Intelligent ant nest sound wave positioning weak signal denoising method combining SDM and LG-BPN

Abstract

An intelligent ant nest sound wave positioning weak signal denoising method combining SDM and LG-BPN comprises the following steps: step 1, obtaining termite sound wave data to obtain a real data sample set; mixing the real data sample set and the synthetic data sample set to obtain a training sample set; 2, generating a large number of training samples by using a shear wave embedding generator SDM according to the training sample set in the step 1, and obtaining an expanded training sample set; step 3, training the LG-BPN denoising network by using the training sample expanded by the shear wave embedding generator SDM as input; and step 4, inputting the collected ant nest sound wave positioning weak signal to be denoised into the trained LG-BPN denoising network, and outputting the denoised ant nest sound wave positioning weak signal. Aiming at the technical problems of low ant nest positioning speed and low positioning precision caused by small sample and low signal-to-noise ratio in ant nest sound wave positioning weak signals, the invention provides the method.

Description

Technical Field

[0001] The present invention relates to the field of signal processing technology, and in particular to an intelligent denoising algorithm for weak signals of ant nest acoustic positioning combining SDM and LG-BPN. Background Art

[0002] Termite damage is a major safety hazard for water conservancy projects and a contributing factor to serious hazards in river embankments. Accurately and efficiently detecting and locating termite nests is crucial for termite control.

[0003] Termites are highly social insects, often numbering tens of thousands within a single nest. Such large numbers of termites continuously emit sound during their activities, which manifests as continuous, long-wavelength, low-frequency pulses in the temporal domain. These pulses decay slowly along their path and exhibit a strong diffraction effect, providing crucial insights into nest location.

[0004] However, a low signal-to-noise ratio (SNR) consistently limits the accuracy of weak signal arrival identification. Intelligently detecting and identifying weak signals from ant nests, and obtaining accurate time difference information, is crucial to improving the accuracy of ant nest acoustic monitoring and positioning.

[0005] Existing technologies, such as those in the papers "Location of Shallow Underground Earthquake Sources Based on Distributed Sensors" and "Indoor Experimental Study of Passive Acoustic Detection Technology for Soil-dwelling Termite Nests," exhibit these technical deficiencies. To address the small sample size and low signal-to-noise ratio of weak acoustic signals from termite nests, this patent proposes an intelligent denoising algorithm for weak acoustic signals from termite nests, combining SDM and LG-BPN, to expand the capacity of weak acoustic monitoring signals and improve their signal-to-noise ratio. Summary of the Invention

[0006] The present invention addresses the technical problem that the weak signal of ant nest acoustic wave positioning has small sample size and low signal-to-noise ratio, which leads to slow ant nest positioning speed and low positioning accuracy.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is: An intelligent denoising method for weak signals of ant nest acoustic positioning combining SDM and LG-BPN, comprising the following steps: Step 1: Obtain termite sound wave data to obtain a real data sample set; mix the real data sample set and the synthetic data sample set to obtain a training sample set; Step 2: Generate a large number of training samples from the training sample set in step 1 using the shearlet embedding generator SDM to obtain an expanded training sample set; Step 3: Use the expanded training samples completed by the shearlet embedding generator SDM as input to train the LG-BPN denoising network; Step 4: Input the collected ant nest acoustic wave positioning weak signal to be denoised into the trained LG-BPN denoising network, and output the denoised ant nest acoustic wave positioning weak signal.

[0008] In step 1, the acoustic wave 2D staggered grid finite difference forward simulation algorithm is used to generate forward simulation signals under different main frequencies and different velocity models to obtain a synthetic signal sample set.

[0009] In step 2, the shearlet embedding generator SDM adopts a UNet structure with N upsampling blocks and N downsampling blocks, as well as skip connections between blocks of the same resolution, and replaces the normal upsampling and downsampling operators with frequency-aware blocks; at the lowest resolution, a frequency bottleneck block is used to better focus on low-frequency and high-frequency components; finally, in order to merge the original signal into different feature pyramids of the encoder.

