled dot products, which computes a set of query, key, and value representations for each location in the sequence.

[0097] Then, by calculating the similarity (dot product) between the query and all keys, and after scaling and normalization, the attention weights of that location to all other locations are obtained. Finally, these weights are used to sum all the values, allowing information from each location to be integrated with features from the global context. This output is then non-linearly enhanced by a feedforward neural network. The processed feature sequence is reassembled into a format with the same spatial dimensions as the original feature map, resulting in a Transformer-processed feature map. This step aims to explicitly model the long-range dependencies between any two spatial locations in the image.

[0098] At each level of the decoder path, the feature map of the previous level is upsampled, and the upsampled result is concatenated with the feature map of the corresponding level processed by Transformer in the channel dimension to generate a fused feature map.

[0099] In the decoder path of the network, each stage receives a feature map from the previous stage. First, an upsampling operation is performed on this feature map, such as using transposed convolution or interpolation, doubling its spatial size. Simultaneously, a feature map corresponding to the current decoder level, already processed by the Transformer, is obtained from the skip connections. This feature map has the same spatial size as the upsampled feature map. Then, these two feature maps are concatenated along the channel dimension, connecting all their channels together to generate a fused feature map with more channels. This operation effectively combines the more abstract semantic information recovered by the decoder with the globally relationally enhanced detail information preserved by the encoder.

[0100] A two-dimensional convolution operation is performed on the fused feature map output from the last layer of the decoder path to map the number of channels to the preset number of output channels K, generating a normalized preliminary predicted spatial energy distribution map of the sound source.

[0101] The final layer output of the decoder path produces a fused feature map whose spatial dimensions have been restored to the same level as the original input image. A final two-dimensional convolution operation is applied to this feature map. This convolution uses a 1x1 kernel, the number of which is preset to a fixed value, namely the number of output channels K (this value is usually consistent with the number of input virtual source channels, for example, set to ten during training). The function of this 1x1 convolution is to linearly combine the features of all channels at each spatial point and map them to K independent output channels. After this operation, the output tensor is a set of K two-dimensional images, called the normalized preliminary predicted sound source spatial energy distribution map. Each channel in this map corresponds to a separated estimated sound source, with pixel values ​​between zero and one, representing the normalized relative energy intensity.

[0102] The normalized preliminary predicted spatial energy distribution map of the sound source is multiplied by the maximum value of the corresponding input case stored, and then rescaled to generate the predicted spatial energy distribution map of the sound source.

[0103] In this step, the maximum value of the original input amplitude for the current processing case, stored in the first step, is read. Then, each pixel value in the normalized preliminary prediction map obtained in the previous step is multiplied by this stored maximum value, and element-wise multiplication is performed to rescale it. After this operation, the values ​​of the prediction map are restored to the original physical amplitude dimensions, generating a predicted spatial energy distribution map of the sound source.

[0104] The prior constraint construction module is used to calculate the spatial entropy loss based on the predicted sound source spatial energy distribution map, perform arrangement matching between the predicted sound source spatial energy distribution map and the preset real source target amplitude tensor and calculate the reconstruction error loss, and obtain the total loss by weighted summation of the spatial entropy loss and the reconstruction error loss. The total loss is used to update the network parameters of the deep neural network to obtain the trained deep neural network and output the updated predicted sound source spatial energy distribution map.

[0105] The prior constraint construction module is used to calculate the spatial entropy loss based on the predicted sound source spatial energy distribution map, perform permutation matching between the predicted sound source spatial energy distribution map and the preset real source target amplitude tensor, calculate the reconstruction error loss, weight and sum the spatial entropy loss and the reconstruction error loss to obtain the total loss, update the network parameters of the deep neural network using the total loss, obtain the trained deep neural network, and output the updated predicted sound source spatial energy distribution map, including:

[0106] The pixel energy value of each output source channel is obtained by squaring the amplitude of each output source channel in the predicted sound source spatial energy distribution map.

[0107] Receive the predicted spatial energy distribution map of the sound source from the depth post-processing module. For each output channel in the map (i.e., each predicted sound source energy distribution map), perform a square operation on the amplitude of each pixel. This operation converts the sound pressure amplitude into an energy value, conforming to the physical principle that sound energy density is proportional to the square of the amplitude, preparing for subsequent calculations of the spatial probability distribution.

