Rotary sound source identification and positioning method based on deep convolutional attention neural network

Abstract

The invention discloses a rotating sound source recognition and positioning method based on a deep convolutional attention neural network, and belongs to the field of rotating sound source recognition, and the method comprises the steps: firstly carrying out the motion compensation and modal expansion of a sound pressure cross-spectrum matrix through a modal decomposition rotating beam forming algorithm, and generating a sound source distribution diagram in a rotating scene; then the sound source distribution map is input into a deep convolution attention neural network model, the model adopts an encoder-decoder structure, a channel and a space attention mechanism are fused, key sound source features are dynamically focused, noise interference is suppressed, and high-precision reconstruction of the sound source distribution map is achieved; and finally, carrying out local maximum value detection on a prediction map output by the network, and extracting a sound source space position. According to the method, the Doppler effect and phase distortion problems in rotating sound source positioning are effectively overcome, and the positioning precision, the anti-noise performance and the calculation efficiency in a complex rotating field are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of rotating sound source recognition, and particularly relates to a rotating sound source recognition and localization method based on a deep convolutional attention neural network. Background Technology

[0002] Rotating machinery is widely used in aerospace, wind power generation, transportation, and other industrial fields. The noise generated during its operation has a significant impact on equipment performance and environmental comfort. Due to the complex effects of spatial position time-varying, Doppler frequency shift, and sound field modulation during the motion of rotating sound sources, traditional sound source localization methods based on the assumption of static location are difficult to apply directly to rotating scenarios. To address these issues, researchers have proposed various improved methods, such as the ROSI algorithm, virtual rotating array methods, and mode decomposition methods, which partially mitigate the interference of the Doppler effect on sound field modeling through motion compensation or reference frame transformation. Furthermore, high-resolution imaging algorithms such as Clean-SC and DAMAS have been introduced to improve the imaging quality of sound sources. In recent years, deep learning technology has been increasingly applied to the field of sound source localization. For example, predictions of static sound source distributions based on CNN, ResNet, and FCN network structures have been achieved, significantly improving localization accuracy and efficiency.

[0003] However, existing methods still have significant limitations in the identification and localization of rotating sound sources: traditional methods make overly idealistic assumptions under high-speed rotation or strong non-steady-state conditions, leading to decreased actual localization accuracy and high computational complexity; while existing deep learning models are mostly designed for static sound sources and lack the ability to represent the spatiotemporal characteristics of rotating sound sources, making it difficult to effectively handle frequency shifts and phase distortions caused by motion. Furthermore, insufficient spatial resolution in the low-frequency band, weak noise suppression capabilities, large model parameters, and poor real-time performance limit their application in complex rotating scenarios. Therefore, there is an urgent need to develop a sound source identification and localization method that can adapt to the characteristics of rotating sound fields and achieves both high accuracy and high efficiency. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for identifying and locating rotating sound sources based on a deep convolutional attention neural network, comprising:

[0005] A sound pressure cross-spectrum matrix is ​​constructed based on the collected sound pressure signals, and a rotating sound source distribution map is generated based on the rotating beamforming algorithm of mode decomposition.

[0006] The rotated sound source distribution map is input into a deep convolutional attention neural network model to obtain a predicted sound source distribution map;

[0007] Local maxima detection is performed based on the predicted sound source distribution map to determine the spatial location of the rotating sound source.

[0008] Optionally, generating the rotating sound source distribution map includes:

[0009] The sound source plane is divided into grids to form a focusing plane, and each grid point is used as the location of a potential sound source.

[0010] Multiple microphone sensors are arranged in the sound field to construct a measurement plane that is parallel to the sound source plane and coaxial with its center;

[0011] The time-domain sound pressure signal collected by the microphone sensor is received, and the sound pressure cross-spectrum matrix is ​​calculated.

[0012] The rotating beamforming algorithm based on mode decomposition processes the acoustic pressure cross spectrum matrix to generate a rotating sound source distribution map.

