Audio denoising method based on lightweight robust unsupervised feature selection of algorithm

Abstract

The application discloses an audio denoising method based on algorithm lightweight robust unsupervised feature selection, comprising collecting noisy audio signals in real-time scenes as audio data to be denoised; based on the audio data to be denoised, a robust unsupervised feature selection model based on joint subspace learning is constructed, and a target optimization function is constructed through multi-task complementary learning; an alternating optimization method is used to iteratively solve the target optimization function step by step; based on algorithm expansion technology, the iterative solution process is converted into a modular neural network structure to construct a lightweight unsupervised feature selection network; based on the constructed lightweight unsupervised feature selection network, the audio data to be denoised is iteratively calculated to output an optimal feature weight matrix; based on the optimal feature weight matrix, the importance of features is evaluated, and a target feature subset corresponding to key audio features is selected; and the final audio data after feature selection is output. The application uses the above method to realize efficient denoising and quality improvement of audio data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of audio denoising, in particular to an audio denoising method based on algorithmic lightweight robust unsupervised feature selection. BACKGROUND

[0002] Audio denoising refers to a technology of suppressing or removing noise components from a noise-polluted audio signal through signal processing or machine learning methods, generating clearer and more usable audio while preserving the integrity and perceptual quality of the original target audio information as much as possible. In application scenarios such as real-time conference systems and simultaneous interpretation, audio denoising is very important, and a real-time, low-latency and high-quality audio denoising method is the key to improving user experience and downstream task performance.

[0003] Audio quality can be affected by many environmental noises, which can be divided into three categories. The first category is stationary noise (fan noise, engine noise, etc.); the second category is non-stationary noise (human voice interference, traffic noise, etc.); and the third category is extreme noise (microphone impact sound, sudden burst sound, etc.).

[0004] In order to remove the influence of the above different types of noise on audio, the current main method is to use audio denoising methods based on statistical characteristics (such as spectral subtraction, Wiener filtering, MMSE estimation, etc.).

[0005] These methods can play a certain denoising role in specific scenarios. However, the existing technology has obvious shortcomings:

[0006] The audio denoising method based on statistical characteristics cannot accurately estimate the noise in complex or non-stationary conference scenarios, because the characteristics of the noise no longer meet the preset stationary or specific distribution assumption, which leads to distortion of the speech signal and a significant decrease in the intelligibility and naturalness of the audio. The audio is processed in long frames, and the processing process is time-consuming, which makes it difficult to meet the real-time requirements of audio denoising in real-time scenarios such as conference systems.

[0007] The mainstream unsupervised feature selection method UFSHS directly constructs a fixed similarity graph in the original high-dimensional data space, is greatly affected by outliers, cannot effectively handle non-stationary and extreme noise, and has complex iterative process calculations, making it difficult to adapt to dynamic real-time scenarios. SUMMARY

[0008] The purpose of the present application is to provide an audio denoising method based on algorithmic lightweight robust unsupervised feature selection, which solves the problems of poor noise robustness, complex calculation and insufficient real-time performance of existing methods: through joint subspace learning and multi-task complementary learning, the modeling ability of the model for global and local features of the data is improved; combined with sparse learning to enhance noise adaptability; using algorithm expansion technology to simplify complex calculations and speed up model operation, realizing efficient audio denoising and quality improvement.

[0009] To achieve the above objectives, this invention provides a lightweight, robust, unsupervised feature selection-based audio denoising method, comprising the following steps:

[0010] S1. Collect noisy audio signals in real-time scenes as noise data to be denoised;

[0011] S2. Based on the noise-to-be-denoised frequency data in S1, construct a robust unsupervised feature selection model based on joint subspace learning, and construct the objective optimization function through multi-task complementary learning.

[0012] S3. Use the alternating optimization method to solve the objective optimization function step by step iteratively;

[0013] S4. Based on algorithm expansion technology, the iterative solution process of S3 is transformed into a modular neural network structure to construct a lightweight unsupervised feature selection network.

[0014] S5. A lightweight unsupervised feature selection network built on S4 is used as input. After iterative calculation, the network outputs the optimal feature weight matrix. The importance of features is evaluated based on the optimal feature weight matrix, and a target feature subset is selected. The audio data is then denoised based on the target feature subset, and the audio data after feature selection is output, thus completing the audio denoising.

[0015] Preferably, S2 is as follows:

[0016] S21. The robust unsupervised feature selection model based on joint subspace learning includes an adaptive graph manifold learning module, a clustering module, a reconstruction module, and a sparse learning module.

[0017] S22. Construct the objective optimization function through multi-task complementary learning, as shown below:

[0018] ;

[0019] ;

[0020] in, It is a graph manifold matrix. It is the number of time steps of the frequency to be denoised; It is a cluster center matrix. It is a low-dimensional feature dimension. It is the number of cluster categories; It is a clustering representation matrix; It is a noise matrix. It is the frequency feature dimension to be denoised; It is the feature weight matrix; It is a unit vector; It is the identity matrix; It is a nonlinear canonical parameter; These are the regularization parameters of the clustering model; These are encoder regularization parameters; It is a sparse regularization parameter; It is the noise sparsity regularization parameter; These are auxiliary variables used to ensure the clustering representation matrix Satisfies the nonnegativity constraint; It is a vector Norm; It is a matrix Norm; It is a sparse norm; It is the noise-reducing frequency data matrix; , They are the first , The noise-reducing frequency feature vectors at each time step; , Represents the first element in the total length of the noise-reducing frequency signal. , One time step; It is a graph manifold matrix The Line number Column elements; It is the feature weight matrix The transpose of .