[0108] Perform arithmetic addition on the energy values ​​of all pixels in each output source channel to obtain the total energy of each output source channel.

[0109] For the pixel energy map of each channel obtained in the previous step, calculate the sum of the energy values ​​of all pixels in the map. This sum represents the total energy of the predicted sound source corresponding to that channel across the entire spatial range, and it will be used as a normalization factor in the next calculation.

[0110] Divide the energy value of each pixel in each output source channel by the total energy of that channel to obtain the normalized spatial energy distribution of each output source channel.

[0111] The energy value of each pixel in each channel is divided by the total energy of that channel, performing an element-wise division operation. After this operation, each channel is transformed into a normalized spatial probability distribution map, where the sum of all pixel values ​​is 1. This step directly implements the following formula for calculating the... The source is in the location probability The formula is as follows:

[0112] ;

[0113] Where, the denominator represents the first... The predicted sound source channel is in the first... The amplitude at each spatial grid point is the sum of the energy of all pixels in that channel, with the denominator being the total energy of all pixels in that channel.

[0114] Multiplying the normalized spatial energy distribution of each output source channel by its natural logarithm yields the entropy component of each channel.

[0115] For each channel's normalized probability distribution, the natural logarithm of the probability value for each pixel is calculated, and then the two are multiplied together. This operation is a core component of calculating Shannon entropy.

[0116] Perform arithmetic addition on all entropy components of each output source channel and take the opposite to obtain the spatial entropy value of each output source channel.

[0117] Summing all entropy components for each channel, then inverting the sum (multiplying by -1). This step is formulated as follows, yielding the... Spatial entropy of a sound source:

[0118] ;

[0119] This value quantitatively describes the spatial concentration of sound source energy. The smaller the value, the more concentrated the energy.

[0120] Perform arithmetic addition on the spatial entropy values ​​of all output source channels to obtain the spatial entropy loss term.

[0121] The total spatial entropy loss term is calculated by taking all... The spatial entropy values ​​of the output channels are summed. The formula for this step is as follows, and it constitutes the spatial entropy loss term:

[0122] ;

[0123] This loss term is used to penalize energy-dispersed source estimations during training, encouraging the network to generate spatially compact source distribution maps.

[0124] Secondly, the reconstruction error loss term is calculated. This step aims to measure the pixel-level difference between the predicted and ground truth images, while also addressing the permutation ambiguity caused by the uncertain order of the network output channels.

[0125] Based on the predicted spatial energy distribution map of the sound source and the amplitude tensor of the real source target, the Hungarian matching algorithm is used to find the permutation matrix that minimizes the L1 norm distance between the two.

[0126] Simultaneously holding both the predicted image and the actual sound source target image. Since the order of the neural network output channels is arbitrary, it is necessary to find the optimal match between the predicted and actual channels. The Hungarian matching algorithm is invoked, which searches all possible channel permutations to find one that minimizes the L1 norm distance (i.e., the sum of the absolute differences of all corresponding pixels) between the rearranged predicted image and the actual image.

[0127] The core formula corresponding to this optimization objective is:

[0128] ;

[0129] in, It is a prediction of the spatial energy distribution map of the sound source (three-dimensional tensor). It is the magnitude tensor of the real source target (three-dimensional tensor). It is the permutation matrix to be solved. Represents all possible A set of permutation matrices, L1 norm (sum of absolute values ​​over all elements) represents the formula for finding a permutation. This minimizes the difference between the rearranged true image and the predicted image.

[0130] The permutation matrix is ​​used to rearrange the predicted spatial energy distribution map of the sound source.

[0131] Based on the optimal permutation matrix found by the Hungarian algorithm in the previous step, the channel order of the predicted sound source spatial energy distribution map is rearranged to obtain a predicted map that is aligned with the channel order of the actual target.

[0132] The L1 norm distance between the rearranged predicted sound source spatial energy distribution map and the pre-acquired real source target amplitude tensor is calculated to obtain the Hungarian L1 loss term.

[0133] The absolute differences between all corresponding pixels in the aligned predicted image and the ground truth target image are calculated, and these absolute values ​​are summed to obtain a scalar loss value. The minimum value in this calculation is denoted as the Hungarian L1 loss term, which measures the pixel-level reconstruction error between the prediction and the ground truth.