[0010] The shearlet embedding generator SDM is specifically: Input features of the shearlet embedding generator SDM Input to the first convolution block and the first shearlet downsampling module respectively; The output end of the first convolution block is connected to the input end of the first downsampling module; the output feature of the first downsampling module is feature F1; The output end of the first shear wave downsampling module is connected to the input end of the second shear wave downsampling module, and the output features of the first shear wave downsampling module and the second shear wave downsampling module are feature F2 and feature F3 respectively; The features F1 and F2 are added together to output feature F4; feature F4 is input to the second shear wave downsampling module; the output feature of the second shear wave downsampling module is F5; The output features F3 and F5 of the second shear wave downsampling module are added together to output feature F6; Feature F6 is input to the first frequency bottleneck module. The output of the first frequency bottleneck module is connected to the input of the attention module, and the output of the attention module is connected to the input of the second frequency bottleneck module. The output of the second frequency bottleneck module is connected to the input of the first upsampling module, and the output of the first upsampling module is connected to the input of the second upsampling module. The output end of the second downsampling module is jump-connected to the input end of the first upsampling module; the output end of the first downsampling module is jump-connected to the input end of the second upsampling module; the output end of the second upsampling module is connected to the input end of the activation function and group normalization module, and the output end of the activation function and group normalization module is connected to the input end of the second convolution block; the output end of the second convolution block outputs the output features of the entire shearlet embedding generator SDM.

[0011] The first downsampling module includes N serially connected residual blocks, wherein the input end of the first residual block is connected to the output end of the first convolution block, and the output end of the Nth residual block is connected to the input end of the frequency-aware downsampling module; The second downsampling module includes N serially connected residual blocks, wherein the input end of the first residual block is connected to the first shear wave downsampling module and the output end of the first downsampling module, and the output end of the Nth residual block is connected to the input end of the frequency-aware downsampling module.

[0012] The first upsampling module includes N serially connected residual blocks, wherein the input end of the first residual block is connected to the output ends of the second frequency domain bottleneck module and the second downsampling module, and the output end of the Nth residual block is connected to the input end of the frequency-aware upsampling module; The second upsampling module includes N serially connected residual blocks, wherein the input end of the first residual block is connected to the output ends of the first upsampling module and the first downsampling module, and the output end of the Nth residual block is connected to the input end of the frequency-aware upsampling module.

[0013] The frequency domain bottleneck module first divides the feature map into low-frequency subbands and high-frequency subbands through discrete shearlet transform, and then passes the low-frequency subbands as input to the residual block for processing. The processed low-frequency subbands and the original high-frequency subbands are output through inverse shearlet transform.

[0014] The LG-BPN denoising network consists of two parallel branches, one for reconstructing local and the other for reconstructing global context. The LG-BPN denoising network extracts features from the input through a 1-layer Conv layer, and then inputs the features into the local and global branches respectively. The local branch first uses a 9×9 DSPMC module to extract dense features through downsampling to break spatial correlation; then the feature map is processed by dilated convolution with a dilation rate of 2; The global branch first passes through a 21×21 DMPMC module with a larger receptive field, and then is processed by a DTB module; finally, the local and global information are fused to obtain the final output.

[0015] The kernel formula used by the DSPMC module is: (5); It is the core of DSPMC. is the kernel of ordinary convolution, is a mask used to filter highly correlated pixels.

[0016] The DTB module uses channel attention to perform long-range context aggregation and make full use of global information. The attention formula used by the DTB module is as follows: (6); (7); + (8); X is the input, represents element-wise multiplication, LN represents layer normalization, and Represents a 3×3 dilated convolution.