[0021] Preferably, in S3, the objective optimization function is solved by iteratively updating the graph manifold matrix, cluster center matrix, cluster representation matrix, feature weight matrix, and noise matrix step by step. The iterative process is as follows:

[0022] S31. Update the graph manifold matrix: ;

[0023] S32. Update the cluster center matrix: ;

[0024] S33. Update the clustering representation matrix:

[0025] Define auxiliary variables ;in, It is an auxiliary variable matrix The Line number List the elements; It is a clustering representation matrix The Line number List the elements;

[0026] Calculate the input matrix: ;in, It is a constant;

[0027] right Perform singular value decomposition: ;in, It is a left singular matrix; It is a singular value matrix; It is a right singular matrix;

[0028] Output ;

[0029] S34. Update the feature weight matrix:

[0030] ;

[0031] in, It is the feature weight matrix The Line number List the elements; It is about finding the trace of a matrix; It is a diagonal matrix, diagonal elements ;in, It is the feature weight matrix The Row vectors; It is a graph manifold matrix The graph Laplacian matrix is ​​calculated as follows: ; , It is a degree matrix The One diagonal element;

[0032] S35, Update the noise matrix:

[0033] Calculate outlier detection parameters: ;

[0034] in, It is the Laplace nucleus; It is the variance of the features in a single time step; It is the interquartile range within a single time step; It is the difference between the reconstructed audio data and the audio data to be denoised; Representing the Residual audio data at each time step;

[0035] Calculate the Laplace fraction: ;

[0036] in, It is a sample The Laplace fraction, when At that time, representing Contains outliers, and indexes Save to collection middle; It is the threshold for identifying outliers;

[0037] Iteratively update the outlier threshold: ;

[0038] in, It is the number of iterations; It is the noise threshold ultimately used to update the noise matrix; It is the first t The noise threshold for the next iteration; This is the noise threshold currently calculated based on the Laplace kernel function, specifically expressed as:

[0039] ;

[0040] Update the noise matrix using a soft thresholding function:

[0041] ;

[0042] in, It is a noise matrix The Noise data at each time step; It is a hard threshold function.

[0043] Preferably, S4 is as follows:

[0044] S41. Design the core lightweight component, the Kronecker linear layer; decompose the large weight matrix of the original linear layer into a weighted combination of Kronecker products of multiple low-rank submatrices, as shown below:

[0045] ;

[0046] in, , It is a trainable parameter matrix. , It is any positive integer; It is the Kronecker product; It is the Kronecker product, specifically expressed as: ;

[0047] S42, Lightweight Unsupervised Feature Selection Network It consists of several modules, among which... The integer is used; the module contains 7 core sub-modules, namely clustering module, weight score module, feature representation module, high-order matrix representation module, graph representation module, soft threshold selection module and weight update module.

[0048] Preferably, the clustering module in S42 is used to process the clustering information calculation during the update process of the cluster center matrix, cluster representation matrix and feature weight matrix;

[0049] First, to address the complexity of calculating the inverse matrix, an inverse matrix network is constructed.

[0050] The matrix inversion is transformed into an optimization problem; the above optimization problem is solved based on the momentum gradient descent method, and the momentum parameters and step size parameters of the momentum gradient descent are transformed into a learnable form; a subnetwork of 4 Kronecker linear layers is used to obtain the inverse matrix of the matrix, wherein the Kronecker linear layers are used to reduce the number of fully connected parameters in the learnable form;

[0051] Secondly, to address the complexity issue of singular value decomposition, an orthogonal approximation network is constructed.

[0052] The singular value decomposition is transformed into an orthogonal constraint optimization problem; the orthogonal matrix of the singular value decomposition is approximately output through a self-attention network and trainable parameters.

[0053] Finally, two Kronecker linear layers are introduced to learn the feature representations of the audio data matrix, cluster center matrix, and cluster representation matrix, respectively, and the product is output. ;in, It is a clustering module calculation The approximate result.

[0054] Preferably, the weight score module, feature representation module, and higher-order matrix representation module in S42 are as follows:

[0055] The weight score module is used to calculate the scaling factor during the feature weight matrix update process. It is implemented through a cascaded structure of Kronecker linear layers + activation functions + Kronecker linear layers + Softplus layers to learn the internal structure information of the feature weight matrix and output the result. ;in, It is a scaling factor Approximate results;

[0056] The feature representation module simplifies large-scale matrix calculations while preserving high-order information of the data, generating interactive information that characterizes the internal dynamics of the audio data, corresponding to the feature weight matrix update process. The calculation is as follows:

[0057] The audio data is learned in the feature space through two Kronecker linear layers. Different representations in;

[0058] After multiplying the output matrices of the two Kronecker linear layers, the output is... ;in, It is a feature representation Approximate results;

[0059] The higher-order matrix representation module is used to generate higher-order matrices that represent the interaction between weight information and data information, corresponding to the feature weight matrix update process. The calculation is as follows:

[0060] By introducing two Kronecker linear layers, the feature information of the feature weight matrix and the audio data matrix are learned respectively;

[0061] Then obtain the interaction results of feature information and output them. ;in, It is a product of higher-order matrices. The approximate result.