[0134] The total loss function is obtained by performing arithmetic addition on the results of multiplying the Hungarian L1 loss term and the spatial entropy loss term by the preset regularization weight coefficient.

[0135] The reconstruction error term (Hungarian L1 loss) is combined with the spatial compactness constraint term (spatial entropy loss). Specifically, the spatial entropy loss term is multiplied by a preset regularization weight coefficient, and then added to the Hungarian L1 loss term. This step is formulated as follows, forming the total loss function used for training:

[0136] ;

[0137] Among them, hyperparameters This is used to balance the goals of reconstruction accuracy and spatial compactness. Its optimal value can be determined through ablation experiments (e.g., using the value with the highest median spatial correlation on coherent source configuration experimental data).

[0138] The gradient is calculated using the total loss function, and the network parameters of the deep neural network in the deep post-processing module are updated. After training, the predicted spatial energy distribution map of the sound source generated by the trained deep neural network is output.

[0139] During the backpropagation process of network training, the total loss function calculated in the aforementioned steps is used as the optimization objective. The gradient of this loss with respect to all network weight parameters is calculated using automatic differentiation, and an optimizer (such as Adam) is used to update the parameters according to the gradient direction. Through iterative optimization, the network simultaneously learns the ability to accurately reconstruct sound sources and output a compact distribution, thereby achieving high-quality sound source separation. After training, the network outputs the predicted spatial energy distribution map of the sound sources generated by the trained deep neural network.

[0140] The preset regularization weight coefficients are obtained through the following steps:

[0141] Multiple network models are initialized and trained using different candidate regularized weight coefficients from the set {0,1,2,5,8,10,20,30,40,50,100}.

[0142] To find the optimal regularization weight coefficients, a grid search experiment is performed. This step selects a predefined set of candidate values ​​{0,1,2,5,8,10,20,30,40,50,100}. This set is designed to cover a wide range from no regularization to strong regularization (λ=100), and centers on values ​​validated in the documentation to ensure that inflection points in performance changes are captured. For each candidate value in the set, a complete model training process is performed: a new network model instance is created using the same random seed or initialization strategy, the value is substituted into the total loss function, and then training is performed on the same simulated training dataset for a specified number of epochs. This results in a set (e.g., 11) of trained models with identical structures but varying performance due to differences in the strength of spatial entropy constraints imposed during training.

[0143] On the experimental validation dataset with coherent source configuration, the median spatial correlation of the output results of each trained network model is calculated.

[0144] To evaluate the model's generalization and separation capabilities under different values, it was tested on a specific and more challenging evaluation scenario—the "coherent source configuration" experimental validation dataset. This dataset, derived from real-world experiments, consists of spatially coherent multiple sound sources, unlike the simple point sources used in training, effectively testing the model's ability to handle unknown source features. For each trained model, it was applied to all test samples on this dataset. For each separated sound source in each sample, the spatial correlation between its output distribution and the true distribution was calculated. Then, for all spatial correlation values ​​obtained by a model on the dataset (across all samples and all sources), the median spatial correlation value was calculated. This median spatial correlation value, as a robust statistic, represents the model's average separation accuracy at the corresponding value.

[0145] The candidate regularization weight coefficients that maximize the median spatial correlation are selected as the preset regularization weight coefficients.

[0146] Collect and compare the median spatial correlation values ​​corresponding to all candidate values. Perform a maximum value search operation to identify the value that produces the highest median spatial correlation. This value represents the regularization weight that achieves the best balance between "promoting spatial compactness" and "maintaining reconstruction accuracy" under the experimental conditions. Select and fix this value as a preset regularization weight coefficient for all subsequent training and applications involving spatial entropy loss. For example, when λ=5, the highest average spatial correlation is achieved with the coherent source configuration, so λ=5 can be determined as the optimal preset value. This method determines the most effective strength of the regularization term in the total loss function in a data-driven manner.

[0147] The source arrangement resolution module is used to receive the predicted sound source spatial energy distribution map output by the trained deep neural network and the preset real source target amplitude tensor, perform arrangement matching, and generate the sound source estimation result with the order aligned.

[0148] The source arrangement resolution module receives the predicted sound source spatial energy distribution map output by the trained deep neural network and the preset real source target amplitude tensor, performs arrangement matching, and generates sequentially aligned sound source estimation results, including:

[0149] A channel distribution from the predicted sound source spatial energy distribution map is selected sequentially and paired with a channel distribution from the pre-obtained real source target amplitude tensor to form a calculation pair.

