Termite sound wave signal intelligent denoising method based on TFDDU-Net model

Abstract

The invention discloses a termite sound wave signal intelligent denoising method based on a TFDDU-Net model. The method comprises the following steps: collecting termite sound wave signal sample data; constructing an improved generative adversarial network; collecting environment noise, adding the environment noise into termite sound wave sample data, inputting the termite sound wave sample data into a generator, and generating a simulated termite sound wave signal data set; constructing a termite acoustic wave denoising model by adopting a TFDDU-Net network; training a termite sound wave denoising model by using the simulated termite sound wave signal data set; and the termite sound wave signal to be denoised passes through the termite sound wave denoising model to realize denoising of the termite sound wave signal. According to the termite acoustic wave generator, the Unet + + GAN architecture is adopted, so that the generator can recover low-frequency and complex termite acoustic wave signals, and the limitation of a traditional generator in signal processing is overcome. Three loss functions and a TFCAB module are designed to effectively extract and fuse time-frequency signal features, the robustness of the model to complex signals is improved, meanwhile, the denoising performance is enhanced, and it is ensured that details and structures of the signals are better reserved.

Description

Technical Field

[0001] The present invention belongs to the field of acoustic signal processing and artificial intelligence technology, and particularly relates to an intelligent denoising method for termite acoustic wave signals based on a TFDDU-Net model. Background Art

[0002] As a novel biodetection method, termite acoustic detection technology offers unique advantages in identifying and locating termite activity. The core of this technology lies in the effective extraction and analysis of acoustic signals. However, during the actual acquisition process, acoustic signals are susceptible to environmental noise, equipment interference, and signal attenuation, resulting in valid signals being drowned out by noise. In particular, the acoustic signals generated by termite activity have weak amplitudes, a wide frequency range, and randomly dispersed signal sources. This makes it difficult for traditional acoustic denoising methods to fully suppress noise while maintaining signal integrity.

[0003] In recent years, deep learning technology has made progress in acoustic signal processing fields such as speech enhancement and industrial noise detection. By constructing an end-to-end feature mapping model, deep neural networks can effectively separate signal noise in specific scenarios. However, existing deep learning denoising models still have limitations when processing weak signals from multiple sources: on the one hand, conventional network structures are unable to capture low-frequency features and have difficulty covering the core frequency band of termite sound waves; on the other hand, simple convolution operations have difficulty distinguishing useful signals with similar spectral characteristics from complex background noise. Traditional voiceprint denoising solutions have poor adaptability to termite sound waves with sparse signal sources and uneven energy distribution. After denoising, there are still problems with effective signal distortion or residual noise. Summary of the Invention

[0004] In response to the problem that termite acoustic wave signals have low frequency and small data volume, making it difficult to recover effective signals from noise, this paper establishes a termite acoustic wave denoising model and uses the generative adversarial network Unet++GAN architecture to solve the problem of insufficient termite signal samples. The Unet++ network framework replaces the traditional GAN ​​generator architecture and is based on the three loss function systems of multimodal feature information mapping: time series information loss, frequency domain information loss and time-frequency fusion information loss. It can efficiently fuse time-frequency features in the same dimension, significantly improving the denoising performance of complex termite acoustic wave signals. A time-frequency dual cross-attention module is designed to effectively extract time-frequency signal features and fuse time-frequency information, further improving the denoising performance of termite acoustic wave signals and the robustness of the denoising model.

[0005] To solve the above technical problems, the technical solution provided by the present invention is an intelligent denoising method for termite acoustic signals based on the TFDDU-Net model, comprising the following steps: Step 1: Collect termite acoustic signal sample data; Step 2: Constructing an improved generative adversarial network, wherein the improved generative adversarial network includes a generator and a discriminator; Furthermore, the generator of the improved generative adversarial network adopts the Unet++ structure.

[0006] Furthermore, the improved generative adversarial network Unet++GAN structure specifically includes: The collected termite sound wave signals Noise-containing termite sound waves are generated by randomly adding noise ; The noise includes Gaussian noise and uniform noise.

[0007] Input noisy termite sound waves to the generator , and generate false signals through nonlinear transformation , the calculation formula of the generation process is: ; (1) ; (2) Where UGen is the generator of the Unet++ structure, and Dis is the discriminator.