[0062] Preferably, the soft threshold selection module in S42 is used to update the outlier matrix, as detailed below:

[0063] By introducing a mask To achieve smooth selection;

[0064] Two Kronecker linear layers are used to obtain interaction information and output... ;in, It is calculated by the soft threshold selection module. The approximate result.

[0065] Preferably, the graph representation module in S42 is used to characterize the collaborative changes of features of the original data under the low-dimensional representation graph, and its specific content is as follows:

[0066] First, represent the low-dimensional dimension. Converted into a three-dimensional tensor, i.e. , ;in, and element values ​​and Consistent;

[0067] Next, calculate the graph manifold matrix: ;

[0068] in, For feature dimension Summation, Normalize the rows;

[0069] Finally, two Kronecker linear layers are introduced to learn the directional differences of the original features under graph constraints and the basic orientation of the original features. An outer product is then calculated to approximate the covariant behavior of the data under graph constraints, and the output is... ;in, It is an approximate calculation result of the covariant behavior of data under graph constraints.

[0070] Preferably, the weight update module in S42 is used to update the feature weight matrix. Normalization, the specific details are as follows:

[0071] Define a diagonal matrix The The diagonal elements are the feature weight matrix. The reciprocal of;

[0072] Diagonal matrix Multiplying with the feature weight matrix achieves the effect of multiplying the feature weight matrix. -norm Normalization.

[0073] Preferably, the feature-selected audio data output by S5 is as follows:

[0074] ;

[0075] in, It outputs the audio data after feature selection; It is the number of audio features selected in unsupervised feature selection; It is a feature selection function that evaluates the importance of features in the noise-reducing frequency data matrix based on the feature weight matrix, and selects the features before... Key audio features; It is the optimal feature weight matrix, specifically represented as: ;in, This is the solution function for the robust unsupervised feature selection audio denoising method, calculated by... The optimal feature weight matrix can be obtained.

[0076] Therefore, the present invention employs the aforementioned algorithm-based lightweight robust unsupervised feature selection audio denoising method, which has the following beneficial effects:

[0077] (1) Improved denoising accuracy and noise robustness: This invention constructs a graph structure in a low-dimensional space through joint subspace learning, and combines reconstruction constraints and clustering constraints to obtain a more accurate feature subset, thereby reducing speech signal distortion; Outlier modeling and By using sparse norms, non-stationary and extreme noise can be effectively handled, and stationary noise weights can be suppressed, allowing the model to focus more on the main speech components and become more robust to noise.

[0078] (2) Meets the high-efficiency computing requirements of real-time scenarios: Based on the algorithm expansion technology, this invention transforms complex iterative calculations and matrix solving into neural networks, converts regular parameters into trainable parameters, and avoids complex parameter tuning; adopts a lightweight architecture, separates the physical implementation of audio denoising from the real-time requirements, simplifies the calculation process and improves the calculation efficiency while maintaining interpretability, and adapts to dynamic scenarios such as real-time meetings and simultaneous interpretation.

[0079] (3) Enhance the practicality and scene adaptability of the model: This invention integrates multiple modules such as codec and adaptive graph manifold to carry out multi-task complementary learning, taking into account the data reconstruction quality, local geometric structure and global distribution pattern; avoids high-dimensional computation overflow, improves the utilization rate of feature information, and does not require assumption of noise distribution in unsupervised scenarios, which can adapt to the audio denoising needs in complex noise environments.

[0080] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0081] Figure 1 This is a flowchart of the lightweight, robust, unsupervised feature selection-based audio denoising method of the present invention.

[0082] Figure 2 This is a structural diagram of the lightweight unsupervised feature selection network in an embodiment of the present invention. Detailed Implementation

[0083] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0084] This invention presents a lightweight, robust, unsupervised feature selection-based audio denoising method, comprising the following steps:

[0085] S1. Collect noisy audio signals in real-time scenes as noise data to be denoised;

[0086] S2. Based on the noise-to-be-denoised frequency data in S1, construct a robust unsupervised feature selection model based on joint subspace learning, and construct the objective optimization function through multi-task complementary learning.

[0087] S3. Use the alternating optimization method to solve the objective optimization function step by step iteratively;

[0088] S4. Based on algorithm expansion technology, the iterative solution process of S3 is transformed into a modular neural network structure to construct a lightweight unsupervised feature selection network corresponding to the robust unsupervised feature selection model.

[0089] S5. A lightweight unsupervised feature selection network built on S4 is used as input. After iterative calculation, the network outputs the optimal feature weight matrix. The importance of features is evaluated based on the optimal feature weight matrix, and a target feature subset is selected. The audio data is then denoised based on the target feature subset, and the audio data after feature selection is output, thus completing the audio denoising.