[0150] The system receives a predicted spatial energy distribution map of the sound sources (containing K channels) and a pre-stored amplitude tensor of the real source targets (also containing K channels) from the deep post-processing module. To address the problem of uncertain channel order in the neural network output (orthogonal ambiguity), it is necessary to find the optimal one-to-one correspondence between the predicted and real channels. For this purpose, a double-loop traversal operation is performed. First, using the first channel of the predicted map as a reference, it is paired sequentially with the 1st, 2nd, up to the Kth channel of the real map, forming K computational pairs. Then, the process is repeated for the second channel of the predicted map, forming another K computational pairs. This process continues until all K channels of the predicted map have undergone this operation, ultimately generating a total of K×K computational pairs. Each computational pair contains a predicted sound source distribution map and a real sound source distribution map for subsequent difference comparisons.

[0151] For each calculation pair, perform element-by-element subtraction on the two sets of distribution data to obtain the difference distribution data.

[0152] For each computational pair, pixel-level numerical comparisons are performed. The pixel values ​​at the same spatial coordinates in the predicted distribution map and the true distribution map are subtracted. This process is repeated for each pixel in both maps, generating a new two-dimensional dataset of the same dimension, called the difference distribution data. This data records the numerical deviation between the prediction and the actual values ​​at each point in space.

[0153] Perform the absolute value operation on all elements in each difference distribution data to obtain the absolute difference distribution data.

[0154] Since the resulting differences can be positive or negative, to measure the magnitude of the differences without them canceling out due to the sign, an absolute value operation is performed on each difference distribution data point. This involves iterating through each element in the data (i.e., the difference at each pixel) and converting it to a non-negative absolute value, thus obtaining the absolute difference distribution data. This operation prepares for calculating the L1 norm distance.

[0155] The summation of all elements in each absolute difference distribution data is performed to obtain a scalar distance value.

[0156] To quantify the overall difference between two images in a computational pair, a global summation operation is performed on the absolute difference distribution data. This involves iterating through all pixels with absolute differences in the data and summing them all to obtain a single scalar value. This value represents the L1 norm distance between that specific predicted channel and that specific true channel, indicating the pixel-level total error of the prediction result based on this correspondence.

[0157] Arrange all the calculated scalar distance values ​​in the order of the selected predicted sound source channels as rows and the order of the selected real source target channels as columns to generate the first distance matrix.

[0158] All K×K scalar distance values ​​are organized into a matrix according to a systematic rule. The row indices of the matrix correspond to the order of the selected prediction channels (row 1 corresponds to prediction channel 1, row 2 corresponds to prediction channel 2, ...), and the column indices correspond to the order of the real channels selected later in step (1) (column 1 corresponds to real channel 1, column 2 corresponds to real channel 2, ...). Thus, the element in the i-th row and j-th column of the matrix represents the L1 distance value when "predicted channel i" is paired with "real channel j". This matrix is ​​called the first distance matrix or cost matrix, which fully defines the "cost" required to allocate all prediction channels to all real channels for every possible allocation scheme.

[0159] Input the first distance matrix into the Hungarian matching algorithm, perform the optimal allocation solution, and generate the second allocation matrix.

[0160] Taking the first distance matrix as input, the standard Hungarian matching algorithm (also known as the Kuhn-Munkres algorithm) is invoked. This algorithm is designed to solve this type of linear assignment problem. The algorithm operates automatically, aiming to find an optimal one-to-one assignment scheme (i.e., each predicted channel uniquely corresponds to one real channel, and each real channel is also uniquely corresponding to one predicted channel) that minimizes the sum of the distance values ​​(costs) corresponding to all selected assignment pairs. The output of the algorithm is a new matrix called the second assignment matrix (or permutation matrix). This matrix is ​​typically also a K×K matrix, but its elements are only 0 or 1, where the position of the number 1 indicates an optimal correspondence (e.g., if the 2nd row and 3rd column is 1, it means that the optimal solution is to assign predicted channel 2 to real channel 3).

[0161] Based on the second allocation matrix, a permutation operation is performed on all channels of the predicted sound source spatial energy distribution map to generate sequentially aligned sound source estimation results.

