A method and system for reducing noise of gas leakage sound signal based on hourglass model
Through the neural network method based on the hourglass model, the gas leakage acoustic signals are characterized by encoding and separation, and combined with local and global modeling technology, the problem of insufficient signal noise reduction capability in the existing technology is solved, achieving better noise reduction effect and signal recognition accuracy.
Patent Information
- Application Number
- CN202310546193.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-15
- Publication Date
- 2025-03-21
- Estimated Expiration
- 2043-05-15
AI Technical Summary
The existing gas leakage acoustic signal noise reduction method has the problem of poor signal noise reduction capability and insufficient modeling capability of feature sequences.
The neural network method based on the hourglass model is adopted to feature encoding and separation of the acoustic signals through one-dimensional convolution layer and one-dimensional deconvolution layer. Combined with local modeling and global modeling technology, the target signal and background noise signal are separated.
It achieves better noise reduction effect, retains the phase information and amplitude information of the signal, improves the signal-to-noise ratio, and improves the accuracy of subsequent identification networks.
Smart Images
Figure CN116825120B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of acoustic signal fault diagnosis, and in particular to a method and system for reducing noise of a gas leakage acoustic signal based on an hourglass model. Background Art
[0002] During the transportation of gas in chemical pipelines, leakage may occur, which will affect the production efficiency of the chemical industry and cause huge economic losses. Therefore, it is necessary to monitor gas leakage in real time. Gas leakage detection usually requires feature extraction of gas leakage acoustic signals first, and then determine whether there is a gas leakage. Therefore, it is necessary to study reliable acoustic signal noise reduction algorithms to improve detection accuracy.
[0003] The existing acoustic signal denoising algorithms are divided into mask-based time-frequency domain denoising methods and original waveform-based time-domain denoising methods.
[0004] The mask-based time-frequency domain denoising method first needs to extract time-frequency domain features from the mixed sound signal, then perform feature separation to obtain the time-frequency features of different target signals, and finally reconstruct the sound signal through short-time Fourier inverse transform. Since this method extracts time-frequency features through short-time Fourier transform, the phase information is lost, and the signal is modeled and separated only through amplitude estimation, the denoising effect is poor compared to the time-domain denoising method. In addition, the denoising performance of the mask-based time-frequency domain denoising method mainly depends on the accuracy of the mask estimation. If the mask estimation is inaccurate, the separated target signal will still be mixed with background noise, which will have an adverse effect on subsequent detection.
[0005] The time-domain denoising method based on the original waveform directly denoises the mixed sound signal, and uses encoders and decoders instead of Fourier transform and inverse Fourier transform to extract the sound signal features and separate the target signal and background noise, which can avoid the problems of phase information loss and amplitude decoupling caused by short-time Fourier transform. Compared with the time-frequency domain denoising method, this method has a faster denoising speed and better denoising effect, but the design of the separation module is more complicated, and the modeling capabilities of the existing local modeling module and global modeling module still have room for improvement. Summary of the invention
[0006] In order to solve the problems of poor signal noise reduction capability and insufficient feature sequence modeling capability in existing gas leakage noise reduction methods, the present invention provides a gas leakage acoustic signal noise reduction method and system based on an hourglass model.
[0007] A method for reducing noise of gas leakage sound signal based on an hourglass model, comprising:
[0008] S1: collecting gas leakage sound signals in real time in a non-contact manner to construct a training set and a test set, wherein the gas leakage sound signals include a target signal and background noise;
[0009] S2: using a neural network model to establish an hourglass model, training and testing the hourglass model through a training set and a test set to obtain a final hourglass model, wherein the final hourglass model includes a one-dimensional convolution layer and a one-dimensional deconvolution layer, and performing feature encoding and segmentation on the acoustic signal through the one-dimensional convolution layer to obtain a coded feature sequence;
[0010] S3: Separate the coding feature sequence to obtain the feature sequence of the target signal and the feature sequence of the background noise signal;
[0011] S4: Finally, the characteristic sequence of the target signal and the characteristic sequence of the background noise signal obtained in step S3 are restored and decoded through a one-dimensional deconvolution layer, and the final characteristic sequence of the target signal and the characteristic sequence of the background noise signal are output, thereby achieving noise reduction of the gas leakage sound signal appearing in the actual working environment.
[0012] Furthermore, in step S2, the process of feature encoding and segmentation is as follows: the acoustic signal is encoded using a one-dimensional convolutional layer and activated using a ReLU activation function to obtain a coded feature sequence, and then the feature sequence is segmented to facilitate subsequent feature separation.
[0013] Further, in step S3, local modeling and global modeling are used to separate the coding feature sequence;
[0014] In this step, the dilated causal convolution with dilation coefficients of 1, 2, and 4 is used to model the local information in the encoded feature sequence to obtain the feature sequence after local modeling. The calculation process of local modeling can be expressed as:
[0015]
[0016] In the formula, F(x) represents the feature sequence output after the expansion causal convolution calculation, f(i) is the convolution kernel, and x k-d·i is the feature sequence in the convolution window corresponding to f(i), d represents the expansion coefficient, i represents the position of the convolution window corresponding to f(i), and k is the size of the one-dimensional convolution kernel;
[0017] By modeling the global information between the encoded feature sequences, the feature sequence after global modeling is obtained. The specific process of global modeling is:
[0018] (1) First, two one-dimensional average pooling operations are performed to extract the features in the vertical and horizontal directions respectively, and then the two feature vectors with spatial perception information are concatenated.
