Distributed fiber optic sensing speech enhancement method based on soil geology and environmental information

Abstract

The present invention relates to the technical field of distributed optical fiber acoustic sensing, and in particular to a distributed optical fiber sensing voice enhancement method based on soil geology and environmental information. The method comprises the following steps: S1, collecting spatiotemporal acoustic wave data; S2, performing a voice enhancement operation on the spatiotemporal acoustic wave data collected in step S1, and specifically comprises the following steps: S21, extracting features from the spatiotemporal acoustic wave data; S22, constructing a clue mask network based on soil geology information and soil environmental information; and S23, constructing a complex spectrum convolutional recurrent neural network, which comprises a clue mask network and a backbone network, and inputting the features extracted from the spatiotemporal acoustic wave data in step S21 into the complex spectrum convolutional recurrent neural network to obtain voice enhancement data of the spatiotemporal acoustic wave data. The present invention fully utilizes the external deployment environment information of the distributed optical fiber and better improves the performance of the distributed optical fiber sensing system.

Description

Technical Field

[0001] The present invention relates to the field of distributed optical fiber acoustic sensing technology, and in particular to a distributed optical fiber sensing voice enhancement method based on soil geology and environmental information. Background Art

[0002] Distributed fiber-optic acoustic sensing, a novel acoustic detection technology, utilizes a large area of ​​distributed optical fiber as a sensing unit. By detecting subtle bends and twists in the fiber caused by physical events in the surrounding environment, as well as phase changes in the Rayleigh backscattered light within the fiber, it achieves real-time location and perception of external vibration or sound sources. Distributed fiber-optic acoustic sensing technology boasts high sensitivity, high information richness, and multidimensional signal characteristics, and has been widely applied in perimeter security, water / oil and gas pipeline monitoring, and earthquake monitoring. However, in distributed acoustic sensing systems, the complexity of speech events and the unpredictable nature of noise mean that the collected spatiotemporal data is severely affected by random noise, resulting in poor data quality and the inability to extract valid information. Therefore, it is necessary to remove the influence of background noise from the collected spatiotemporal data (i.e., data enhancement / speech enhancement) to improve data quality and enhance the performance of distributed acoustic sensing systems.

[0003] Traditional speech enhancement methods, including Wiener filtering, spectral subtraction, and principal component analysis, perform well in steady-state noise scenarios, but their performance is severely limited in non-steady-state, random scenarios. In recent years, deep learning-based speech enhancement technology has emerged. By training neural networks on large datasets of clean and noisy audio, it can capture the nonlinear relationship between noise and clean signal characteristics, improve data quality, and achieve speech enhancement.

[0004] Prior art CN115622626A discloses a distributed acoustic wave sensing voice information recognition system and method. This system designs a distributed sensing fiber system, uses a sensing fiber to receive voice signals, uses a laser output unit to emit a narrowband laser signal, uses a circulator to detect the voice signal, and finally, receives backscattered signals through an acquisition unit and acquisition circulator. Finally, a convolutional recursive network unit performs complex domain mapping on the scattered signals to reconstruct the voice signal. However, this method focuses on the modularization of the distributed fiber system and does not fully utilize the external deployment environment information of the distributed fiber, thus failing to achieve a higher-performance distributed fiber sensing system design.

[0005] Therefore, there is an urgent need to provide a distributed fiber optic sensing voice enhancement method based on soil geology and environmental information. Compared with the existing technology, it can make full use of the external deployment environment information of the distributed fiber optic and better improve the performance of the distributed fiber optic sensing system. Summary of the Invention

[0006] The present invention solves the technical problems existing in the prior art and provides a distributed optical fiber sensing voice enhancement method based on soil geology and environmental information.