[0162] Based on the optimal correspondence encoded by the second allocation matrix, the original predicted sound source spatial energy distribution map with arbitrary channel order is rearranged. Specifically, the second allocation matrix is ​​checked to find the position (i,j) of each element that is 1. This indicates that the i-th channel in the original prediction map should be placed in the position of the j-th channel in the output new map. Following this mapping, all K predicted channels are rearranged to generate a new prediction map whose channel order is perfectly aligned with the amplitude tensor of the true source target. After this step, the arrangement ambiguity is eliminated, and the prediction results can be directly used for accurate channel-by-channel comparison with the true values ​​and subsequent loss calculations.

[0163] Extract the two-dimensional distribution data of the currently selected channel from the predicted spatial energy distribution map of the sound source.

[0164] Based on the predicted channel index i specified in the "computation pair" being processed, data is extracted from the three-dimensional predicted sound source spatial energy distribution tensor. Specifically, based on index i, the i-th channel slice of the tensor is located and accessed to obtain a two-dimensional array of size [H, W] (H is the spatial height and W is the width). This array represents the two-dimensional distribution data of the current predicted channel, and each element stores the estimated sound pressure or energy amplitude at that spatial grid point.

[0165] Extract the two-dimensional distribution data of the currently selected channel from the amplitude tensor of the real source target.

[0166] Simultaneously, based on the real channel index j specified in the current "computation pair", data is extracted from the three-dimensional real source target amplitude tensor. Specifically, based on index j, the j-th channel slice of the tensor is located and accessed to obtain a two-dimensional array of the same size [H, W]. This array represents the two-dimensional distribution data of the current real channel, with each element storing the known real sound source amplitude at the corresponding spatial location.

[0167] The values ​​at the same spatial pixel positions of the two extracted two-dimensional distribution data are subtracted to obtain difference distribution data with the same dimension.

[0168] Element-wise subtraction is performed to quantify the local differences between the two images. Specifically, each identical spatial coordinate (h, w) in both 2D arrays is simultaneously traversed (where h is the row index and w is the column index). At each position, the value of the predicted array at that position is subtracted from the value of the true array at that position. After traversing all H×W positions, a new 2D array of size [H, W] is generated. This new array is the difference distribution data, where each element records the deviation (positive, negative, or zero) between the predicted and true values ​​at a specific spatial point. This data is an intermediate result necessary for subsequent calculation of the L1 norm distance between the channel pairs (i.e., the sum of the absolute deviations at all positions).

[0169] The source quantity estimation module is used to perform channel energy statistics on the predicted sound source spatial energy distribution map output by the trained deep neural network to determine the number of effective sound sources.

[0170] The source quantity estimation module is used to perform channel energy statistics on the predicted sound source spatial energy distribution map output by the trained deep neural network to determine the number of effective sound sources, including:

[0171] The sum of squared amplitudes of all pixels in each channel of the predicted sound source spatial energy distribution map is calculated to obtain the total energy value of each channel;

[0172] The system receives the output from the deep post-processing module—a predicted spatial energy distribution map of the sound source (a 3D tensor containing K channels, each a 2D image). For each channel of this tensor (indexed from 1 to K), the following operations are performed: First, the energy value of each pixel (sound pressure amplitude) in the 2D image of that channel is squared; then, the energy values ​​of all pixels are summed globally. This summation result is the total energy value of that channel. It represents the total energy estimated by the network and attributed to the "sound source" corresponding to that channel. This calculation is performed sequentially for all K channels to obtain K total energy values, forming an energy list.

[0173] Based on the total energy value of all channels, the maximum value is identified as the energy reference benchmark;

[0174] Among the calculated total energy values ​​for the K channels, a maximum value lookup operation is performed. By iterating through and comparing all energy values, the largest value is found. This maximum value is selected as the energy reference.

[0175] The relative energy ratio of each channel is obtained by comparing the total energy value of each channel with the energy reference benchmark.

[0176] Perform a normalized comparison. For each channel (i=1,…,K), divide its total energy value by the found energy reference benchmark (maximum value), performing a division operation once. The result is called the relative energy ratio of that channel, with a value between 0 and 1 (the ratio of the maximum value channel itself is 1). This ratio reflects how weak the energy of the "sound source" represented by that channel is relative to the strongest sound source.

[0177] The relative energy ratio is compared with a preset energy threshold to filter out channels whose relative energy ratio is greater than the threshold.