[0090] Example

[0091] like Figure 1 As shown, the present invention provides a lightweight, robust, unsupervised feature selection-based audio denoising method, comprising the following steps:

[0092] S1. Collect noisy audio signals in real-time scenes as the noise-to-denoise frequency data, as shown below:

[0093] ;

[0094] in, This represents the number of frequency features to be denoised; This represents the number of time steps in the frequency to be denoised; This represents the frequency data to be denoised; Representing the Noise-reducing frequency data at each time step.

[0095] To improve the quality of feature subsets in unsupervised feature selection models, this embodiment employs an unsupervised feature selection method to simultaneously learn the global distribution information and local manifold information of the data, and utilizes sparse learning to enhance the model's robustness to outliers such as noise, thereby selecting the optimal feature subset.

[0096] S2. Based on the noise-removing frequency data in S1, a robust unsupervised feature selection model based on joint subspace learning is constructed, and the objective function is constructed through multi-task complementary learning.

[0097] S21. The robust unsupervised feature selection model based on joint subspace learning includes an adaptive graph manifold learning module, a clustering module, a reconstruction module, and a sparse learning module.

[0098] The adaptive graph manifold module maps the noise-to-frequency data to a low-dimensional space, reducing the computational complexity of model graph construction, removing some noise effects, and avoiding trivial solutions by introducing nonlinear terms.

[0099] The clustering module ensures that low-dimensional representations maintain the global distribution pattern of data in low-dimensional space, guaranteeing that different types of audio data can be separated and avoiding audio feature confusion.

[0100] The reconstruction module ensures the quality constraints of the low-dimensional representation, guaranteeing that the low-dimensional representation can effectively represent the basic characteristics of the original data without distortion.

[0101] The sparse learning module considers the effects of different types of noise, and utilizes [specific methods] to address extreme and non-stationary noise. and Constraints enhance the model's ability to model complex noise without requiring initial assumptions.

[0102] In this way, through multi-task complementary learning, the model can ensure intra-class consistency of the main frequency data, while achieving the separability of noise and main frequency data.

[0103] S22. Construct the objective optimization function through multi-task complementary learning, as shown below:

[0104] ;

[0105] ;

[0106] in, It is a graph manifold matrix that represents the correlation information between data time steps; It is a cluster center matrix. m is the low-dimensional feature dimension, and c is the number of cluster categories; It is a clustering representation matrix; This is a noise matrix, representing the extreme noise in the original data; It is the feature weight matrix; It is a unit vector, and all its elements have a value of 1. It is an identity matrix, with its diagonal elements being 1 and the rest being 0; It is a nonlinear canonical parameter; These are the regularization parameters of the clustering model; These are the regularization parameters for robust encoders; It is the row sparsity regularization parameter of the weight matrix; It is the noise sparsity regularization parameter; These are auxiliary variables used to ensure the clustering representation matrix Satisfies the nonnegativity constraint; Represents vector Norm; Represents the matrix Norm; Represents the row sparsity norm of the matrix; 、 They are the first 、 The noise-reducing frequency feature vectors at each time step; , Represents the first of the total length of the audio signal 、 One time step; It is a graph manifold matrix The Line number Column elements; It is the feature weight matrix The transpose of .

[0107] S3. The objective function is solved step by step iteratively using an alternating optimization method.

[0108] Because the objective function contains orthogonal and sparse constraints, it is non-convex, making direct solution computationally complex. Therefore, this embodiment uses an alternating optimization method to iteratively update the graph manifold matrix, cluster center matrix, cluster representation matrix, feature weight matrix, and noise matrix step by step. The specific update process is as follows:

[0109] S31, Update the graph manifold matrix : .

[0110] S32, Update the cluster center matrix : .

[0111] S33, Update the clustering representation matrix :

[0112] Define auxiliary variables ;in, Auxiliary variable The Line number The element at the column position; It is a clustering representation matrix The Line number The element at the column position.

[0113] Calculate the input matrix : ;in, It is a constant greater than 0, and is generally set to 0 to ensure the effectiveness of the solution. .

[0114] right Perform singular value decomposition: ;

[0115] in, It is a left singular matrix; It is a singular value matrix; It is a right singular matrix.

[0116] Output .

[0117] S34. Update the feature weight matrix:

[0118] ;

[0119] in, This represents finding the trace of a matrix; It is a diagonal matrix, with diagonal elements , It is a weight matrix The Row vectors; It is a graph manifold matrix The graph Laplacian matrix is ​​calculated as follows: ; , It is a degree matrix The One diagonal element, and the rest of the elements have a value of 0; It is a weight matrix The Line number The value of the element at the column position.

[0120] S35, Update the noise matrix:

[0121] Calculate outlier detection parameters: ;

[0122] in, It is the Laplace nucleus; It is the variance of the features in a single time step; It is the interquartile range within a single time step; It is the difference between the reconstructed audio data and the audio data to be denoised; Representing the Each time step residual audio data.

[0123] Calculate the Laplace fraction: ;

[0124] in, It is a sample The Laplace fraction, when At that time, representing Contains outliers, and indexes Save to collection middle; It is the threshold for identifying outliers.

[0125] Iteratively update the outlier threshold: ;

[0126] in, It is the number of iterations; It is the noise threshold ultimately used to update the noise matrix; It is the first The noise threshold for the next iteration; This is the noise threshold for the current Laplace kernel function solution, specifically expressed as:

[0127] .