[0007] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] The distributed optical fiber sensing voice enhancement method based on soil geology and environmental information includes the following steps:

[0009] S1, collect spatiotemporal acoustic wave data;

[0010] S2, performing a speech enhancement operation on the spatiotemporal sound wave data collected in step S1, specifically comprising the following steps:

[0011] S21, extracting features from spatiotemporal sound wave data;

[0012] S22, constructing a clue mask network based on soil geological information and soil environmental information;

[0013] S23. Construct a complex spectrum convolutional recurrent neural network, which includes a clue mask network and a backbone network. Input the features extracted from the spatiotemporal sound wave data in step S21 into the complex spectrum convolutional recurrent neural network to obtain speech enhancement data of the spatiotemporal sound wave data.

[0014] Furthermore, S21 specifically includes the following steps:

[0015] S211. Performing a short-time Fourier transform on the time domain signal of the space-time acoustic wave data to obtain a real part and an imaginary part of the short-time Fourier transform;

[0016] S212, setting a Mel filter bank, passing the real part and the imaginary part obtained in step S211 through the Mel filter bank respectively to obtain a Mel short-time Fourier transform spectrum;

[0017] S213, taking the logarithm of the Mel short-time Fourier transform spectrum to obtain a logarithmic Mel time-frequency spectrum graph;

[0018] S214. Perform first-order differences and second-order differences on the logarithmic Mel-spectrogram to obtain corresponding first-order differences and second-order differences of the logarithmic Mel-spectrogram, thereby obtaining features to be extracted from the spatiotemporal sound wave data. The features to be extracted from the spatiotemporal sound wave data include the logarithmic Mel-spectrogram, the first-order differences of the logarithmic Mel-spectrogram, and the second-order differences of the logarithmic Mel-spectrogram.

[0019] Furthermore, in step S213, the logarithmic Mel-time spectrum is obtained by the following formula:

[0020]

[0021] In the above formula, M(f,t) represents the value of the Mel spectrogram at frequency f and time t, |H j (f)| represents the response value of the jth Mel filter at frequency f, |X j (t)| represents the amplitude of the frequency domain signal corresponding to the jth Mel filter at time t, n represents the total number of Mel filters, and i is 1-n.

[0022] Furthermore, S22 specifically includes the following steps:

[0023] S221. Fusing features of soil geological information and soil environmental information. The soil geological information includes soil depth and soil moisture content, and the soil environmental information includes surface vegetation conditions. The soil depth, soil moisture content, and surface vegetation conditions are binary-converted to obtain a one-hot encoding.

[0024] S222. Construct a clue mask network based on one-hot encoding.

[0025] Furthermore, the specific method of constructing the clue mask network in step S222 is: setting a Transformer encoder, setting the one-hot encoding representation as l; inputting l into the Transformer encoder, and obtaining the output E of the Transformer encoder k , E k Integrate with l to perform multiplication query and get E k ′, E k 、E k 'Use one-dimensional convolution layer to project to the decoder dimension N d 、N e , respectively get the projection code P k 、P k ′, then, P k 、P k ′ are respectively passed into the self-attention module and the cross-attention module, P k After being processed by the self-attention module and then passed through the feedforward neural network, it is input into the cross-attention module. The output of the cross-attention module is the target mask in the projection decoder space. The target mask passes through the feedforward neural network and the one-dimensional convolution layer in turn, and is projected back to the Transformer encoder dimension to obtain m k ′, and finally obtain the clue mask vector, thus obtaining the clue mask network.

[0026] Furthermore, the clue mask vector is calculated as follows:

[0027] m k =m k ′+E k ';

[0028] In the above formula, mk Represents the clue mask vector.

[0029] Furthermore, the backbone network includes an encoder, a long short-term memory network and a decoder. The encoder and the decoder are quickly linked through the long short-term memory network. In step S21, the features extracted from the spatiotemporal sound wave data are input into the encoder, and the encoder processes and outputs them. The output of the encoder is input into the clue mask network to obtain a clue mask vector, and then the clue mask vector and the clue mask network are multiplied and integrated. The result of the multiplication query integration is passed through the long short-term memory network and the decoder in sequence. The output of the decoder is the speech enhancement data of the spatiotemporal sound wave data.