[0017] Compared with the prior art, the present invention has the following technical effects: 1) The present invention adopts a joint shearlet embedding generator SDM and LG-BPN denoising network to complete the denoising of weak signals of ant nest acoustic positioning. By expanding the real noise and synthetic noise sample sets through the shearlet embedding generator SDM, only a small sample training set is needed to complete the training of the LG-BPN denoising network. LG-BPN adopts two parallel branches to effectively extract local details and capture global features. The local branch adopts dilated convolution to achieve a denser receptive field, thereby improving local texture recovery. The global branch learns global features through the DTB module, aggregates global context, and reduces computational complexity. The LG-BPN denoising network performs well in the denoising task of synthetic and real training samples, greatly improving the denoising effect of weak signals of ant nest acoustic positioning. 2) The shearlet embedding generator (SDM) in this invention uses the shearlet transform to achieve more accurate frequency domain analysis and multi-scale analysis. It expands the dataset through the diffusion generative model, solving the small sample problem and completing the denoising task with only a small sample set. 3) In this paper, LG-BPN adopts dilated convolution to extract rich local features and uses channel attention to focus on global features, further optimizing the network's feature capture capability, thereby significantly improving the network's denoising effect and completing the weak signal denoising task of ant nest acoustic localization with high quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention will be further described below with reference to the accompanying drawings and examples: Figure 1 Workflow diagram of the combined SDM and LG-BPN method; Figure 2 Schematic diagram of the structure of the shear wave embedding generator SDM in the present invention; Figure 3 for Figure 2 Schematic diagram of the mid-frequency domain perception downsampling module; Figure 4 for Figure 2 Schematic diagram of the mid-frequency domain perception upsampling module; Figure 5 for Figure 2Schematic diagram of the shear wave downsampling module; Figure 6 for Figure 2 Schematic diagram of the bottleneck block in the mid-frequency domain; Figure 7 This is the LG-BPN network structure diagram in the present invention; Figure 8 for Figure 7 DTB network structure diagram; Figure 9 This is a schematic diagram of the weak signal denoising results of ant nest acoustic positioning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] like Figure 1 As shown in the figure, a method for denoising weak signals of ant nest acoustic localization by combining SDM and LG-BPN is proposed. The specific steps are as follows: Step 1: Conduct acoustic monitoring in an ant-infested area in China to obtain on-site termite acoustic data and obtain a real data sample set. Use the acoustic 2D staggered grid finite difference forward modeling algorithm to generate forward simulation signals under different main frequencies and different velocity models to obtain a synthetic signal sample set. Mix the real data sample set and the synthetic data sample set to obtain a training sample set. Step 2: Generate a large number of training samples from the training sample set in step 1 using the Shearlet Embedding Generator (SDM) to obtain an expanded training sample set. The principle of the Shearlet Embedding Generator (SDM) is as follows: Figure 2 Figure 1 is a schematic diagram of the Shearlet Embedding Generator (SDM). It uses a UNet architecture with N upsampling blocks and N downsampling blocks, as well as skip connections between blocks of the same resolution. Frequency-aware blocks replace the normal upsampling and downsampling operators. At the lowest resolution, a frequency bottleneck block is used to better focus on low- and high-frequency components. Finally, to merge the original signal into the different feature pyramids of the encoder, the present invention introduces frequency residual connections using the Shearlet downsampling layer.

[0020] Input features of the shearlet embedding generator SDM Input to the first convolution block and the first shearlet downsampling module respectively; The output end of the first convolution block is connected to the input end of the first downsampling module; the output feature of the first downsampling module is feature F1; The output end of the first shear wave downsampling module is connected to the input end of the second shear wave downsampling module, and the output features of the first shear wave downsampling module and the second shear wave downsampling module are feature F2 and feature F3 respectively; The features F1 and F2 are added together to output feature F4; feature F4 is input to the second shear wave downsampling module; the output feature of the second shear wave downsampling module is F5; The output features F3 and F5 of the second shear wave downsampling module are added together to output feature F6; Feature F6 is input to the first frequency bottleneck module. The output of the first frequency bottleneck module is connected to the input of the attention module, and the output of the attention module is connected to the input of the second frequency bottleneck module. The output of the second frequency bottleneck module is connected to the input of the first upsampling module, and the output of the first upsampling module is connected to the input of the second upsampling module. The output end of the second downsampling module is jump-connected to the input end of the first upsampling module; the output end of the first downsampling module is jump-connected to the input end of the second upsampling module; the output end of the second upsampling module is connected to the input end of the activation function and group normalization module, and the output end of the activation function and group normalization module is connected to the input end of the second convolution block; the output end of the second convolution block outputs the output features of the entire shearlet embedding generator SDM.