[0013] Optionally, the deep convolutional attention neural network model is trained with a rotated sound source distribution map as input and a target sound source distribution map as a supervision label, and outputs a predicted sound source distribution map.

[0014] Optionally, the deep convolutional attention neural network model includes an encoder and a decoder, wherein the encoder is used to extract deep features of the sound source distribution map, and the decoder is used to reconstruct the sound source distribution map.

[0015] Optionally, the encoder includes multiple encoding modules, each encoding module including a convolutional block and a downsampling layer. The convolutional block includes a depth-separable convolutional layer, a normalization layer and an activation function, and a convolutional block attention mechanism is introduced at the output.

[0016] Optionally, the depth-separable convolutional layers in the convolutional block consist of a sequence of depthwise convolutions and pointwise convolutions, used to extract local spatial features and perform inter-channel feature fusion.

[0017] Optionally, the convolutional block attention mechanism includes a channel attention module and a spatial attention module, which are used to weight the feature map in the channel dimension and the spatial dimension, respectively.

[0018] Optionally, the decoder includes multiple upsampling modules. Each upsampling module performs upsampling through bilinear interpolation and concatenates it with the feature map of the corresponding layer of the encoder before processing it through a double convolutional layer to restore spatial resolution.

[0019] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0020] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0021] Compared with the prior art, the present invention has the following advantages and technical effects:

[0022] This invention proposes a method for identifying and locating rotating sound sources based on a deep convolutional attention neural network. It innovatively combines traditional acoustic modeling with deep learning feature extraction, significantly improving the resolution and computational efficiency of rotating sound source imaging. By using a rotating beamforming algorithm based on mode decomposition, this invention effectively optimizes the sound field feature representation before model training. This method utilizes RBF maps to perform nonlinear spatial mapping of the sound field, overcoming the problems of time-varying sound field distribution and complex spatial modulation during the motion of rotating sound sources, thus providing high-quality input features for subsequent network feature extraction.

[0023] This invention combines deep learning technology, using the DCANN model and rotating beamforming mapping as input, to extract and reconstruct the spatial features of sound sources through an encoder-decoder structure. While maintaining the advantages of skip connections, the model introduces a hybrid attention mechanism, fusing channel and spatial attention modules. This enables dynamic focusing on key sound source regions and effective suppression of noise interference, significantly improving feature representation capabilities and localization accuracy. Simultaneously, the network employs a lightweight deep convolutional structure, greatly reducing the number of model parameters and computational complexity, thus improving training speed and real-time imaging performance.

[0024] The method of this invention has significant advantages in spatial resolution and main lobe width control in the low-frequency band, effectively suppressing sidelobe effects and enhancing imaging clarity. This invention also demonstrates superior performance in imaging accuracy, noise resistance, and computational efficiency. This method can achieve high-precision sound source localization even with a limited number of microphone arrays, significantly improving the practicality and robustness of rotating sound source localization, and providing a new technical approach for the identification of noise sources and sound field reconstruction in complex rotating machinery. Attached Figure Description

[0025] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0026] Figure 1 This is a flowchart of the rotating sound source identification and localization method based on a deep convolutional attention neural network according to an embodiment of the present invention;

[0027] Figure 2 This is a framework diagram of the rotating sound source identification and localization method based on a deep convolutional attention neural network according to an embodiment of the present invention;

[0028] Figure 3 This is a schematic diagram of the specific structure of the module in an embodiment of the present invention;

[0029] Figure 4 The diagram shows the location results of the dual sound sources according to an embodiment of the present invention; wherein, (a) is the location result diagram when the dual sound sources are located at [-0.21, 0.17], [0.17, -0.15], f=1600 Hz, and z=0.5 m; (b) is the location result diagram when the dual sound sources are located at [-0.21, 0.17], [0.17, -0.15], f=3000 Hz, and z=0.5 m; and (c) is the location result diagram when the dual sound sources are located at [-0.21, 0.17], [0.17, -0.15], f=6000 Hz, and z=0.5 m. Detailed Implementation

[0030] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0031] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0032] Example 1