[0178] A preset energy threshold (e.g., 0.1, 0.05, a decimal between 0 and 1) is used as the judgment threshold. This threshold is set based on prior knowledge or experimental experience and is used to distinguish between valid signals and noise residue. The relative energy ratio of each channel is compared with this preset threshold one by one. If the relative energy ratio of a channel is greater than the threshold, it is determined that the channel is likely to correspond to a real valid sound source; otherwise, the channel is considered to have weak energy, possibly originating from noise or separation residue, and should be considered invalid. Through this comparison operation, all channels that meet the criteria are filtered out from K channels.

[0179] The number of filtered channels is counted, and the count is output as an estimate of the number of effective sound sources.

[0180] Count all selected channels and determine their total number. The integer obtained from this count is the estimated number of valid sound sources in the module's final output. This value directly answers the key question of "how many meaningful sound sources exist in the scene," achieving an implicit and adaptive estimation of the number of sound sources.

[0181] The second objective of this invention is to provide a sound source separation method based on orthogonalized source components and spatial entropy constraints. The sound source separation system based on orthogonalized source components and spatial entropy constraints, as described in any one of the above embodiments, includes the following steps:

[0182] S1. Acquire acoustic measurement data, perform feature decomposition on the acoustic measurement data, and generate orthogonal virtual source components;

[0183] S2. Receive the orthogonalized virtual source components, perform nonlinear mapping through a deep neural network, and output a predicted spatial energy distribution map of the sound source.

[0184] S3. Calculate the spatial entropy loss based on the predicted sound source spatial energy distribution map, perform arrangement matching between the predicted sound source spatial energy distribution map and the preset real source target amplitude tensor and calculate the reconstruction error loss, and obtain the total loss by weighted summation of the spatial entropy loss and the reconstruction error loss. Use the total loss to update the network parameters of the deep neural network to obtain the trained deep neural network and output the updated predicted sound source spatial energy distribution map.

[0185] S4. Receive the predicted sound source spatial energy distribution map output by the trained deep neural network and the preset real source target amplitude tensor, perform arrangement matching, and generate the sound source estimation results with the order aligned.

[0186] S5. Perform channel energy statistics on the predicted sound source spatial energy distribution map output by the trained deep neural network to determine the number of effective sound sources. The above describes and illustrates the basic principles, main features, and advantages of this invention. Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made without departing from the spirit and scope of the invention, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A control system for a sound source separation system based on orthogonalized source components and spatial entropy constraints, characterized in that: It includes an orthogonal source generation module, a deep post-processing module, a prior constraint construction module, a source permutation resolution module, and a source quantity estimation module; The orthogonalization source generation module is used to collect acoustic measurement data, perform feature decomposition on the acoustic measurement data, and generate orthogonalization virtual source components. The deep post-processing module is used to receive orthogonalized virtual source components, perform nonlinear mapping through a deep neural network, and output a predicted spatial energy distribution map of the sound source. The prior constraint construction module is used to calculate the spatial entropy loss based on the predicted sound source spatial energy distribution map, perform arrangement matching between the predicted sound source spatial energy distribution map and the preset real source target amplitude tensor and calculate the reconstruction error loss, and obtain the total loss by weighted summation of the spatial entropy loss and the reconstruction error loss. The total loss is used to update the network parameters of the deep neural network to obtain the trained deep neural network and output the updated predicted sound source spatial energy distribution map. The source arrangement resolution module is used to receive the predicted sound source spatial energy distribution map output by the trained deep neural network and the preset real source target amplitude tensor, perform arrangement matching, and generate the sound source estimation result with the order aligned. The source quantity estimation module is used to perform channel energy statistics on the predicted sound source spatial energy distribution map output by the trained deep neural network to determine the number of effective sound sources.