[0021] The first downsampling module includes N serially connected residual blocks, wherein the input end of the first residual block is connected to the output end of the first convolution block, and the output end of the Nth residual block is connected to the input end of the frequency-aware downsampling module; The second downsampling module includes N serially connected residual blocks, wherein the input end of the first residual block is connected to the first shear wave downsampling module and the output end of the first downsampling module, and the output end of the Nth residual block is connected to the input end of the frequency-aware downsampling module.

[0022] The first upsampling module includes N serially connected residual blocks, wherein the input end of the first residual block is connected to the output ends of the second frequency domain bottleneck module and the second downsampling module, and the output end of the Nth residual block is connected to the input end of the frequency-aware upsampling module; The second upsampling module includes N serially connected residual blocks, wherein the input end of the first residual block is connected to the output ends of the first upsampling module and the first downsampling module, and the output end of the Nth residual block is connected to the input end of the frequency-aware upsampling module.

[0023] Figure 3 It is a frequency-aware downsampling module. Figure 4 It is a frequency-aware upsampling module. DST is discrete shearlet transform, IST is inverse shearlet transform, hi-sub is high-frequency subband, and lo-sub is low-frequency subband. It uses the inherent properties of shearlet transform to better achieve upsampling and downsampling, and enhance the perception of high-frequency information. The downsampling block accepts input features , latent variables z and temporal embedding t are processed through a series of layers to return downsampled features and high-frequency subbands. The returned high-frequency subbands serve as additional input for upsampling features based on frequency cues in the upsampling block.

[0024] Figure 3 The input and latent variable z in the frequency-aware downsampling module enter AdaGN, then pass through the Conv module, and then enter the DST module. The high-frequency subband at the output of the DST module enters the frequency-domain-aware upsampling module, while the low-frequency subband is connected to the time embedding t and enters the AdaGN module. The latent variable z also enters the AdaGN module, and the output of the AdaGN module is connected to the input of the Conv module.

[0025] Figure 4 The input and latent variable z in the frequency-aware upsampling module enter the AdaGN module, then pass through the Conv module, and then enter the IST module. At the same time, the high-frequency subband enters the IST module. The output of the IST module is connected with the time embedding t and enters the AdaGN module. The latent variable z also enters the AdaGN module. The output of the AdaGN module is connected to the input of the Conv module.

[0026] Figure 5 This is the shearlet downsampling module. The shearlet downsampling module first uses discrete shearlet transform to obtain high-frequency and low-frequency information, then concatenates them and outputs them through a convolution.

[0027] Figure 6 This is the frequency-domain bottleneck module. It first shuffles the feature map into low-frequency and high-frequency subbands using the discrete shearlet transform. The low-frequency subbands are then passed as input to the residual block for processing. The processed low-frequency subbands and the original high-frequency subbands are then output using the inverse shearlet transform. This frequency bottleneck operation allows for focused learning of intermediate features in the low-frequency subbands while preserving high-frequency details.

[0028] The adversarial loss function of the Shearlet Diffusion Model is as follows: (1); (2); In formulas 1 and 2, D is the discriminator. is a pseudo sample at time step t, is the corresponding perturbation sample.

[0029] The reconstruction loss function is as follows: (3); In formula 3 It is a clean sample. , is the unperturbed sample generated by the generator G, with latent variables z~N(0,1).

[0030] The overall loss function is as follows: (4); In formula 4 is a weighting hyperparameter.

[0031] Step 3: Use the expanded training samples from SDM as input to train the LG-BPN denoising network. Figure 7 As shown in Figure 1, the LG-BPN network consists of two parallel branches, one for reconstructing local and the other for global context. LG-BPN extracts features from the input through a single Conv layer, which are then fed into the local and global branches. The local branch first uses a 9×9 DSPMC module to downsample the densely extracted features to break spatial correlations. The feature map then undergoes dilated convolution with a dilation rate of 2. The global branch first passes through a 21×21 DMPMC module with a larger receptive field, followed by DTB processing. Finally, the local and global information are fused to produce the final output.