[0019] (2) Then, according to the height H and width W of the input feature vector, the pooling kernel size is set to (H, 1) and (1, W);
[0020] (3) The information of different channels is then encoded along the vertical and horizontal directions and decomposed into weight matrices in the horizontal and vertical directions;
[0021] (4) Finally, the input local modeling feature sequence is weighted to obtain a feature sequence containing global information.
[0022] Furthermore, in step (1), the calculation process of the feature vector is:
[0023]
[0024] In the formula, x c is the input feature vector, c is the number of channels of the input feature vector, [,] is the concatenation operation of the feature vector, f is the feature map containing spatial information encoded along the X-axis and Y-axis, F1 is a nonlinear transformation, δ is a nonlinear activation function, H and W represent the height H and width W of the feature vector respectively, i represents the convolution window position corresponding to f in the X-axis direction, j represents the convolution window position corresponding to f in the Y-axis direction, h represents the height of the corresponding convolution window, and w represents the width of the corresponding convolution window.
[0025] Furthermore, in step (3), the calculation process of the weight matrix in the horizontal and vertical directions is:
[0026]
[0027]
[0028] In the formula, F h is a 1×1 convolution operation in the horizontal direction, F w is a 1×1 convolution operation in the vertical direction, δ is a nonlinear activation function, σ is a Sigmoid activation function, is the weighted output feature in the vertical direction, is the weighted output feature in the horizontal direction, and are two independent tensors after the feature map f is decomposed.
[0029] A gas leakage acoustic signal denoising system based on an hourglass model mainly comprises: an acquisition module, a feature encoding module, a feature separation module and a feature decoding module;
[0030] An acquisition module is used to acquire gas leakage sound signals in real time in a non-contact manner to construct a training set and a test set, wherein the gas leakage sound signals include a target signal and background noise;
[0031] The feature encoding module is used to encode the original sound signal through a one-dimensional convolution layer and use R eLU activation function is used to activate the image, and the coded feature sequence is obtained, and the coded feature sequence is segmented and input into the feature separation module and the feature decoding module; in this module, a neural network model is used to establish an hourglass model, and the hourglass model is trained and tested through a training set and a test set to obtain a final hourglass model, which includes a one-dimensional convolution layer and a one-dimensional deconvolution layer;
[0032] A feature separation module is used to separate the segmented coding feature sequence to obtain a feature sequence of the target signal and a feature sequence of the background noise signal;
[0033] The feature decoding module is used to restore and decode the feature sequence of the target signal and the feature sequence of the background noise signal through a one-dimensional deconvolution layer, and output the final feature sequence of the target signal and the feature sequence of the background noise signal, thereby achieving noise reduction of the gas leakage sound signal that occurs in the actual working environment.
[0034] Furthermore, in the feature encoding module, the process of feature encoding and segmentation is as follows: the acoustic signal is encoded using a one-dimensional convolutional layer and activated using a ReLU activation function to obtain a coded feature sequence, and then the feature sequence is segmented to facilitate subsequent feature separation.
[0035] Furthermore, in the feature separation module, local modeling and global modeling are used to separate the coding feature sequences;
[0036] This module uses dilated causal convolution with dilation coefficients of 1, 2, and 4 to model the local information in the encoded feature sequence and obtain the feature sequence after local modeling. The calculation process of local modeling can be expressed as:
[0037]
[0038] In the formula, F(x) represents the feature sequence output after the expansion causal convolution calculation, f(i) is the convolution kernel, and x k-d·i is the feature sequence in the convolution window corresponding to f(i), d represents the expansion coefficient, i represents the position of the convolution window corresponding to f(i), and k is the size of the one-dimensional convolution kernel;
[0039] By modeling the global information between the encoded feature sequences, the feature sequence after global modeling is obtained. The specific process of global modeling is:
[0040] (1) First, two one-dimensional average pooling operations are performed to extract the features in the vertical and horizontal directions respectively, and then the two feature vectors with spatial perception information are concatenated.
[0041] (2) Then, according to the height H and width W of the input feature vector, the pooling kernel size is set to (H, 1) and (1, W);
[0042] (3) The information of different channels is then encoded along the vertical and horizontal directions and decomposed into weight matrices in the horizontal and vertical directions;
[0043] (4) Finally, the input local modeling feature sequence is weighted to obtain a feature sequence containing global information.