[0008] The generator Unet++Generator structure in Unet++GAN specifically includes: The noisy signal Input generator Unet++Generator, the calculation process is: ; (3) In formula (3), is the initial feature map, CL is the convolution layer Conv and LeakyReLU activation function; In each In the structure, a represents the level of network depth, and b represents the position of the feature map in the jump connection; The calculation process of the encoding stage is: ; (4) ; (5) Where Maxpooling is the maximum pooling operation, [·] represents the connection layer, and Up is the upsampling layer; The discriminator structure in Unet++GAN specifically includes: Input signal to the discriminator , the signal Including real signals and false signals ; Will Perform a one-dimensional convolution operation. The calculation formula of the operation process is: ; (6) ; (7) ; (8) Where, is the convolution operation, is the output after the convolution operation, BN is batch normalization, LR is the LeakyReLU activation function, and AAPL is the adaptive average pooling layer.

[0009] Step 3: Collect environmental noise, add the environmental noise to the termite sound wave sample data obtained in step 1, and input it into the generator in step 2 to generate a simulated termite sound wave signal dataset; Step 4: Use the TFDDU-Net network to build a termite sound wave denoising model. The TFDDU-Net network includes a time domain subnetwork TNet for processing time domain signals and a frequency domain subnetwork FNet for processing frequency domain signals. The time domain sub-network TNet trains the initial time domain signal, and the frequency domain sub-network FNet trains the frequency domain signal after wavelet packet transformation.

[0010] The termite sound wave denoising model includes a time domain subnetwork TNet, a frequency domain subnetwork FNet and a time-frequency dual cross attention module TFCAB.

[0011] The TFDDU-Net network structure specifically includes: Input noisy timing signal , whose tensor is B,1,L, representing B one-dimensional signals of length L; The signal is decomposed by wavelet packet transform (WPT), and one layer is decomposed using Haar wavelet basis to obtain two sub-band approximate coefficients and detail coefficients; Equation (9) is the decomposition process; ; (9) In formula (9), is the low-frequency approximation coefficient, is the high-frequency detail coefficient; The approximate coefficient and detail coefficient Splicing in the length dimension to obtain the frequency domain signal , Formula (10) is the splicing process; ; (10) In formula (10), the frequency domain signal The tensor is B,1,L; ; (11) ; (12) Formulas (11) and (12) are the input and output of the time domain sub-network TNet and the frequency domain sub-network FNet respectively; and The downsampling part in : ; (13) ; (14) ; (15) ; (16) Where, for Perform preliminary convolution to extract feature maps. for The output of each downsampled layer, For Perform preliminary convolution to extract feature maps. for The downsampled output of each layer; and The upsampling part in : ; (17) ; (18) ; (19) ; (20) ;(twenty one) ;(twenty two) ;(twenty three) ;(twenty four) In formulas (18) (20) and (22) (24), Add represents the skip connection operation. in Represents and At the same depth of the model; In formula (17) (29) (21) (23) For the above operation; In formula (17) and formula (21), and The final downsampled data is obtained by upsampling and ; In formula (18) and formula (22), the last downsampled data of the two sub-networks and the upsampled results of the data are skipped to obtain and ; In formula (19) and formula (23), the skip connection result of each layer is convolved and upsampled to obtain the upsampled result of the layer. and ; In formula (19) and formula (23), and Will correspond to Features reconstructed by layer upsampling and And the corresponding Features extracted by layer downsampling and Fusion is performed to obtain and ; and The fusion part of the two networks is as follows: ; (25) In formula (25), TFCAB is a time-frequency dual cross attention module, To downsample the output of FNet and TNet and The result is input into the TFCAB module for feature extraction and fusion; ; (26) In formula (26) To upsample the output of FNet and TNet each time and The result is input into the TFCAB module for feature extraction and fusion; ; (27) ; (28) In formula (27) and The skip connection is ; Formula (28) combines the result of each upsampling and downsampling feature extraction fusion of FNet and TNet with the result of convolution and upsampling of the previous jump connection; ; (29) ; (30) ; (31) In formulas (29), (30) and (31), the results of the last jump connection of each sub-network are convolved to obtain the final output signal; ; (32) Formula (32) contains three loss functions, They are frequency domain loss function, time domain loss function and time-frequency fusion loss function respectively.