[0128] Update the noise matrix using a soft thresholding function:

[0129] ;

[0130] in, It is a noise matrix The j Noise data at each time step; It is a hard threshold function.

[0131] S4. Based on algorithmic expansion technology, the iterative solution process of S3 is transformed into a modular neural network structure to construct a lightweight unsupervised feature selection network.

[0132] The iterative algorithm of the robust unsupervised feature selection model is transformed into a lightweight neural network using deep unrolling technology, enabling adaptive parameter tuning. This allows for fast audio denoising in demanding real-time scenarios such as headsets and conference systems, without requiring additional hardware resources like GPUs. This embodiment eliminates the complex parameter tuning process and simplifies the original, complex computational process through a lightweight unsupervised feature selection network.

[0133] S41. Design the core lightweight component, the Kronecker Linear Layer (KLL).

[0134] The Kronecker Linear Layer (KLL) approximates the large weight matrix in the original linear layer using the Kronecker product.

[0135] Traditional linear layers suffer from a large number of parameters, which contradicts the design of lightweight unsupervised feature selection networks. This embodiment uses the Kronecker product as an optimization technique. In large matrix-vector computations, by decomposing the large matrix into a weighted combination of multiple low-rank submatrices, the number of parameters is significantly reduced while preserving the original high-order feature interaction capabilities.

[0136] The general calculation expression for a linear layer is as follows:

[0137] ;

[0138] in, It is the input data for the linear layer; This is the weight matrix of the linear layer, and it consists of the sum of Kronecker products, specifically expressed as: ;in, It is the Kronecker product; It is the Kronecker product, specifically expressed as: ;in, , It is a trainable parameter matrix. , It is any positive integer.

[0139] Therefore, the calculation expression for KLL can be obtained as follows:

[0140] .

[0141] S42, Lightweight Unsupervised Feature Selection Network It consists of several modules, among which... Each module contains 7 core sub-modules: clustering module, weight score module, feature representation module, high-order matrix representation module, graph representation module, soft threshold selection module, and weight update module.

[0142] Specifically, except In addition, all sub-modules represent feature weight matrices. The information computation parts at different levels in the update expression are converted into trainable parameters, reducing high-order matrix calculations and parameter tuning. While preserving the interpretability of the method, the model's real-time performance is significantly improved, achieving fast and accurate audio denoising performance in complex dynamic scenes.

[0143] (1) The clustering module CM, which includes an inverse matrix network and an orthogonal approximation network, is used to process cluster information calculations during the updating of the cluster center matrix, cluster representation matrix, and feature weight matrix. The specific structure of CM is as follows: Figure 2 As shown in the CM module.

[0144] The updating of the cluster center matrix and cluster representation matrix involves two computationally complex processes: matrix inversion and singular value decomposition, which limits the speed of the model when calculating large matrices. Therefore, this invention simplifies the computation by approximating these two computations as a new optimization problem and transforming them into an inverse matrix network and an orthogonal approximation network.

[0145] First, to address the complex problem of calculating the inverse matrix, we construct the inverse matrix network INet, as detailed below:

[0146] The first step is to transform the problem of finding the inverse matrix into an optimization problem, as shown below:

[0147] ;

[0148] in, The input is the target matrix to be inverted; It is the inverse matrix corresponding to the target matrix.

[0149] The second step is to solve the above optimization problem using the momentum gradient descent method, as shown below:

[0150] ;

[0151] in, It is an optimization problem about finding the inverse matrix. The gradient;

[0152] ;

[0153] in, It is the momentum term; It is a momentum parameter;

[0154] ;

[0155] in, It is the step size parameter.

[0156] The third step involves converting the momentum and step size parameters into a learnable form consisting of a fully connected layer, an activation function, and another fully connected layer. Since the fully connected layer has a large number of parameters, KLL is used instead of the fully connected layer in the learnable form, resulting in a momentum gradient descent parameter network. By stacking the four parameterized sub-networks, the inverse matrix network INet is obtained. Figure 2 As shown in the INet and MGDPNet modules, It is the first The input to the INet corresponding to each block is an identity matrix. This is the inverse matrix output by INet.

[0157] Secondly, to address the complexity issue of singular value decomposition, an orthogonal approximation network, ONet, is constructed.

[0158] The first step is to transform the singular value decomposition into an orthogonal approximation problem, as shown below:

[0159] ;

[0160] in, It is a matrix And through Replacement matrix F Calculation; It is the target matrix that satisfies the orthogonal constraint condition.

[0161] The second step is to use the augmented Lagrange multiplier method of PLAM to solve the orthogonal approximation problem.

[0162] make The solution process is as follows:

[0163] ;

[0164] ;

[0165] ;

[0166] ;

[0167] ;

[0168] in, It is the augmented Lagrangian function of the objective function; For Lagrange multipliers; For regularization parameters; This is the step size parameter; It is the augmented Lagrangian function with respect to the matrix. The gradient; It is the first The inverse matrix output by each inverse matrix module; It is the input matrix. .