[0030] Furthermore, both the encoder and the decoder include four complex two-dimensional convolutions, complex batch normalization operations, and real PReLU activation functions. The specific expression of the complex two-dimensional convolution is:

[0031]

[0032] y[i] represents the output sequence, w[k] represents the convolution kernel, r represents the void rate of the convolution kernel, K represents the size of the convolution kernel, k is 1-K, x[i-rk] represents the sequence after the input sequence is processed, and x[i] represents the input sequence.

[0033] Furthermore, the LSTM network processes the real and imaginary parts of the input separately. The processing is specifically expressed by the following formula:

[0034] F rr =LSTM r (X r );

[0035] F ir =LSTM r (X i );

[0036] F ri =LSTM i (X r );

[0037] F ii =LSTM i (X i );

[0038] F out =(F rr -F ii )+j(F ir +F ri );

[0039] In the above formula, X rrepresents the real part of the input feature of the long short-term memory network, X i Represents the imaginary part of the input feature of the long short-term memory network, LSTM r , LSTM i Respectively represent the long and short-term memory networks that process the real and imaginary parts, F out Represents the output of the long short-term memory network.

[0040] Furthermore, a loss function is set in the complex spectrum convolutional recurrent neural network to assist in speech enhancement processing. The loss function is obtained by training the optimizer by combining the noise signal-to-noise ratio and the scale-invariant signal-to-noise ratio. The loss function is expressed as:

[0041] L=L snr +λL sisnr ;

[0042]

[0043] In the above formula, L represents the loss function, L snr represents the noise signal-to-noise ratio, L sisnr represents the scale-invariant signal-to-noise ratio, λ represents the hyperparameter, S represents the real audio waveform, represents the target audio waveform output by the complex spectrum convolutional recurrent neural network, S m represents the mth segment of real audio waveform in S, express The mth target audio waveform in .

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] This method constructs auxiliary clues by integrating multi-dimensional soil geological and environmental information, including soil quality, soil moisture content, and surface vegetation, and uses a clue mask network to comprehensively generate clue mask vectors. Furthermore, during the feature extraction phase, a three-dimensional complex log-mel spectrum is selected, including the log-mel spectrum corresponding to the real and imaginary parts of the short-time Fourier transform, the first-order difference of the log-mel spectrum, and the second-order difference of the log-mel spectrum. This feature information is input into a complex convolutional recurrent neural network. Through training and testing with a large amount of distributed sensor acoustic wave data, the method achieves noise reduction and enhancement of speech signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram showing the connection relationship between the spatiotemporal data acquisition system for distributed buried optical fibers and the distributed buried optical fiber acoustic sensing voice enhancement processing system of the present invention.

[0047] Figure 2 It is a flow chart of the present invention for extracting features from spatiotemporal sound wave data.

[0048] Figure 3 It is a flowchart of the present invention for constructing a clue mask network based on one-hot encoding.

[0049] Figure 4 Schematic diagram of the complex spectrum convolutional recursive neural network of the present invention. DETAILED DESCRIPTION

[0050] The technical solution of the present invention will be clearly described below in conjunction with the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0051] The present invention provides a distributed optical fiber sensing voice enhancement method based on soil geology and environmental information, comprising the following steps:

[0052] S1. Collect spatiotemporal acoustic wave data using a spatiotemporal data acquisition system for distributed buried optical fibers. Figure 1 As shown, it includes a laser, an acousto-optic modulator, an erbium-doped fiber amplifier, a circulator, a distributed buried optical fiber, a phase demodulation unit, a photodetector and a spatiotemporal data acquisition card. The laser uses a narrowband laser, and the laser emits a light source, which enters the acousto-optic modulator and is modulated into a pulse light by the acousto-optic modulator. The pulse light passes through the erbium-doped fiber amplifier, is enhanced, and then enters the circulator. After passing through the circulator, it propagates through the distributed buried optical fiber. When a physical event occurs, that is, an acoustic wave signal is generated, the acoustic wave vibration will pass through the complex geological structure and eventually act on the distributed buried optical fiber, causing the phase of the backward Rayleigh scattered light to change. The carried phase passes through the phase modulation unit and is then received by the photodetector. An analog-to-digital converter is set, and the analog-to-digital converter converts the phase received by the photodetector into a digital signal and sends it to the spatiotemporal data acquisition card, thereby collecting spatiotemporal acoustic wave data.