[0033] This embodiment provides a method for identifying and locating rotating sound sources based on a deep convolutional attention neural network, including:

[0034] Step 1: Collect and construct the sound pressure cross-spectrum matrix (CSM), perform motion compensation and modal expansion on the cross-spectrum matrix based on the RBF algorithm of modal decomposition, and then calculate and generate the sound source distribution map in the rotating scene;

[0035] Step 2: Input the rotating sound source distribution map into the DCANN model, and obtain the transformed rotating sound source distribution map through network inference; the DCANN is trained with the rotating sound source distribution map as input and the target sound source distribution map as supervision label to obtain parameters;

[0036] Step 3: Perform local maxima detection on the predicted sound source distribution map to obtain the sound source location.

[0037] Furthermore, obtaining the rotating sound source distribution map includes:

[0038] Step 1.1: Divide the plane where the sound source is located into a grid, discretize the sound source calculation area, form a focusing plane, and use each grid point on this plane as the potential sound source location for subsequent beamforming calculations;

[0039] Step 1.2: Arrange multiple microphone sensors in the sound field formed by several sound sources to construct a planar microphone array as a measurement plane. The measurement plane is set parallel to the sound source plane, and the center of the measurement plane is coaxial with the rotation center of the rotating sound source to ensure the uniformity and symmetry of the spatial sampling of the sound field of the rotating sound source by the measurement array.

[0040] Step 1.3: Receive time-domain sound pressure signal data collected by the microphone array, wherein the sound pressure signal data consists of the target rotating sound source signal and the ambient noise signal;

[0041] Step 1.4: Preprocess the time-domain sound pressure signal data to eliminate noise interference, calculate the sound pressure cross spectrum matrix, and generate a rotating sound source distribution map using the RBF algorithm based on mode decomposition.

[0042] Furthermore, obtaining the predicted rotating sound source distribution map includes:

[0043] Step 2.1: Input the rotating sound source distribution map into the initial convolutional layer of the DCANN model. The initial convolutional layer uses a single 3×3 convolutional kernel to perform local receptive field modeling and feature extraction on the input image, thereby obtaining the low-level feature information of the rotating sound source distribution map.

[0044] Step 2.2: Input the basic features into the encoder of the DCANN model. The encoder consists of four coding modules composed of convolutional blocks (CBAM-ConvNeXt modules) and downsampling layers. Each coding module is used to extract deep feature information of the sound source distribution map layer by layer.

[0045] The CBAM-ConvNeXt module consists of a depthwise separable convolutional layer, a LayerNorm normalization layer, a fully connected mapping layer, and a GELU activation function. It also introduces a convolutional block attention mechanism (CBAM) at the output to achieve local feature enhancement and adaptive global feature extraction, thereby improving the expressive power of multi-scale spatial features.

[0046] The Downsample layer uses a convolution operation with a stride of 2 to achieve spatial downsampling, which is used to compress the feature map size, expand the receptive field and reduce computational complexity, thereby obtaining a more hierarchical representation of sound source features;

[0047] Step 2.3: In the encoder module, each convolutional block sequentially includes a depth-separable convolutional layer, a LayerNorm normalization layer, and a GELU activation function, which are used to extract multi-scale spatial features and perform nonlinear mapping on the input sound source distribution features.

[0048] The depth-separable convolutional layer consists of a 7×7 depthwise convolution and a 1×1 pointwise convolution in sequence;

[0049] Depthwise convolution is used to extract local spatial features independently within each channel while preserving spatial structure information; pointwise convolution is used to perform linear combination between channels to achieve feature fusion and reduce the number of parameters.

[0050] After the convolution operation is completed, the LayerNorm normalization layer calculates the mean and variance of the feature map in the channel dimension to achieve standardization of the features of each channel, thereby reducing the difference in feature distribution, improving network stability and accelerating training convergence.

[0051] After normalization, the GELU activation function is used to perform nonlinear mapping on the feature map, so that the model changes smoothly in the low response region and maintains saturation characteristics in the high response region, thereby improving the feature representation ability.