[0169] The third step is to transfer the contents of the second step. and The computation is replaced by a self-attention network; and The parameters are converted into trainable parameters; by stacking the parameterized update network three times, the orthogonal approximation network ONet is obtained. A single layer representation of ONet is as follows: Figure 2 As shown in ONet(1), where This represents a randomly initialized input matrix. This represents the approximate orthogonal matrix of the ONet(1) output.

[0170] Then, the contribution information of the clustering model to the update of the feature weight matrix is ​​combined. Two KLLs are introduced to learn the feature representation and cluster reconstruction of the original data. The feature representations are then calculated, and the product of the two feature representations is used as the final output. ,in, It is a clustering module calculation The approximate result.

[0171] (2) The Weight Score Module (WSM) is used to calculate the scaling factor during the feature weight matrix update process. Its specific structure is as follows:Figure 2 The WSM module is shown in the image.

[0172] Because the calculation of the scaling factor involves negative exponents and square roots, the network solution becomes unstable. Therefore, this embodiment implements a learnable nonlinear transformation scaling factor structure. This nonlinear transformation scaling factor structure is constructed by cascading KLL+ activation functions+KLL+Softplus to learn the internal structure information of the feature weight matrix, where the scaling factor is as follows: ;in, It is a scaling factor The approximate result.

[0173] (3) Feature Representation Module (FRM): This module simplifies large-scale matrix calculations and preserves high-order information of the data, generating interactive information that characterizes the internal dynamics of the audio data. It corresponds to the feature weight matrix update process. The calculation.

[0174] This module learns the audio data in the feature space using two KLLs. Different representations in the code are then multiplied together to output the result. ,in, It is a feature representation The approximate result is as follows. The calculation process is as follows: Figure 2 The FRM module is shown in the image.

[0175] (4) Higher-order matrix representation module (HRM) is used to generate a higher-order matrix representing the interaction between weight information and data information, corresponding to the higher-order matrix multiplication in the feature weight matrix update process. The calculation part.

[0176] Directly calculating higher-order matrices is computationally complex, increasing the time consumed by model computation. Therefore, two KLLs are introduced to learn the feature weight matrices respectively. and data matrix The feature information is obtained; then the interaction result of the two feature information is obtained, and the output is... ;in, It is a product of higher-order matrices. The approximate result.

[0177] The higher-order matrix representation module reduces computational complexity and improves model efficiency by decreasing the number of multiplications. The specific structure is as follows: Figure 2 As shown in the HRM module.

[0178] (5) Soft threshold selection module SSM, used to update the outlier matrix, corresponding to the contribution part of the outlier matrix in the feature weight matrix update process. The specific structure is as follows Figure 2The SSM module is shown in the image.

[0179] Traditional outlier matrices update variables through hard selection, but hard selection can cause the network to fail at the threshold. The derivative is discontinuous at a certain point, leading to instability in training. To address this, this embodiment introduces a mask. This achieves smooth selection; simultaneously, two KLL methods are used to obtain the final interaction information and output it. ,in, It is calculated by the soft threshold selection module. The approximate result is shown below, where the mask is as follows:

[0180] ;

[0181] in, Trainable parameters, replacing the original hard threshold. .

[0182] Outlier matrix update: ;

[0183] in, It is element-wise multiplication.

[0184] (6) The graph representation module GRM is used to characterize the co-variation of features of the original data under a low-dimensional representation graph, simplifying the process. The calculation process, specifically the structure is as follows: Figure 2 As shown in GRM.

[0185] Traditional element-wise solving has high computational complexity and cannot meet the needs of real-time processing of large-scale data. Therefore, this embodiment achieves efficient computation through tensor transformation, graph manifold simplification, and Kronecker linear layers.

[0186] The first step is to represent the low-dimensional dimension. Converted into a three-dimensional tensor, i.e. , ,in, and Element value and Consistent.

[0187] The second step is to calculate the graph manifold matrix: ;

[0188] in, For feature dimension Summation, Normalize the rows.

[0189] The third step, to simplify the process, introduces two KLLs: learning the directional differences of the original features under graph constraints and the basic orientation of the original features; then, the outer product is calculated to approximate the covariant behavior of the data under graph constraints, and the output is... ,in, It is an approximate calculation result of the covariant behavior of data under graph constraints.

[0190] (7) Weight update module Used to perform scalar operations on the feature weight matrix. Normalization corresponds to the weight update rule. part.

[0191] definition It concerns the feature weight matrix. of The diagonal matrix of the reciprocal; and Perform product operations to achieve the following: of Normalization.

[0192] Through the above transformation, the contribution calculation of different information in the weight update rule is converted into a trainable neural network for updating.

[0193] S5. A lightweight unsupervised feature selection network built upon S4 iteratively calculates the optimal feature weight matrix from the audio data to be denoised. Based on this optimal feature weight matrix, feature importance is evaluated, and a subset of target features corresponding to key audio features is selected. The audio data after feature selection is output, completing the audio denoising process. As shown below:

[0194] ;

[0195] in, It outputs the audio data after feature selection; It is the number of audio features selected in unsupervised feature selection; It is a feature selection function, designed to handle noise-reduced frequency data. according to Before selecting the feature weight score of the root One feature; It is the optimal feature weight matrix. ;in, This is the solution function for the robust unsupervised feature selection audio denoising method, calculated by... The optimal feature weight matrix can be obtained.