[0053] S2, enhance the spatiotemporal acoustic wave data collected in step S1, specifically using Figure 1 The distributed buried optical fiber acoustic sensing speech enhancement processing system in the embodiment of the present invention performs an enhancement operation on spatiotemporal acoustic wave data; specifically, the following steps are included:

[0054] S21, such as Figure 2 As shown, feature extraction is performed on the spatiotemporal acoustic wave data collected in step S1, which helps facilitate subsequent model training. Specifically, the following steps are included:

[0055] S211, acquiring the time domain signal of the space-time acoustic wave data from the space-time data acquisition card, and performing short-time Fourier transform to obtain the real part and imaginary part of the short-time Fourier transform; the specific steps are:

[0056] The noisy spatiotemporal sound wave data is represented as y, y=x+v, where x represents the denoised spatiotemporal sound wave data and v represents the noise signal.

[0057] Performing short-time Fourier transform on both sides of the expression of the noisy space-time sound wave data yields Y=X+V, where Y, X, and V are the short-time Fourier transforms of y, x, and v, respectively.

[0058] Further expanded into the real and imaginary parts of the complex number, expressed as:

[0059] Y r +iY i =X r +iX i +V r +iV i ;

[0060] In the above formula, Y r represents the real part of Y, iY i represents the imaginary part of Y; X r represents the real part of X, iX i represents the imaginary part of X; V r represents the real part of V, iV i Represents the imaginary part of V.

[0061] S212. Set a Mel filter bank, and pass all the real and imaginary parts obtained by the short-time Fourier transform in step S211 through the set Mel filter bank to obtain a Mel short-time Fourier transform spectrum.

[0062] S213, taking the logarithm of the Mel short-time Fourier transform spectrum obtained in step S212 to obtain a logarithmic Mel time spectrum graph, which is specifically obtained by the following formula:

[0063]

[0064] In the above formula, M(f, t) represents the value of the Mel spectrogram at frequency f and time t, |H j (f)| represents the response value of the jth Mel filter at frequency f, |X j (t)| represents the amplitude of the frequency domain signal corresponding to the jth Mel filter at time t, n represents the total number of Mel filters, and i is 1-n.

[0065] S214. Perform first-order and second-order differences on the logarithmic Mel-spectrogram obtained in step S213 to obtain corresponding first-order differences of the logarithmic Mel-spectrum and second-order differences of the logarithmic Mel-spectrum, thereby obtaining the features to be extracted from the spatiotemporal sound wave data. The features to be extracted from the spatiotemporal sound wave data include the logarithmic Mel-spectrogram, the first-order differences of the logarithmic Mel-spectrum, and the second-order differences of the logarithmic Mel-spectrum.

[0066] S22, constructing a clue mask network based on soil geology and environmental information, specifically including the following steps:

[0067] S221. Fusing the features of soil geological information and soil environmental information to obtain corresponding one-hot encoding. Soil geological information includes soil depth and soil water content, and soil environmental information includes surface vegetation conditions. Soil depth, soil water content, and surface vegetation conditions can all be represented by numbers. The soil depth, soil water content, and surface vegetation conditions represented by the data are converted into binary to form a one-hot encoding. The digital representation of soil depth, soil water content, and surface vegetation conditions is shown in Table 1:

[0068] Table 1

[0069]

[0070]

[0071] A first moisture content threshold and a second moisture content threshold are set. The first moisture content threshold is smaller than the second moisture content threshold. When the moisture content of the area is smaller than the first moisture content threshold, it is a low moisture environment. When the moisture content of the area is greater than or equal to the first moisture content threshold and smaller than the second moisture content threshold, it is a medium moisture environment. When the moisture content of the area is greater than or equal to the second moisture content threshold, it is a high moisture environment. A first unit vegetation threshold and a second unit vegetation threshold are set. The first unit vegetation threshold is smaller than the second unit vegetation threshold. When the number of unit vegetation in the area is smaller than the first unit vegetation threshold, it is no vegetation. When the number of unit vegetation in the area is greater than or equal to the first unit vegetation threshold and smaller than the second unit vegetation threshold, it is medium vegetation. When the number of unit vegetation in the area is greater than the second unit vegetation threshold, it is dense vegetation.