[0052] Step 2.4: Introduce a Convolutional Block Attention (CBAM) mechanism into the convolutional blocks of each encoding module to adaptively enhance feature representation capabilities in both channel and spatial dimensions;

[0053] The CBAM includes a channel attention module and a spatial attention module.

[0054] The channel attention module extracts channel-level global information through global average pooling and global max pooling, generates channel weight coefficients through a 1×1 convolution with shared weights and a Sigmoid activation function, and weights each channel of the input feature to enhance key channel features and suppress irrelevant channels.

[0055] The spatial attention module performs average pooling and max pooling operations on the channel-weighted feature map along the channel dimension, concatenates the resulting feature maps and extracts spatial correlation through 7×7 convolution, and uses the Sigmoid activation function to generate a spatial attention map, which is then weighted along the spatial dimension to highlight the feature response of the salient region of the sound source.

[0056] By embedding the CBAM mechanism in each coding module, the model can adaptively focus on the key channels and spatial positions of the rotating sound source at different levels, thereby improving the ability to express the characteristics of the rotating sound source under complex sound fields and suppressing noise interference.

[0057] Step 2.5: The decoder includes four upsampling modules that use bilinear interpolation for upsampling and are connected to dual convolutional layers, used to gradually restore spatial resolution and reconstruct the sound source distribution map;

[0058] Each upsampling module upsamples the deep features from the previous layer using bilinear interpolation and concatenates them with the feature map output from the corresponding encoder layer to form a fused feature map.

[0059] The fused feature map is processed by a dual convolution module to restore local details and high-level semantic features, thereby achieving a fine reconstruction of the spatial structure of the rotating sound source.

[0060] Step 2.6: Compress the output of the last decoder module through a 1×1 convolution to obtain a single-channel reconstruction result image;

[0061] The reconstructed result map is a distribution map of rotating sound sources after data conversion, which includes energy distribution information of rotating sound sources in the spatial domain.

[0062] Furthermore, obtaining the position of the rotating sound source includes:

[0063] Step 3.1: Perform local peak analysis on the predicted sound source distribution map, search and extract representative extreme points in each row of the distribution matrix, and select the two peaks with the highest response intensity as candidate sound source locations.

[0064] Step 3.2: Perform neighborhood consistency check on the candidate peak. By comparing the sound pressure amplitude of the candidate point with that of each point in its surrounding area, determine whether the candidate point has a unique maximum response in the local area. If the verification result is valid, the point is confirmed as a valid local maximum and arranged in order according to its amplitude.

[0065] Step 3.3: Based on the sorted maximum value information, coordinate mapping and geometric transformation methods are used to convert the matrix index positions into actual spatial coordinates, and then the specific positions of each sound source in the physical coordinate system are calculated.

[0066] Example 2

[0067] like Figure 1 As shown, this embodiment provides a method for identifying and locating rotating sound sources based on a deep convolutional attention neural network, including:

[0068] A sound pressure cross-spectrum matrix is ​​constructed based on the collected sound pressure signals, and a rotating sound source distribution map is generated based on the rotating beamforming algorithm of mode decomposition.

[0069] The rotated sound source distribution map is input into a deep convolutional attention neural network model to obtain a predicted sound source distribution map;

[0070] Local maxima detection is performed based on the predicted sound source distribution map to determine the spatial location of the rotating sound source.

[0071] Step 1: Acquire time-domain sound pressure signals under a rotating sound source field using an array of sensors, and calculate the sound pressure cross-spectrum matrix. A rotating beamforming (RBF) algorithm based on modal decomposition is used to perform motion compensation and modal expansion on the cross-spectrum matrix to eliminate the Doppler effect of the rotating sound source, thereby generating a sound source distribution map under the rotating scene. Step 2: Input the rotating sound source distribution map into the DCANN model, and obtain the transformed rotating sound source distribution map through network inference. The DCANN model is trained using the rotating sound source distribution map as input and the target sound source distribution map as the supervised label, achieving nonlinear transformation and fine reconstruction of the sound source distribution map through end-to-end feature mapping learning. Step 3: Perform local maxima detection on the predicted sound source distribution map output by the DCANN model, extract the sound source energy peak points, and thus determine the spatial location of the sound source, achieving accurate localization of the rotating sound source.