[0196] Therefore, the present invention employs the aforementioned algorithm-based lightweight robust unsupervised feature selection audio denoising method, which has the following beneficial effects:

[0197] (1) This invention integrates codecs, adaptive graph manifolds, clustering, and sparse learning methods into a unified framework, constraining the discriminative power of low-dimensional representations from three complementary perspectives: reconstruction quality, local geometry, and global distribution patterns. Through this local-global collaborative optimization approach, the model can simultaneously balance intra-class consistency and inter-class separability, achieving the preservation of high-quality audio features and the removal of noise outliers. Furthermore, based on... The outlier modeling method can effectively learn the distribution of extreme noise and non-stationary noise, ensuring the stability and reliability of feature selection results.

[0198] (2) This invention transforms the regularization parameters in the robust unsupervised feature selection method into trainable parameters through an algorithmic expansion method, thereby avoiding the complex grid search parameter tuning process. Furthermore, this method accelerates model computation through a learnable approach by converting the complex process of finding the inverse matrix and the orthogonal matrix of the target matrix into a neural network. Compared to UFSHS, which uses iterative solutions, the parameter optimization method in this invention eliminates the adjustment process of regularization parameters and simplifies the complex computational patterns during iteration, achieving efficient parameter tuning and model acceleration performance.

[0199] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A lightweight, robust, unsupervised feature selection-based audio denoising method, characterized in that, Includes the following steps: S1. Collect noisy audio signals in real-time scenes as noise data to be denoised; S2. Based on the noise-to-be-denoised frequency data in S1, construct a robust unsupervised feature selection model based on joint subspace learning, and construct the objective optimization function through multi-task complementary learning. S21. The robust unsupervised feature selection model based on joint subspace learning includes an adaptive graph manifold learning module, a clustering module, a reconstruction module, and a sparse learning module. S22. Construct the objective optimization function through multi-task complementary learning, as shown below: ； ； in, It is a graph manifold matrix. It is the number of time steps of the frequency to be denoised; It is a cluster center matrix. It is a low-dimensional feature dimension. It is the number of cluster categories; It is a clustering representation matrix; It is a noise matrix. It is the frequency feature dimension to be denoised; It is the feature weight matrix; It is a unit vector; It is the identity matrix; It is a nonlinear canonical parameter; These are the regularization parameters of the clustering model; These are encoder regularization parameters; It is a sparse regularization parameter; It is the noise sparsity regularization parameter; These are auxiliary variables used to ensure the clustering representation matrix Satisfies the nonnegativity constraint; It is a vector Norm; It is a matrix Norm; It is a sparse norm; It is the noise-reducing frequency data matrix; , They are the first , The noise-reducing frequency feature vectors at each time step; , Represents the first of the total length of the audio signal , One time step; It is the first of the graph manifold matrix. Line 1 Column elements; It is the feature weight matrix The transpose of the matrix; S3. Use the alternating optimization method to solve the objective optimization function step by step iteratively; The objective optimization function is solved by iteratively updating the graph manifold matrix, cluster center matrix, cluster representation matrix, feature weight matrix, and noise matrix step by step. The iterative process is as follows: S31. Update the graph manifold matrix: ; S32. Update the cluster center matrix: ; S33. Update the clustering representation matrix, the details of which are as follows: Define auxiliary variables: ; in, It is an auxiliary variable matrix The Line 1 List the elements; It is a clustering representation matrix The Line 1 List the elements; Calculate the input matrix : ; in, It is a constant; right Perform singular value decomposition: ; in, It is a left singular matrix; It is a singular value matrix; It is a right singular matrix; The conclusion is ; S34. Update the feature weight matrix: ； in, It is the feature weight matrix The Line 1 List the elements; This represents finding the trace of a matrix; It is a diagonal matrix, diagonal elements , It is the feature weight matrix The Row vectors; It is a graph manifold matrix The graph Laplacian matrix is ​​calculated as follows: ; , It is a degree matrix The One diagonal element; S35, Update the noise matrix: Calculate outlier detection parameters: ; in, It is the Laplace nucleus; It is the variance of the features in a single time step; It is the interquartile range within a single time step; It is the difference between the reconstructed audio data and the audio data to be denoised; Representing the Residual audio data at each time step; Calculate the Laplace fraction: ; in, It is a sample The Laplace fraction, when At that time, representing Contains outliers, and indexes Save to collection middle; It is the threshold for identifying outliers; Iteratively update the outlier threshold: ; in, It is the number of iterations; It is the noise threshold ultimately used to update the noise matrix; It is the first t The noise threshold for the next iteration; This is the noise threshold currently calculated based on the Laplace kernel function, specifically expressed as: ； Update the noise matrix using a soft thresholding function: ； in, It is a noise matrix The Noise data at each time step; It is a hard threshold function; S4. Based on algorithm expansion technology, the iterative solution process of S3 is transformed into a modular neural network structure to construct a lightweight unsupervised feature selection network. S41. Design the core lightweight component, the Kronecker linear layer; decompose the large weight matrix of the original linear layer into a weighted combination of Kronecker products of multiple low-rank submatrices, as shown below: ； in, , It is a trainable parameter matrix. , It is any positive integer; It is the Kronecker product; It is the Kronecker product, specifically expressed as: ; S42, Lightweight Unsupervised Feature Selection Network It consists of several modules, among which... The integer is used; the module contains 7 core sub-modules, namely clustering module, weight score module, feature representation module, high-order matrix representation module, graph representation module, soft threshold selection module and weight update module; S5. A lightweight unsupervised feature selection network built on S4 is used as input. After iterative calculation, the network outputs the optimal feature weight matrix. The importance of features is evaluated based on the optimal feature weight matrix, and a target feature subset is selected. The audio data is then denoised based on the target feature subset, and the audio data after feature selection is output, thus completing the audio denoising.