[0012] The frequency domain loss function The specific calculation formula is: ; Where, is the frequency domain representation, is the output of FNet; is the total number of elements of the WPT domain signal; The time domain loss function The specific calculation formula is: ; Where, is the i-th sampling point of the real clean signal y, is the i-th sampling point of TNet’s output; is the weight coefficient; The time-frequency fusion loss function The specific calculation formula is: ; Step 5: Use the simulated termite acoustic signal dataset to train the termite acoustic denoising model. Add the outputs of the time-domain subnetwork TNet and the frequency-domain subnetwork FNet to the time-frequency dual cross attention module TFCAB for feature extraction and fusion. Use a predefined loss function to guide network training to obtain the trained termite acoustic denoising model. The time-frequency dual cross attention module TFCAB specifically includes: The spectrum-time attention block STAB includes an embedding layer, a dual cross attention module, a layer normalization and a convolution layer. The specific process is shown in formulas (33) and (34): ; (33) ; (34) In formula (33), , are the time domain and frequency domain signals of the input module, For the embedding layer operation, is layer normalization, For double cross attention; In formula (34), Indicates the output after passing through the DCA module, and then After the embedding layer and layer normalization operation , Skip connection, followed by normalization and convolution operations; ; (35) ; (36) ; (37) ; (38) In formula (35) It is the time-frequency dual cross attention feature map. Requires input of Formula (36), Formula (37) and Formula (38) are The calculation method is respectively based on the learned weight matrix With input Calculated.

[0013] ; (39) ; (40) Formulas (39) and (40) are the calculation methods of attention. and represents the feature output extracted by spatial self-attention, and are the transpose of the quires matrix and keys matrix of spatial self-attention, is the values ​​matrix of spatial self-attention, activation function Normalize the input vector, is the size of each vector.

[0014] Step 6: The one-dimensional termite acoustic wave signal to be denoised is passed through the trained termite acoustic wave denoising model to achieve denoising of the termite acoustic wave signal.

[0015] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention adopts the Unet++GAN network structure to effectively expand the sample size under the condition of limited sample size through generative adversarial networks. Unet++'s multi-level and multi-scale feature extraction and dense skip connections enable the generator to better recover low-frequency, small-amplitude, and complex termite acoustic signals, improving signal recovery capabilities and overcoming the limitations of traditional generators in signal processing. It effectively avoids the signal structure destruction and effective component loss problems commonly seen in traditional denoising methods, and achieves complete reconstruction of bioacoustic signals in low signal-to-noise ratio environments. (2) The present invention designs a collaborative constraint mechanism of three loss functions. Under the multiple effects of time series information loss to ensure the dynamic continuity of the signal, frequency domain information loss to enhance the spectrum feature representation, and time-frequency fusion information loss to optimize cross-domain consistency, it significantly improves the complementary and collaborative capabilities of time domain and frequency domain features, solves the noise residual problem caused by the splitting of time-frequency features, and improves the accuracy and effect of signal denoising.

[0016] (3) The present invention adopts the TFDDU-Net model based on the Transformer encoding structure and designs a time-frequency dual cross attention module TFCAB to achieve the physical separation of noise features and signal features. It can effectively extract and fuse time-frequency signal features, improve the robustness of the model to complex signals, and enhance the denoising performance, ensuring that the details and structure of the signal are better preserved. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention will be further described below with reference to the accompanying drawings and examples.

[0018] Figure 1 Schematic diagram of the flow of the intelligent denoising algorithm for termite acoustic signals based on the TFDDU-Net model in an embodiment of the present invention.

[0019] Figure 2 Schematic diagram of the Unet++GAN process according to an embodiment of the present invention.

[0020] Figure 3 This is a network structure diagram of the Unet++Generator generator according to an embodiment of the present invention.

[0021] Figure 4 This is a diagram of the Discriminator network structure of an embodiment of the present invention.

[0022] Figure 5 This is a network structure diagram of TFDDU-Net according to an embodiment of the present invention.

[0023] Figure 6 FIG. 4 is a flow chart of the TFCAB module according to an embodiment of the present invention.

[0024] Figure 7Schematic diagram of a termite plus noise wave signal according to an embodiment of the present invention.

[0025] Figure 8 This is the U-Net network denoising result of an embodiment of the present invention.