[0072] Specifically:

[0073] Step 1: Collect sound pressure signals to construct a cross-spectral matrix (CSM), and use the rotating beamforming (RBF) algorithm of mode decomposition to perform motion compensation and mode expansion to generate a sound source distribution map in a rotating scene.

[0074] Step 1 specifically involves:

[0075] Step 1.1: In a sound field formed by K sound sources, M microphone sensors are arranged, forming a planar array defined as the measurement plane. The measurement plane is parallel to the sound source plane, with a distance of z between them. The microphone array is arranged in a ring to acquire spatial sound pressure information.

[0076] Step 1.2: The plane containing the sound source is meshed, and the continuous sound source distribution is discretized into several discrete sound source units, forming a computational plane composed of multiple grid points, defined as the focusing plane. The focusing plane contains N grid points, each of which is also called a focal point. Each focal point serves as the potential sound source location and is used for the estimation and reconstruction of sound source energy.

[0077] Step 1.3: Receive the sound pressure signal data X collected by the sensor array, the data including the target sound source signal and the background noise signal. Process the sound pressure signal to calculate the sound pressure cross-spectral matrix C, which is used to characterize the spatial correlation between the sensors.

[0078] ;

[0079] in, H denotes the conjugate transpose of a matrix.

[0080] Step 1.4: Calculate the turning vector of the rotating sound source.

[0081] ;

[0082] in, The Green's function for sound propagation in a rotating coordinate system in a free field is expressed as:

[0083] ;

[0084] The range of the number of terms in an infinite series expansion can be approximated as follows, provided that the truncation error is small: , ,in , , , , Indicates the radius of rotation. This indicates rounding down to the nearest integer. It is the rotational angular frequency. and These are the spherical coordinate vectors of grid points in the microphone and scanning plane, respectively. The expression for the normalization factor is:

[0085] ;

[0086] The wavenumber after frequency shift caused by the rotation effect is:

[0087] ;

[0088] for Step Legendre functions of the first kind

[0089] ;

[0090] ;

[0091] in, for The first-order Bessel function of the sphere of the first kind yes Rank Ball-like Hankel function.

[0092] Step 1.4: At grid points The output of RBF is:

[0093] ;

[0094] Step 1.5: Output Normalization is performed to perform acoustic imaging, resulting in a sound source distribution map.

[0095] Step 2: Input the rotated sound source distribution map into the DCANN model, and obtain the transformed sound source distribution map through network inference. DCANN uses the rotated sound source distribution map as input and the target sound source distribution map as labels for supervised training to obtain model parameters. The DCANN model structure is as follows: Figure 2 As shown.

[0096] Step 2 specifically involves:

[0097] Step 2.1: Input the rotating sound source distribution map into the initial convolutional layer of the DCANN model. The initial convolutional layer uses a single 3×3 convolutional kernel to perform local receptive field modeling and feature extraction on the input image, thereby obtaining the low-level feature information of the rotating sound source distribution map.

[0098] Step 2.2: Input the basic features into the encoder of the DCANN model. The encoder consists of four cascaded encoding modules, each containing a CBAM-ConvNeXt module and a downsample layer, used to extract deep features of the sound source distribution map layer by layer and construct a hierarchical feature representation.

[0099] The expression formula for the ConvNeXt module:

[0100] ;

[0101] in Represents depthwise convolution. and These represent channel-expanded convolution and channel-compressed convolution, respectively.

[0102] The CBAM-ConvNeXt module consists of depthwise separable convolutional layers, LayerNorm normalization layers, fully connected mapping layers, and GELU activation functions. A convolutional block attention mechanism (CBAM) is introduced at its output to enhance local details and adaptively extract global semantic features, thereby improving the expressive power of multi-scale spatial features.