2. The lightweight, robust, unsupervised feature selection-based audio denoising method according to claim 1, characterized in that, The clustering module in S42 is used to process clustering information calculations during the updating of the cluster center matrix, cluster representation matrix, and feature weight matrix. First, to address the complexity of calculating the inverse matrix, an inverse matrix network is constructed. The matrix inversion is transformed into an optimization problem; the above optimization problem is solved based on the momentum gradient descent method, and the momentum parameters and step size parameters of the momentum gradient descent are transformed into a learnable form; a subnetwork of 4 Kronecker linear layers is used to obtain the inverse matrix of the matrix, wherein the Kronecker linear layers are used to reduce the number of fully connected parameters in the learnable form; Secondly, to address the complexity issue of singular value decomposition, an orthogonal approximation network is constructed. The singular value decomposition is transformed into an orthogonal constraint optimization problem; the orthogonal matrix of the singular value decomposition is approximately output through a self-attention network and trainable parameters. Finally, two Kronecker linear layers are introduced to learn the feature representations of the audio data matrix, cluster center matrix, and cluster representation matrix, respectively, and the product is output. ;in, It is a clustering module calculation The approximate result.

3. The lightweight, robust, unsupervised feature selection-based audio denoising method according to claim 1, characterized in that, The S42 module, comprising the weighted score module, feature representation module, and higher-order matrix representation module, contains the following components: The weight score module is used to calculate the scaling factor during the feature weight matrix update process. It learns the internal structure information of the feature weight matrix through a cascaded structure of Kronecker linear layers + activation functions + Kronecker linear layers + Softplus layers, and outputs the result. ;in, It is a scaling factor Approximate results; The feature representation module simplifies large-scale matrix calculations while preserving high-order information of the data, generating interactive information that characterizes the internal dynamics of the audio data, corresponding to the feature weight matrix update process. The calculation is as follows: The audio data is learned in the feature space through two Kronecker linear layers. Different representations in; After multiplying the output matrices of the two Kronecker linear layers, the output is... ;in, Representative features Approximate results; The higher-order matrix representation module is used to generate higher-order matrices that represent the interaction between weight information and data information, corresponding to the feature weight matrix update process. The calculation is as follows: By introducing two Kronecker linear layers, the feature information of the feature weight matrix and the audio data matrix are learned respectively; Then obtain the interaction results of feature information and output them. ;in, It is a product of higher-order matrices. The approximate result.

4. The lightweight, robust, unsupervised feature selection-based audio denoising method according to claim 1, characterized in that, The soft threshold selection module in S42 is used to update the outlier matrix, and its specific functions are as follows: By introducing a mask To achieve smooth selection; Two Kronecker linear layers are used to obtain interaction information and output... ;in, It is calculated by the soft threshold selection module. The approximate result.

5. The lightweight, robust unsupervised feature selection-based audio denoising method according to claim 1, characterized in that, The S42 graph representation module is used to characterize the collaborative changes of features of the original data under a low-dimensional representation graph. The specific content is as follows: First, represent the low-dimensional dimension. Converted into a three-dimensional tensor, i.e. , ;in, and element values ​​and Consistent; Next, calculate the graph manifold matrix: ; in, For feature dimension Summation, Normalize the rows; Finally, two Kronecker linear layers are introduced to learn the directional differences of the original features under graph constraints and the basic orientation of the original features. An outer product is then calculated to approximate the covariant behavior of the data under graph constraints, and the output is... ;in, It is an approximate calculation result of the covariant behavior of data under graph constraints.

6. The lightweight, robust, unsupervised feature selection-based audio denoising method according to claim 1, characterized in that, The weight update module in S42 is used to update the feature weight matrix. Normalization, the specific details are as follows: Define a diagonal matrix The The diagonal elements are the feature weight matrix. The reciprocal of; Diagonal matrix Multiplying with the feature weight matrix achieves the effect of multiplying the feature weight matrix. -norm Normalization.

7. The lightweight, robust, unsupervised feature selection-based audio denoising method according to claim 1, characterized in that, The feature-selected audio data output by S5 is shown below: ； in, It outputs the audio data after feature selection; It is the number of audio features selected in unsupervised feature selection; It is a feature selection function that evaluates the importance of features in the noise-reducing frequency data matrix based on the feature weight matrix, and selects the features before... Key audio features; It is the optimal feature weight matrix, specifically represented as: ;in, This is the solution function for the robust unsupervised feature selection audio denoising method, calculated by... The optimal feature weight matrix can be obtained.

Images