[0072] S222, based on the one-hot encoding obtained in step S221, construct a clue mask network; the specific process is as follows Figure 3 As shown, the encoder is set, the encoder is the Transformer encoder, the unique hot encoding is input into the encoder, and the output E of the encoder is obtained. k , in Represents the dimension of the encoder, and the output E of the encoder k Perform multiplication query integration with the one-hot encoding l to obtain E k ′, that is, E k ′=E k ⊙l, E k 、E k 'Use one-dimensional convolution layer to project to the decoder dimension N d 、N e , respectively get the projection code P k 、Pk ′, then, P k 、P k ′ are respectively passed into the self-attention module and the cross-attention module, P k After being processed by the self-attention module, it passes through the feedforward neural network and inputs the value into the cross attention module to calculate the decoded representation and obtain the target mask in the projection decoder space. The target mask in the projection decoder space passes through the feedforward neural network and the one-dimensional convolution layer in turn, and is projected back to the encoder dimension to obtain m k ′, Finally, the clue mask vector is obtained, thereby obtaining the clue mask network; the clue mask vector is calculated by the following formula:

[0073] m k =m k ′+E k ';

[0074] In the above formula, m k Represents the clue mask vector.

[0075] S23, construct a complex spectrum convolutional recurrent neural network, such as Figure 4 As shown, the logarithmic Mel-spectrogram, Mel-spectrogram first-order difference, and logarithmic Mel-spectrogram second-order difference obtained in step S21 are input into the complex spectrum convolutional recurrent neural network to achieve speech enhancement of spatiotemporal sound wave data.

[0076] The complex spectrum convolutional recurrent neural network includes a clue mask network and a backbone network. The backbone network includes an encoder, a long short-term memory network and a decoder, wherein the encoder and the decoder are connected through a U-Net fast link. The first-order difference of the Mel spectrum and the second-order difference of the logarithmic Mel spectrum are input into the encoder, and the encoder processes and outputs them. The output of the encoder is input into the clue mask network to obtain a clue mask vector, and then the clue mask vector and the clue mask network are multiplied and integrated. The result of the multiplication query integration is then passed through the long short-term memory network and the decoder in sequence. The output of the decoder is the speech enhancement data of the spatiotemporal sound wave data.

[0077] The encoder consists of four complex two-dimensional convolutions, complex batch normalization operations, and real PReLU activation functions. The complex two-dimensional convolution uses a smooth transition of the void convolution. The features after the complex two-dimensional convolution are then processed by the complex batch normalization operation and the real PReLU activation function in turn. The specific expression of the complex two-dimensional convolution is:

[0078]

[0079] y[i] represents the output sequence, w[k] represents the convolution kernel, r represents the void rate of the convolution kernel, K represents the size of the convolution kernel, k is 1-K, x[i-rk] represents the sequence after the input sequence is processed, and x[i] represents the input sequence.

[0080] The long short-term memory network is used to connect the encoder and the decoder. The input of the long short-term memory network is the product of the output of the encoder and the clue mask vector output by the clue mask network. The real and imaginary parts of the input of the long short-term memory network are processed separately. The processing process is specifically expressed by the following formula:

[0081] F rr =LSTM r (X r );

[0082] F ir =LSTM r (X i );

[0083] F ri =LSTM i (X r );

[0084] F ii =LSTM i (X i );

[0085] F out =(F rr -F ii )+j(F ir +F ri );

[0086] In the above formula, X r represents the real part of the input feature of the long short-term memory network, X i Represents the imaginary part of the input feature of the long short-term memory network, LSTM r , LSTM i Respectively represent the long and short-term memory networks that process the real and imaginary parts, F out Represents the output of the long short-term memory network.