[0026] Figure 9 This is the TFDDU-Net network denoising result of an embodiment of the present invention. DETAILED DESCRIPTION

[0027] like Figure 1 As shown in FIG, an intelligent denoising algorithm for termite acoustic signals based on the TFDDU-Net model includes the following steps: Step 1: Collect termite acoustic signal sample data on site; Step 2: Build an improved generative adversarial network, which includes a generator and a discriminator; enhance the performance of the generator by integrating the Unet++ model; Using an improved generative adversarial network, multiple generators are independently trained. Data is fed into each generator to generate a large and rich dataset of simulated termite acoustic signals. Real environmental noise is collected and added to create a noise signal. The dataset is preprocessed and divided into training, validation, and test sets.

[0028] like Figure 2 As shown in the figure, the improved generative adversarial network Unet++GAN structure specifically includes: The collected termite sound wave signals Noise-containing termite sound waves are generated by randomly adding noise ; Noise includes Gaussian noise and uniform noise.

[0029] Input noisy termite sound waves to the generator , and generate false signals through a series of nonlinear transformations , the calculation formula of the generation process is: ; (1) ; (2) In the formula, UGen is the generator of Unet++ structure, and Dis is the discriminator.

[0030] The generator receives data and generates samples that are as realistic as possible. ; The discriminator receives the true signal and the false signal generated by the generator , and try to tell whether the data is real or generated by the generator, The discriminator determines the authenticity of the data. The model is based on the evaluation results Whether the network structure or training parameters of the generator and discriminator are adjusted correctly to improve the quality and diversity of generated data.

[0031] like Figure 3 As shown in Figure 1, the generator Unet++Generator structure in Unet++GAN specifically includes: The noisy signal Input generator Unet++Generator, the calculation process is: ; (3) In formula (3), is the initial feature map, CL is the convolution layer Conv and LeakyReLU activation function; In each In the structure, a represents the level of network depth, and b represents the position of the feature map in the jump connection; The calculation process of the encoding stage is: ; (4) ; (5) Where Maxpooling is the maximum pooling operation, [·] represents the connection layer, and Up is the upsampling layer; Max pooling is used between the front and back layers to compress data features, aiming to reduce the number of parameters and expand the receptive field of the model.

[0032] like Figure 4 As shown in Figure 2, the discriminator structure in Unet++GAN specifically includes: Input signal to the discriminator ,Signal Including real signals and false signals ; Will Perform a one-dimensional convolution operation. The calculation formula of the operation process is: ; (6) ; (7) ; (8) Where, is the convolution operation, The output after the convolution operation is obtained after the output is subjected to three convolution layers, nonlinear activation and batch normalization operations. , BN is batch normalization, LR is LeakyReLU activation function, and AAPL is adaptive average pooling layer.

[0033] Is a probability value that indicates the likelihood that the input data is real data. This value will be used to generate the loss function of the adversarial network and guide the training of the model through gradient feedback.

[0034] Step 3: Collect environmental noise, add the environmental noise to the termite sound wave sample data obtained in step 1, and input it into the generator in step 2 to generate a simulated termite sound wave signal dataset; Step 4: Use the TFDDU-Net network to build a termite sound wave denoising model. The TFDDU-Net network includes a time domain subnetwork TNet for processing time domain signals and a frequency domain subnetwork FNet for processing frequency domain signals. The time domain sub-network TNet trains the initial time domain signal, and the frequency domain sub-network FNet trains the frequency domain signal after wavelet packet transform.

[0035] The termite sound wave denoising model includes a time domain subnetwork TNet, a frequency domain subnetwork FNet, and a time-frequency dual cross attention module TFCAB.