[0103] The lower downsample layer achieves spatial downsampling through convolution with a stride of 2, which is used to compress the feature map size, expand the receptive field and reduce computational overhead, thereby obtaining a more hierarchical representation of sound source features;

[0104] Step 2.3: In the encoder module, each convolutional block sequentially includes a depth-separable convolutional layer, a LayerNorm normalization layer, and a GELU activation function, which are used to extract multi-scale spatial features and perform nonlinear mapping on the input sound source distribution features.

[0105] The depthwise separable convolutional layer consists of a 7×7 depthwise convolution and a 1×1 pointwise convolution in sequence. The depthwise convolution is used to independently extract local spatial features in each channel to preserve spatial structure information. The pointwise convolution is used to perform linear combination between channels to achieve cross-channel feature fusion and effectively reduce the number of parameters.

[0106] After the convolution operation is completed, the feature map is normalized in the channel dimension by the LayerNorm normalization layer to calculate and eliminate the difference in mean and variance between channels, thereby reducing feature distribution offset, improving network stability and speeding up training convergence.

[0107] After normalization, the GELU activation function is used to perform nonlinear mapping on the feature map, so that the model exhibits smooth changes in the low response region and maintains saturation characteristics in the high response region, thereby enhancing the feature expression ability and the model's nonlinear fitting performance.

[0108] Step 2.4: Introduce a Convolutional Block Attention (CBAM) mechanism into the convolutional blocks of each encoding module to adaptively enhance feature representation capabilities in both channel and spatial dimensions.

[0109] The CBAM consists of a channel attention module and a spatial attention module. The CBAM module expression is as follows:

[0110] ;

[0111] in, This indicates an element-wise weighted operation. and These represent channel attention mapping and spatial attention mapping, respectively.

[0112] The channel attention module extracts channel-level global information through global average pooling and global max pooling, and inputs the results of the two into a 1×1 convolutional layer with shared weights. The channel attention weights are generated by the sigmoid activation function, which weights each channel of the input feature, thereby enhancing key channel features and suppressing irrelevant channel information.

[0113] The spatial attention module performs average pooling and max pooling operations on the channel-weighted feature map along the channel dimension, respectively. The resulting feature maps are then concatenated and input into a 7×7 convolutional layer to extract spatial correlation. A spatial attention map is generated by the Sigmoid activation function, and the feature map is weighted in the spatial dimension to highlight the feature response of the salient region of the sound source.

[0114] By embedding the CBAM mechanism in each coding module, the model can adaptively focus on the key channels and spatial positions of the rotating sound source at different levels, effectively improving the feature representation ability under complex sound fields and suppressing background noise interference.

[0115] Step 2.5: The decoder includes four upsampling modules that use bilinear interpolation for upsampling and are connected to dual convolutional layers, used to gradually restore the spatial resolution of the feature map and reconstruct the sound source distribution map.

[0116] In each upsampling module, the deep features of the previous layer are first upsampled through bilinear interpolation to expand the spatial size of the feature map. Then, the upsampled features are concatenated with the feature map output by the corresponding encoder layer in the channel dimension to form a fused feature map, thereby achieving the supplementation and integration of cross-layer feature information.

[0117] The fused feature map is processed by a dual convolution module to recover detailed features and high-level semantic features based on the fusion of global and local information, thereby achieving a refined reconstruction of the spatial structure of the rotating sound source and improving the model's localization accuracy and reconstruction quality.

[0118] Step 2.6: Input the output feature map of the last decoder module into a 1×1 convolutional layer for channel compression to achieve linear fusion and dimensionality reduction of multi-channel features, and finally obtain the single-channel reconstruction result map.

[0119] The reconstructed result map is a rotating sound source distribution map after data conversion, which contains the energy distribution information of the rotating sound source in the spatial domain. It is used to characterize the intensity characteristics of the sound source at different locations, and realize the spatial imaging and visualization reconstruction of the rotating sound source.

[0120] Step 3: Perform local maxima detection on the predicted sound source distribution map to obtain the sound source location.