[0087] The decoder has the same structure as the encoder, including four complex two-dimensional convolutions, complex batch normalization operations, and real PReLU activation functions. The main function of the decoder is to achieve feature dimensionality reduction, while adding the features previously in the encoder accordingly, and finally obtain speech enhancement data of the spatiotemporal sound wave data.

[0088] In the complex spectrum convolutional recurrent neural network, a loss function is set to assist in speech enhancement processing. The loss function is obtained by training the optimizer by combining the noise signal-to-noise ratio and the scale-invariant signal-to-noise ratio. The optimizer adopts the Adam optimizer, and the loss function is expressed as:

[0089] L=L snr +λL sisnr ;

[0090]

[0091]

[0092] In the above formula, L represents the loss function, L snr represents the noise signal-to-noise ratio, L sisnr represents the scale-invariant signal-to-noise ratio, λ represents the hyperparameter, S represents the real audio waveform, represents the target audio waveform output by the complex spectrum convolutional recurrent neural network, S m represents the mth segment of real audio waveform in S, express The mth target audio waveform in .

[0093] This method constructs auxiliary clues by integrating multi-dimensional soil geological and environmental information, including soil quality, soil moisture content, and surface vegetation, and uses a clue mask network to comprehensively generate clue mask vectors. Furthermore, during the feature extraction phase, a three-dimensional complex log-mel spectrum is selected, including the log-mel spectrum corresponding to the real and imaginary parts of the short-time Fourier transform, the first-order difference of the log-mel spectrum, and the second-order difference of the log-mel spectrum. This feature information is input into a complex convolutional recurrent neural network. Through training and testing with a large amount of distributed sensor acoustic wave data, the method achieves noise reduction and enhancement of speech signals.

[0094] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, rather than to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions of the technical solution of the present invention by ordinary technicians in this field do not deviate from the essence and scope of the technical solution of the present invention.

Claims

1. A distributed optical fiber sensing voice enhancement method based on soil geology and environmental information, characterized in that: The following steps are involved: S1, collect spatiotemporal acoustic wave data; S2, performing a speech enhancement operation on the spatiotemporal sound wave data collected in step S1, specifically comprising the following steps: S21, extracting features from spatiotemporal sound wave data; S22, constructing a clue mask network based on soil geological information and soil environmental information; S22 specifically includes the following steps: S221. Fusing features of soil geological information and soil environmental information. The soil geological information includes soil depth and soil moisture content, and the soil environmental information includes surface vegetation conditions. The soil depth, soil moisture content, and surface vegetation conditions are binary-converted to obtain a one-hot encoding. S222. Based on the one-hot encoding, a clue mask network is constructed through the encoder; S23. Construct a complex spectrum convolutional recurrent neural network, which includes a clue mask network and a backbone network. Input the features extracted from the spatiotemporal sound wave data in step S21 into the complex spectrum convolutional recurrent neural network to obtain speech enhancement data of the spatiotemporal sound wave data.

2. The distributed optical fiber sensing voice enhancement method based on soil geology and environmental information according to claim 1 is characterized in that: S21 specifically includes the following steps: S211. Performing a short-time Fourier transform on the time domain signal of the space-time acoustic wave data to obtain a real part and an imaginary part of the short-time Fourier transform; S212, setting a Mel filter bank, passing the real part and the imaginary part obtained in step S211 through the Mel filter bank respectively, to obtain a Mel short-time Fourier transform spectrum; S213, taking the logarithm of the Mel short-time Fourier transform spectrum to obtain a logarithmic Mel time-frequency spectrum graph; S214. Perform first-order differences and second-order differences on the logarithmic Mel-spectrogram to obtain corresponding first-order differences and second-order differences of the logarithmic Mel-spectrogram, thereby obtaining features to be extracted from the spatiotemporal sound wave data. The features to be extracted from the spatiotemporal sound wave data include the logarithmic Mel-spectrogram, the first-order differences of the logarithmic Mel-spectrogram, and the second-order differences of the logarithmic Mel-spectrogram.