[0036] like Figure 5 As shown in Figure 2, the TFDDU-Net network structure specifically includes: Input noisy timing signal , whose tensor is B,1,L, representing B one-dimensional signals of length L; The signal is decomposed by wavelet packet transform (WPT), and one layer is decomposed using Haar wavelet basis to obtain two sub-band approximate coefficients and detail coefficients; Equation (9) is the decomposition process; ; (9) In formula (9), is the low-frequency approximation coefficient, is the high-frequency detail coefficient; During the decomposition process, the timing signal After a layer of Haar wavelet decomposition, the output is a tensor of shape B,2,L / 2, where 2 represents two sub-bands and L / 2 represents the length of each sub-band. The approximate coefficients and detail coefficients in the tensor are separated and obtained. and , The approximate coefficient and detail coefficient Splicing in the length dimension to obtain the frequency domain signal , Formula (10) is the splicing process; ; (10) In formula (10), the frequency domain signal The tensor is B,1,L; ; (11) ; (12) Formulas (11) and (12) are the input and output of the time domain sub-network TNet and the frequency domain sub-network FNet respectively; and The downsampling part in : ; (13) ; (14) ; (15) ; (16) Where, for Perform preliminary convolution to extract feature maps. for The output of each downsampled layer, For Perform preliminary convolution to extract feature maps. for The downsampled output of each layer; and The upsampling part in : ; (17) ; (18) ; (19) ; (20) ;(twenty one) ;(twenty two) ;(twenty three) ;(twenty four) In formulas (18) (20) and (22) (24), Add represents the skip connection operation. in Represents and At the same depth of the model; In formula (17) (29) (21) (23) For the above operation; In formula (17) and formula (21), and The final downsampled data is obtained by upsampling and ; In formula (18) and formula (22), the last downsampled data of the two sub-networks and the upsampled results of the data are skipped to obtain and ; In formula (19) and formula (23), the skip connection result of each layer is convolved and upsampled to obtain the upsampled result of the layer. and ; In formula (19) and formula (23), and Will correspond to Features reconstructed by layer upsampling and And the corresponding Features extracted by layer downsampling and Fusion is performed to obtain and ; and The fusion part of the two networks is as follows: ; (25) In formula (25), TFCAB is a time-frequency dual cross attention module, To downsample the output of FNet and TNet and The result is input into the TFCAB module for feature extraction and fusion; ; (26) In formula (26) To upsample the output of FNet and TNet each time and The result is input into the TFCAB module for feature extraction and fusion; ; (27) ; (28) In formula (27) and The skip connection is ; Formula (28) combines the result of each upsampling and downsampling feature extraction fusion of FNet and TNet with the result of convolution and upsampling of the previous jump connection; ; (29) ; (30) ; (31) In formulas (29), (30) and (31), the results of the last jump connection of each sub-network are convolved to obtain the final output signal; ; (32) Formula (32) contains three loss functions, They are frequency domain loss function, time domain loss function and time-frequency fusion loss function respectively.

[0037] Frequency domain loss function The specific calculation formula is: ; Where, is the frequency domain representation of the real clean signal y after wavelet packet transformation and splicing, is the output of FNet, that is, the frequency domain signal reconstructed by FNet, is the total number of elements of the WPT domain signal; Time domain loss function The specific calculation formula is: ; Where, is the i-th sampling point of the real clean signal y, is the i-th sampling point of TNet’s output; is the weight coefficient used to balance the contribution of MSE and MAE, usually between 0 and 1; Time-frequency fusion loss function The specific calculation formula is: ; The termite sound wave denoising model is constructed by using To constrain ,use To constrain ,use Constrain the time-frequency fusion module to optimize the final fusion output.

[0038] Step 5: Use the simulated termite acoustic signal dataset to train the termite acoustic denoising model. Add the outputs of the time-domain subnetwork TNet and the frequency-domain subnetwork FNet to the time-frequency dual cross attention module TFCAB for feature extraction and fusion. Use a predefined loss function to guide network training to obtain the trained termite acoustic denoising model. like Figure 6 As shown in Figure 2, the time-frequency dual cross attention module TFCAB specifically includes: The spectrum-time attention block STAB includes an embedding layer, a dual cross attention module, a layer normalization and a convolution layer. The specific process is shown in formulas (33) and (34): ; (33) ; (34) In formula (33), , are the time domain and frequency domain signals of the input module, For the embedding layer operation, is layer normalization, For double cross attention; In formula (34), Indicates the output after passing through the DCA module, and then After the embedding layer and layer normalization operation , Skip connection, followed by normalization and convolution operations; ; (35) ; (36) ; (37) ; (38) In formula (35) It is the time-frequency dual cross attention feature map. Requires input of Formula (36), Formula (37) and Formula (38) are The calculation method is respectively based on the learned weight matrix With input Calculated.

[0039] ; (39) ; (40) Formulas (39) and (40) are the calculation methods of attention. and represents the feature output extracted by spatial self-attention, and are the transpose of the quires matrix and keys matrix of spatial self-attention, is the values ​​matrix of spatial self-attention, activation function Normalize the input vector, is the size of each vector.

[0040] Step 6: The one-dimensional termite acoustic wave signal to be denoised is passed through the trained termite acoustic wave denoising model to achieve denoising of the termite acoustic wave signal.