[0032] DSPMC is a densely sampled block mask convolution that can extract more local information while avoiding misjudgment of noise from strongly correlated adjacent pixels. The kernel calculation formula of DSPMC is: (5); It is the core of DSPMC. is the kernel of ordinary convolution, is a mask used to filter highly correlated pixels.

[0033] like Figure 8 As shown in Figure 2, the DTB module uses channel attention to aggregate long-range context and fully utilize global information. Since DTB only calculates attention on feature channels and does not involve attention calculations in spatial dimensions, its computational complexity is linear with respect to the input dimension, effectively reducing the cost of global information calculation. The DTB attention formula is as follows: (6); (7); + (8); X is the input, represents element-wise multiplication, LN represents layer normalization, and Represents a 3×3 dilated convolution.

[0034] In the present invention, the shearlet embedding generator SDM expands the small sample data set of weak signals for ant nest acoustic positioning through a diffusion generation model. The LG-BPN uses dilated convolution to extract rich local features and the channel attention mechanism to focus on global features, thereby enhancing the network's feature extraction capability and improving the network's accuracy.

[0035] Step 4: Input the collected ant nest acoustic wave positioning weak signal to be denoised into the trained LG-BPN denoising network, and output the denoised ant nest acoustic wave positioning weak signal.

[0036] Example: The present invention denoises the signal in the ant nest acoustic weak signal denoising project. In the ant nest model, multiple acoustic wave sensors and receivers are deployed to receive weak acoustic wave signals. The data set of this experiment consists of 10,000 synthetic signals and 500 real signals, where the synthetic signals are generated by the acoustic wave 2D staggered grid finite difference forward simulation algorithm, and the real signals are obtained by collecting termite acoustic wave data on site. In addition, the experimental platform for training the model is configured with Pycharm, Python3.8 and Pytorch1.12 deep learning framework, and the hardware is Intel(R) Xeon(R) Silver 4314 processor, 94G memory and NVIDIA GeForce RTX 3090 GPU with 24G video memory. The denoising results of weak signals of ant nest acoustic positioning are shown as follows: Figure 9 shown.

[0037] from Figure 9 The results clearly demonstrate that the proposed method is highly effective in denoising weak acoustic signals in an ant nest environment. Compared to traditional denoising methods, this method not only effectively suppresses complex background noise but also significantly improves the signal-to-noise ratio while preserving key signal features, extracting a clearer, more structurally complete target signal waveform. It demonstrates excellent robustness and generalization capabilities.

[0038] In summary, the combination of SDM and LG-BPN provides a novel and practical solution for the denoising of weak signals in ant nest acoustic positioning, which has good engineering application prospects and research value in complex underground structures and strong interference environments.

Claims

1. A method for intelligent denoising of weak signals of ant nest acoustic localization by combining SDM and LG-BPN, characterized in that: The following steps are involved: Step 1: Obtain termite sound wave data to obtain a real data sample set; mix the real data sample set and the synthetic data sample set to obtain a training sample set; Step 2: Generate a large number of training samples from the training sample set in step 1 using the shearlet embedding generator SDM to obtain an expanded training sample set; Step 3: Use the expanded training samples completed by the shearlet embedding generator SDM as input to train the LG-BPN denoising network; Step 4: Input the collected ant nest acoustic wave positioning weak signal to be denoised into the trained LG-BPN denoising network, and output the denoised ant nest acoustic wave positioning weak signal.

2. The method according to claim 1, characterized in that In step 1, the acoustic wave 2D staggered grid finite difference forward simulation algorithm is used to generate forward simulation signals under different main frequencies and different velocity models to obtain a synthetic signal sample set.

3. The method according to claim 1, characterized in that In step 2, the shearlet embedding generator SDM adopts a UNet structure with N upsampling blocks and N downsampling blocks, as well as skip connections between blocks of the same resolution, and replaces the normal upsampling and downsampling operators with frequency-aware blocks; at the lowest resolution, a frequency bottleneck block is used to better focus on low-frequency and high-frequency components; finally, in order to merge the original signal into different feature pyramids of the encoder.