[0121] Step 3 specifically involves:

[0122] Step 3.1: Perform local peak detection on the sound source distribution map predicted by the model, search and extract representative extreme points in each row of the distribution matrix, and select the two peaks with the highest response intensity as candidate sound source locations.

[0123] Step 3.2: Perform neighborhood consistency verification on the candidate peaks. By comparing the sound pressure amplitude of the candidate point with that of each pixel in its neighborhood, determine whether the point has a unique maximum response in the local region. If the determination result is true, the point is confirmed as a valid local maximum and sorted in order of its amplitude.

[0124] Step 3.3: Based on the sorted local maxima information, coordinate mapping and geometric transformation methods are used to convert the matrix index positions into spatial positions in the actual physical coordinate system, thereby determining the precise coordinates of each sound source in the spatial domain.

[0125] This process integrates local maximum detection, coordinate transformation, and sorting to improve the accuracy of predicting the location of rotating sound sources.

[0126] Figure 3 This is a schematic diagram of the specific structure of the module in an embodiment of the present invention.

[0127] Figure 4 (a) is the localization result diagram at the dual sound source positions [-0.21, 0.17], [0.17, -0.15], f=1600 Hz, z=0.5 m; (b) is the localization result diagram at the dual sound source positions [-0.21, 0.17], [0.17, -0.15], f=3000 Hz, z=0.5 m; (c) is the localization result diagram at the dual sound source positions [-0.21, 0.17], [0.17, -0.15], f=6000 Hz, z=0.5 m.

[0128] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0129] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0130] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for identifying and locating rotating sound sources based on a deep convolutional attention neural network, characterized in that, include: A sound pressure cross-spectrum matrix is ​​constructed based on the collected sound pressure signals, and a rotating sound source distribution map is generated based on the rotating beamforming algorithm of mode decomposition. The rotated sound source distribution map is input into a deep convolutional attention neural network model to obtain a predicted sound source distribution map; Local maxima detection is performed based on the predicted sound source distribution map to determine the spatial location of the rotating sound source.

2. The method according to claim 1, characterized in that, The generated rotating sound source distribution map includes: The sound source plane is divided into grids to form a focusing plane, and each grid point is used as the location of a potential sound source. Multiple microphone sensors are arranged in the sound field to construct a measurement plane that is parallel to the sound source plane and coaxial with its center; The time-domain sound pressure signal collected by the microphone sensor is received, and the sound pressure cross-spectrum matrix is ​​calculated. The rotating beamforming algorithm based on mode decomposition processes the acoustic pressure cross spectrum matrix to generate a rotating sound source distribution map.

3. The method according to claim 1, characterized in that, The deep convolutional attention neural network model takes a rotated sound source distribution map as input and a target sound source distribution map as a supervision label for training, and outputs a predicted sound source distribution map.

4. The method according to claim 1, characterized in that, The deep convolutional attention neural network model includes an encoder and a decoder. The encoder is used to extract deep features of the sound source distribution map, and the decoder is used to reconstruct the sound source distribution map.

5. The method according to claim 4, characterized in that, The encoder includes multiple encoding modules, each encoding module including a convolutional block and a downsampling layer. The convolutional block includes a depth-separable convolutional layer, a normalization layer and an activation function, and introduces a convolutional block attention mechanism at the output.

6. The method according to claim 5, characterized in that, The depth-separable convolutional layers in the convolutional block consist of depthwise convolution and pointwise convolution in sequence, and are used to extract local spatial features and perform inter-channel feature fusion.

7. The method according to claim 5, characterized in that, The convolutional block attention mechanism includes a channel attention module and a spatial attention module, which are used to weight the feature maps in the channel dimension and the spatial dimension, respectively.

8. The method according to claim 4, characterized in that, The decoder includes multiple upsampling modules. Each upsampling module performs upsampling through bilinear interpolation and concatenates it with the feature map of the corresponding layer of the encoder before processing it through a double convolutional layer to restore spatial resolution.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-8.

Images