[0041] like Figure 7 、 Figure 8 、 Figure 9 As shown in the figure, the TFDDU-Net network significantly outperforms the U-Net network in denoising termite acoustic signals. This demonstrates that the network effectively enhances the feature expression capabilities of low-frequency weak signals through the multi-scale feature fusion mechanism of the Unet++GAN architecture. Thanks to the loss function design of the joint optimization of the time-frequency dual domains, the network accurately suppresses the frequency-domain noise component while preserving the continuity of the signal's time-domain waveform. In particular, through the cross-dimensional feature interaction of the time-frequency dual cross-attention module, it achieves strong robustness in extracting weak target signals in complex noise backgrounds, ultimately demonstrating significant advantages in both signal-to-noise ratio improvement and signal detail preservation.

Claims

1. An intelligent denoising method for termite acoustic signals based on the TFDDU-Net model, characterized in that: The following steps are involved: Step 1: Collect termite acoustic signal sample data; Step 2: Constructing an improved generative adversarial network, wherein the improved generative adversarial network includes a generator and a discriminator; Step 3: Collect environmental noise, add the environmental noise to the termite sound wave sample data obtained in step 1, and input it into the generator in step 2 to generate a simulated termite sound wave signal dataset; Step 4: Use the TFDDU-Net network to build a termite sound wave denoising model. The TFDDU-Net network includes a time domain subnetwork TNet, a frequency domain subnetwork FNet, and a time-frequency dual cross attention module TFCAB. Step 5: Use the simulated termite acoustic signal dataset to train the termite acoustic denoising model. Add the outputs of the time domain subnetwork TNet and the frequency domain subnetwork FNet to the time-frequency dual cross attention module TFCAB for feature extraction and fusion. Use a predefined loss function to guide network training to obtain the trained termite acoustic denoising model. Step 6: The termite acoustic wave signal to be denoised is passed through the trained termite acoustic wave denoising model to achieve denoising of the termite acoustic wave signal.

2. The intelligent denoising method for termite acoustic signals based on the TFDDU-Net model according to claim 1 is characterized in that: In step 2, the generator of the improved generative adversarial network adopts the Unet++ structure.

3. The intelligent denoising method for termite acoustic signals based on the TFDDU-Net model according to claim 2 is characterized in that: The improved generative adversarial network Unet++GAN structure specifically includes: The collected termite sound wave signals Noise-containing termite sound waves are generated by randomly adding noise ; Input noisy termite sound waves to the generator , and generate false signals through nonlinear transformation , the calculation formula of the generation process is: ；（1） ；（2） Where UGen is the generator of the Unet++ structure, and Dis is the discriminator.

4. The intelligent denoising method for termite acoustic signals based on the TFDDU-Net model according to claim 3 is characterized in that: The noise includes Gaussian noise and uniform noise.

5. The intelligent denoising method for termite acoustic signals based on the TFDDU-Net model according to claim 4 is characterized in that: In step 2, the generator Unet++Generator structure in Unet++GAN specifically includes: The noisy signal Input generator Unet++Generator, the calculation process is: ；（3） In formula (3), is the initial feature map, CL is the convolution layer Conv and LeakyReLU activation function; In each In the structure, a represents the level of network depth, and b represents the position of the feature map in the jump connection; The calculation process of the encoding stage is: ；（4） ； （5） In the formula, Maxpooling is the maximum pooling operation, [·] represents the connection layer, and Up is the upsampling layer.

6. The intelligent denoising method for termite acoustic signals based on the TFDDU-Net model according to claim 5 is characterized in that: In step 2, the discriminator structure in Unet++GAN specifically includes: Input signal to the discriminator , the signal Including real signals and false signals ; Will Perform convolution operation, the calculation formula of the operation process is: ；（6） ；（7） ；（8） Where, is the convolution operation, is the output after the convolution operation, BN is batch normalization, LR is the LeakyReLU activation function, and AAPL is the adaptive average pooling layer.

7. The intelligent denoising method for termite acoustic signals based on the TFDDU-Net model according to claim 6 is characterized in that: In step 4, the time domain sub-network TNet is used to process time domain signals, and the frequency domain sub-network FNet is used to process frequency domain signals; The time domain sub-network TNet trains the initial time domain signal, and the frequency domain sub-network FNet trains the frequency domain signal after wavelet packet transformation.