2. The sound source separation system based on orthogonalized source components and spatial entropy constraints according to claim 1, characterized in that: The orthogonalization source generation module is used to collect acoustic measurement data, perform feature decomposition on the acoustic measurement data, and generate orthogonalized virtual source components, including: Acquire the sound pressure time-domain signals of each channel of the microphone array, apply a window function to the sound pressure time-domain signals and perform a short-time Fourier transform to obtain the complex sound pressure spectrum matrix at each frequency point; Multiply the complex acoustic compression spectrum matrix by its conjugate transpose and average it over all time frames to obtain the cross-spectral matrix; Perform eigenvalue decomposition on the cross-spectral matrix to obtain an orthogonal matrix composed of eigenvectors and a diagonal matrix composed of eigenvalues; Based on the geometric position of the microphone array and the discrete grid of the source plane, a Green's function matrix is ​​constructed in conjunction with the sound propagation model, and the pseudo-inverse of the Green's function matrix is ​​calculated as the acoustic backpropagation operator. Multiplying the eigenvector matrix by the eigenvalue diagonal matrix yields the weighted eigenvector matrix; The weighted eigenvector matrix is ​​multiplied by the acoustic backpropagation operator and mapped onto the discrete grid of the source plane to generate a spatial distribution map of the orthogonalized virtual source components.

3. The sound source separation system based on orthogonalized source components and spatial entropy constraints according to claim 1, characterized in that: The deep post-processing module receives orthogonalized virtual source components, performs nonlinear mapping through a deep neural network, and outputs a predicted spatial energy distribution map of the sound source, including: Identify the maximum value in the amplitude tensor of the orthogonalized virtual source component of the input, divide the amplitude tensor by the maximum value, and generate the normalized input tensor; The normalized input tensor is input into the Transformer-enhanced U-Net neural network, and convolution and spatial downsampling are performed on each source channel of the normalized input tensor to generate encoder feature maps. At each jump connection in the encoder-decoder path, the encoder feature map of the current level is input into the Transformer module. The correlation weights between positions in the sequence are calculated through a self-attention mechanism based on scaling dot products, and then weighted summation is performed. The enhanced feature sequence is generated through feedforward neural network processing, resulting in a feature map processed by Transformer. At each level of the decoder path, the feature map of the previous level is upsampled, and the upsampled result is concatenated with the feature map of the corresponding level processed by Transformer in the channel dimension to generate a fused feature map. A two-dimensional convolution operation is performed on the fused feature map output from the last layer of the decoder path to map the number of channels to the preset number of output channels K, generating a normalized preliminary predicted spatial energy distribution map of the sound source. The normalized preliminary predicted spatial energy distribution map of the sound source is multiplied by the maximum value of the corresponding input case stored, and then rescaled to generate the predicted spatial energy distribution map of the sound source.

4. The sound source separation system based on orthogonalized source components and spatial entropy constraints according to claim 1, characterized in that: The prior constraint construction module is used to calculate the spatial entropy loss based on the predicted sound source spatial energy distribution map, perform permutation matching between the predicted sound source spatial energy distribution map and the preset real source target amplitude tensor, calculate the reconstruction error loss, weight and sum the spatial entropy loss and the reconstruction error loss to obtain the total loss, update the network parameters of the deep neural network using the total loss, obtain the trained deep neural network, and output the updated predicted sound source spatial energy distribution map, including: The pixel energy value of each output source channel is obtained by squaring the amplitude of each output source channel in the predicted sound source spatial energy distribution map. Perform arithmetic addition on the energy values ​​of all pixels in each output source channel to obtain the total energy of each output source channel; Divide the energy value of each pixel in each output source channel by the total energy of that channel to obtain the normalized spatial energy distribution of each output source channel. Multiply the normalized spatial energy distribution of each output source channel by its natural logarithm to obtain the entropy component of each channel. Perform arithmetic addition on all entropy components of each output source channel and take the inverse to obtain the spatial entropy value of each output source channel; Perform arithmetic addition on the spatial entropy values ​​of all output source channels to obtain the spatial entropy loss term; Based on the predicted spatial energy distribution map of the sound source and the amplitude tensor of the real source target, the Hungarian matching algorithm is applied to find the permutation matrix that minimizes the L1 norm distance between the two. The permutation matrix is ​​used to rearrange the predicted spatial energy distribution map of the sound sources; The L1 norm distance between the rearranged predicted sound source spatial energy distribution map and the pre-acquired real source target amplitude tensor is calculated to obtain the Hungarian L1 loss term. The total loss function is obtained by performing arithmetic addition on the results of multiplying the Hungarian L1 loss term and the spatial entropy loss term by the preset regularization weight coefficient. The gradient is calculated using the total loss function, and the network parameters of the deep neural network in the deep post-processing module are updated. After training, the predicted spatial energy distribution map of the sound source generated by the trained deep neural network is output.