3. The distributed optical fiber sensing voice enhancement method based on soil geology and environmental information according to claim 2 is characterized in that: In step S213, the logarithmic Mel-time spectrum is obtained by the following formula: ； In the above formula, represents the value of the Mel spectrogram at frequency f and time t, represents the response value of the jth Mel filter at frequency f, It represents the amplitude of the frequency domain signal corresponding to the jth Mel filter at time t, n represents the total number of Mel filters, and j is 1-n.

4. The distributed optical fiber sensing voice enhancement method based on soil geology and environmental information according to claim 1 is characterized in that: The specific method of constructing the clue mask network in step S222 is: set the Transformer encoder and set the one-hot encoding representation as ;Will Input into the Transformer encoder and get the output of the Transformer encoder ,Will and Perform multiplication query integration and get ,Will 、 Use one-dimensional convolutional layers to project to the decoder dimension 、 , respectively get the projection code 、 , then, 、 Passed into the self-attention module and cross-attention module respectively, After being processed by the self-attention module and then passed through the feedforward neural network, it is input into the cross-attention module. The output of the cross-attention module is the target mask in the projection decoder space. The target mask passes through the feedforward neural network and the one-dimensional convolution layer in turn, and is projected back to the Transformer encoder dimension to obtain , and finally the clue mask vector is obtained, thus obtaining the clue mask network.

5. The distributed optical fiber sensing voice enhancement method based on soil geology and environmental information according to claim 4 is characterized in that: The clue mask vector is calculated as follows: ； In the above formula, Represents the clue mask vector.

6. The distributed optical fiber sensing voice enhancement method based on soil geology and environmental information according to claim 4 or 5, characterized in that: The backbone network includes an encoder, a long short-term memory network and a decoder. The encoder and the decoder are quickly linked through the long short-term memory network. In step S21, the features extracted from the spatiotemporal sound wave data are input into the encoder, and the encoder processes and outputs them. The output of the encoder is input into the clue mask network to obtain a clue mask vector, and then the clue mask vector and the clue mask network are subjected to multiplication query integration. The result of the multiplication query integration is sequentially passed through the long short-term memory network and the decoder. The output of the decoder is the speech enhancement data of the spatiotemporal sound wave data.

7. The distributed optical fiber sensing voice enhancement method based on soil geology and environmental information according to claim 6 is characterized in that: The encoder and the decoder both include four complex two-dimensional convolutions, complex batch normalization operations, and real PReLU activation functions. The specific expression of the complex two-dimensional convolution is: ； represents the output sequence, Represents the convolution kernel, r represents the void ratio of the convolution kernel, K represents the size of the convolution kernel, and k is 1-K. represents the sequence after the input sequence is processed, Represents the input sequence.

8. The distributed optical fiber sensing voice enhancement method based on soil geology and environmental information according to claim 6 is characterized in that: The long short-term memory network processes the real and imaginary parts of the input separately. The processing process is specifically expressed by the following formula: ； ； ； ； ； In the above formula, represents the real part of the LSTM input feature, represents the imaginary part of the long short-term memory network input feature, 、 Respectively represent the long and short-term memory networks that process the real and imaginary parts, Represents the output of the long short-term memory network.

9. The distributed optical fiber sensing voice enhancement method based on soil geology and environmental information according to claim 6, characterized in that: In the complex spectrum convolutional recurrent neural network, a loss function is set to assist in speech enhancement processing. The loss function is obtained by training the optimizer by combining the noise signal-to-noise ratio and the scale-invariant signal-to-noise ratio. The loss function is expressed as: ； ； ； ； ； In the above formula, represents the loss function, represents the signal-to-noise ratio, represents the scale-invariant signal-to-noise ratio, represents the hyperparameter, Represents the real audio waveform, represents the target audio waveform output by the complex spectrum convolutional recurrent neural network, express The mth segment of real audio waveform, express The mth target audio waveform in .

Images