8. The intelligent denoising method for termite acoustic signals based on the TFDDU-Net model according to claim 7 is characterized in that: In step 4, the TFDDU-Net network structure specifically includes: Input noisy timing signal , whose tensor is B,1,L, representing B one-dimensional signals of length L; The signal is decomposed by wavelet packet transform (WPT), and one layer is decomposed using Haar wavelet basis to obtain two sub-band approximate coefficients and detail coefficients; Equation (9) is the decomposition process; ；（9） In formula (9), is the low-frequency approximation coefficient, is the high-frequency detail coefficient; The approximate coefficient and detail coefficient Splicing in the length dimension to obtain the frequency domain signal , Formula (10) is the splicing process; ；（10） In formula (10), the frequency domain signal The tensor is B,1,L; ；（11） ；（12） Formulas (11) and (12) are the input and output of the time domain sub-network TNet and the frequency domain sub-network FNet respectively; and The downsampling part in : ；（13） ；（14） ；（15） ；（16） Where, for Perform preliminary convolution to extract feature maps. for The output of each downsampled layer, For Perform preliminary convolution to extract feature maps. for The downsampled output of each layer; and The upsampling part in : ；（17） ；（18） ；（19） ；（20） ；（21） ；（22） ；（23） ；（24） In formulas (18) (20) and (22) (24), Add represents the skip connection operation. in Represents and At the same depth of the model; In formula (17) (29) (21) (23) For the above operation; In formula (17) and formula (21), and The final downsampled data is obtained by upsampling and ; In formula (18) and formula (22), the last downsampled data of the two sub-networks and the upsampled results of the data are skipped to obtain and ; In formula (19) and formula (23), the skip connection result of each layer is convolved and upsampled to obtain the upsampled result of the layer. and ; In formula (19) and formula (23), and Will correspond to Features reconstructed by layer upsampling and And the corresponding Features extracted by layer downsampling and Fusion is performed to obtain and ; and The fusion part of the two networks is as follows: ；（25） In formula (25), TFCAB is a time-frequency dual cross attention module, To downsample the output of FNet and TNet and The result is input into the TFCAB module for feature extraction and fusion; ；（26） In formula (26) To upsample the output of FNet and TNet each time and The result is input into the TFCAB module for feature extraction and fusion; ；（27） ；（28） In formula (27) and The skip connection is ; Formula (28) combines the result of each upsampling and downsampling feature extraction fusion of FNet and TNet with the result of convolution and upsampling of the previous jump connection; ；（29） ；（30） ；（31） In formulas (29), (30) and (31), the results of the last jump connection of each sub-network are convolved to obtain the final output signal; ；（32） Formula (32) contains three loss functions They are frequency domain loss function, time domain loss function and time-frequency fusion loss function respectively.

9. The intelligent denoising method for termite acoustic signals based on the TFDDU-Net model according to claim 8 is characterized in that: The frequency domain loss function The specific calculation formula is: ； Where, is the frequency domain representation, is the output of FNet; is the total number of elements of the WPT domain signal; The time domain loss function The specific calculation formula is: ； Where, is the i-th sampling point of the real clean signal y, is the i-th sampling point of TNet’s output; is the weight coefficient; The time-frequency fusion loss function The specific calculation formula is: ； Where, is the i-th sampling point of the Fusion output.

10. The intelligent denoising method for termite acoustic signals based on the TFDDU-Net model according to claim 9 is characterized in that: In step 5, the time-frequency dual cross attention module TFCAB specifically includes: The spectrum-temporal attention block STAB includes an embedding layer, a dual cross attention module, a layer normalization, and a convolutional layer, as shown in formulas (33) and (34): ；（33） ；（34） In formula (33), , are the time domain and frequency domain signals of the input module, For the embedding layer operation, is layer normalization, For double cross attention; In formula (34), is the output after passing through the DCA module, and then After the embedding layer and layer normalization operation , Skip connection, followed by normalization and convolution operations; ；（35） ；（36） ；（37） ；（38） In formula (35) It is the time-frequency dual cross attention feature map; Formula (36), Formula (37) and Formula (38) are The calculation method is respectively based on the learned weight matrix With input Calculated; ；（39） ；（40） Formulas (39) and (40) are the calculation methods of attention. and represents the feature output extracted by spatial self-attention, and are the transpose of the quires matrix and keys matrix of spatial self-attention, is the values ​​matrix of spatial self-attention, activation function Normalize the input vector, is the size of each vector.

Images