4. The method according to any one of claims 1 to 3, characterized in that The shear wave embedding generator SDM is specifically: Input features of the shearlet embedding generator SDM Input to the first convolution block and the first shearlet downsampling module respectively; The output end of the first convolution block is connected to the input end of the first downsampling module; the output feature of the first downsampling module is feature F1; The output end of the first shear wave downsampling module is connected to the input end of the second shear wave downsampling module, and the output features of the first shear wave downsampling module and the second shear wave downsampling module are feature F2 and feature F3 respectively; The features F1 and F2 are added together to output feature F4; feature F4 is input to the second shear wave downsampling module; the output feature of the second shear wave downsampling module is F5; The output features F3 and F5 of the second shear wave downsampling module are added together to output feature F6; Feature F6 is input to the first frequency bottleneck module. The output of the first frequency bottleneck module is connected to the input of the attention module, and the output of the attention module is connected to the input of the second frequency bottleneck module. The output of the second frequency bottleneck module is connected to the input of the first upsampling module, and the output of the first upsampling module is connected to the input of the second upsampling module. The output end of the second downsampling module is jump-connected to the input end of the first upsampling module; the output end of the first downsampling module is jump-connected to the input end of the second upsampling module; the output end of the second upsampling module is connected to the input end of the activation function and group normalization module, and the output end of the activation function and group normalization module is connected to the input end of the second convolution block; the output end of the second convolution block outputs the output features of the entire shearlet embedding generator SDM.

5. The method according to claim 4, characterized in that The first downsampling module includes N serially connected residual blocks, wherein the input end of the first residual block is connected to the output end of the first convolution block, and the output end of the Nth residual block is connected to the input end of the frequency-aware downsampling module; The second downsampling module includes N serially connected residual blocks, wherein the input end of the first residual block is connected to the first shear wave downsampling module and the output end of the first downsampling module, and the output end of the Nth residual block is connected to the input end of the frequency-aware downsampling module.

6. The method according to claim 4, characterized in that The first upsampling module includes N serially connected residual blocks, wherein the input end of the first residual block is connected to the output ends of the second frequency domain bottleneck module and the second downsampling module, and the output end of the Nth residual block is connected to the input end of the frequency-aware upsampling module; The second upsampling module includes N serially connected residual blocks, wherein the input end of the first residual block is connected to the output ends of the first upsampling module and the first downsampling module, and the output end of the Nth residual block is connected to the input end of the frequency-aware upsampling module.

7. The method according to claim 6, characterized in that The frequency domain bottleneck module first divides the feature map into low-frequency subbands and high-frequency subbands through discrete shearlet transform, and then passes the low-frequency subbands as input to the residual block for processing. The processed low-frequency subbands and the original high-frequency subbands are output through inverse shearlet transform.

8. The method according to claim 1, characterized in that The LG-BPN denoising network consists of two parallel branches, one for reconstructing local and the other for reconstructing global context. The LG-BPN denoising network extracts features from the input through a 1-layer Conv, and then inputs the features into the local and global branches respectively. The local branch first uses a 9×9 DSPMC module to extract dense features through downsampling to break spatial correlation; then the feature map is processed by dilated convolution with a dilation rate of 2; The global branch first passes through a 21×21 DMPMC module with a larger receptive field, and then is processed by a DTB module; finally, the local and global information are fused to obtain the final output.

9. The method according to claim 8, characterized in that The kernel formula used by the DSPMC module is: （5）； It is the core of DSPMC. is the kernel of ordinary convolution, is a mask used to filter highly correlated pixels.

10. The method according to claim 8, characterized in that The DTB module uses channel attention to perform long-range context aggregation and make full use of global information. The attention formula used by the DTB module is as follows: （6）； （7）； + （8）； X is the input, represents element-wise multiplication, LN represents layer normalization, and Represents a 3×3 dilated convolution.

Images