5. The sound source separation system based on orthogonalized source components and spatial entropy constraints according to claim 4, characterized in that: The preset regularization weight coefficients are obtained through the following steps: Multiple network models are initialized and trained using different candidate regularized weight coefficients from the set {0,1,2,5,8,10,20,30,40,50,100}. On the experimental validation dataset with coherent source configuration, the median spatial correlation of the output results of each trained network model is calculated; The candidate regularization weight coefficients that maximize the median spatial correlation are selected as the preset regularization weight coefficients.

6. The sound source separation system based on orthogonalized source components and spatial entropy constraints according to claim 1, characterized in that: The source arrangement resolution module receives the predicted sound source spatial energy distribution map output by the trained deep neural network and the preset real source target amplitude tensor, performs arrangement matching, and generates sequentially aligned sound source estimation results, including: A channel distribution from the predicted sound source spatial energy distribution map is selected sequentially and paired with a channel distribution from the pre-acquired real source target amplitude tensor to form a calculation pair; For each calculation pair, perform element-by-element subtraction on the two sets of distribution data to obtain the difference distribution data; Perform the absolute value operation on all elements in each difference distribution data to obtain the absolute difference distribution data; Sum all elements in each absolute difference distribution data point to obtain a scalar distance value; Arrange all the calculated scalar distance values ​​in the order of the selected predicted sound source channels as rows and the order of the selected real source target channels as columns to generate the first distance matrix; Input the first distance matrix into the Hungarian matching algorithm, perform the optimal allocation solution, and generate the second allocation matrix; Based on the second allocation matrix, a permutation operation is performed on all channels of the predicted sound source spatial energy distribution map to generate sequentially aligned sound source estimation results.

7. The sound source separation system based on orthogonalized source components and spatial entropy constraints according to claim 6, characterized in that: For each calculation pair, perform element-wise subtraction on the two sets of distribution data to obtain the difference distribution data, including: Extract the two-dimensional distribution data of the currently selected channel from the predicted spatial energy distribution map of the sound source; Extract the two-dimensional distribution data of the currently selected channel from the amplitude tensor of the real source target; The values ​​at the same spatial pixel positions of the two extracted two-dimensional distribution data are subtracted to obtain difference distribution data with the same dimension.

8. The sound source separation system based on orthogonalized source components and spatial entropy constraints according to claim 1, characterized in that: The source quantity estimation module is used to perform channel energy statistics on the predicted sound source spatial energy distribution map output by the trained deep neural network to determine the number of effective sound sources, including: The sum of squared amplitudes of all pixels in each channel of the predicted sound source spatial energy distribution map is calculated to obtain the total energy value of each channel; Based on the total energy value of all channels, the maximum value is identified as the energy reference benchmark; The relative energy ratio of each channel is obtained by comparing the total energy value of each channel with the energy reference benchmark. The relative energy ratio is compared with a preset energy threshold to filter out channels whose relative energy ratio is greater than the threshold. The number of filtered channels is counted, and the count is output as an estimate of the number of effective sound sources.

9. A sound source separation method based on orthogonalized source components and spatial entropy constraints, based on the sound source separation system based on orthogonalized source components and spatial entropy constraints as described in any one of claims 1-8, characterized in that: Includes the following steps: S1. Acquire acoustic measurement data, perform feature decomposition on the acoustic measurement data, and generate orthogonal virtual source components; S2. Receive the orthogonalized virtual source components, perform nonlinear mapping through a deep neural network, and output a predicted spatial energy distribution map of the sound source. S3. Calculate the spatial entropy loss based on the predicted sound source spatial energy distribution map, perform arrangement matching between the predicted sound source spatial energy distribution map and the preset real source target amplitude tensor and calculate the reconstruction error loss, and obtain the total loss by weighted summation of the spatial entropy loss and the reconstruction error loss. Use the total loss to update the network parameters of the deep neural network to obtain the trained deep neural network and output the updated predicted sound source spatial energy distribution map. S4. Receive the predicted sound source spatial energy distribution map output by the trained deep neural network and the preset real source target amplitude tensor, perform arrangement matching, and generate the sound source estimation results with the order aligned. S5. Perform channel energy statistics on the predicted sound source spatial energy distribution map output by the trained deep neural network to determine the number of effective sound sources.

Images