[0044] Furthermore, in step (1), the calculation process of the feature vector is:
[0045]
[0046] In the formula, x c is the input feature vector, c is the number of channels of the input feature vector, [,] is the concatenation operation of the feature vector, f is the feature map containing spatial information encoded along the X-axis and Y-axis, F1 is a nonlinear transformation, δ is a nonlinear activation function, H and W represent the height H and width W of the feature vector respectively, i represents the convolution window corresponding to f in the X-axis direction, j represents the convolution window corresponding to f in the Y-axis direction, h represents the height of the corresponding convolution window, and w represents the width of the corresponding convolution window.
[0047] Furthermore, in step (3), the calculation process of the weight matrix in the horizontal and vertical directions is:
[0048]
[0049]
[0050] In the formula, F h is a 1×1 convolution operation in the horizontal direction, F w is a 1×1 convolution operation in the vertical direction, δ is a nonlinear activation function, σ is a Sigmoid activation function, is the weighted output feature in the vertical direction, is the weighted output feature in the horizontal direction, and are two independent tensors after the feature map f is decomposed.
[0051] The beneficial effects brought by the technical solution provided by the present invention are:
[0052] 1. The present invention uses a one-dimensional convolutional neural network instead of a short-time Fourier transform to extract the time-frequency domain features of the signal to obtain a coding feature sequence, which can simultaneously retain phase information and amplitude information, improve the feature expression ability of the coding sequence, and thus obtain a better noise reduction effect; then, a local modeling module and a global modeling module are used to locally and globally model the signal, and the use of dilated causal convolution for local modeling can better extract the contextual information of the time series and improve the modeling ability of the sequence within the segment; finally, the modeled feature sequence is decoded to separate the target signal, improve the signal-to-noise ratio, and improve the accuracy of the subsequent recognition network.
[0053] 2. The gas leakage acoustic signal denoising method provided by the present invention does not require preprocessing, occupies less memory and has a short inference time. By mixing background noises with different signal-to-noise ratios, the hourglass model can achieve good denoising effects when denoising weak leakage signals. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:
[0055] Figure 1 It is a flow chart of a method for reducing noise of a gas leakage acoustic signal based on an hourglass model in an embodiment of the present invention.
[0056] Figure 2 It is a structural diagram of the hourglass model in an embodiment of the present invention.
[0057] Figure 3 Schematic diagram of coding sequence segmentation in an embodiment of the present invention.
[0058] Figure 4 Schematic diagram of sequence global modeling in an embodiment of the present invention. DETAILED DESCRIPTION
[0059] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.
[0060] The embodiment of the present invention provides a gas leakage acoustic signal denoising method and system based on an hourglass model. The overall idea is to collect gas leakage acoustic signals in real time in a non-contact manner, use a one-dimensional convolutional neural network to feature encode and segment the acoustic signals, then perform local and global modeling on the encoded feature sequence to separate the target signal and background noise features, and finally restore the target signal and background noise signal through deconvolution, thereby achieving denoising of the gas leakage acoustic signal.
[0061] Please refer to Figure 1-2 , Figure 1 is a flow chart of a method for reducing noise of a gas leakage acoustic signal based on an hourglass model in an embodiment of the present invention. Figure 2is a structural diagram of an hourglass model in an embodiment of the present invention. The gas leakage sound signal denoising method based on the hourglass model comprises the following steps:
[0062] S1: The gas leakage acoustic signal is collected in real time in a non-contact manner. The acoustic signal is a type of non-contact collection signal. The gas leakage acoustic signal is collected by a microphone. The specific process is to collect the pure acoustic signal (i.e., the target signal) and the background noise signal generated when the gas is leaking in a non-contact manner, mix the pure acoustic signal with the background noise signal at different signal-to-noise ratios to obtain the gas leakage acoustic signal containing background noise, and use the gas leakage acoustic signals containing different background noises to construct a training set and a test set;
[0063] S2: using a neural network model to establish an hourglass model to achieve noise reduction of the gas leakage acoustic signal; training and testing the hourglass model through a training set and a test set to obtain a final hourglass model, the final hourglass model includes a one-dimensional convolution layer and a one-dimensional deconvolution, and the acoustic signal is feature encoded and segmented through the one-dimensional convolution layer; this step is specifically as follows:
[0064] The acoustic signal is encoded using a one-dimensional convolutional layer and activated using a ReLU activation function to obtain a coded feature sequence, which is then segmented to facilitate subsequent feature separation. The segmentation process is as follows: Figure 3 shown.
[0065] The feature encoding in this step inherits the idea of short-time Fourier transform. A one-dimensional convolutional neural network is used instead of short-time Fourier transform to extract the time-frequency features of the acoustic signal. Different transformation windows are simulated by changing the convolution kernel size and step size to select and extract feature sequences with different time-frequency resolutions according to the signal characteristics. Both phase information and amplitude information can be retained at the same time, thereby improving the feature expression capability of the coding sequence.
[0066] S3: Separate the coding feature sequence through local modeling (TCN) and global modeling (CA) to obtain the feature sequence of the target signal and the feature sequence of the background noise signal;
[0067] This step uses dilated causal convolution with dilation coefficients of 1, 2, and 4 to model the local information in the encoded feature sequence and obtain the feature sequence after local modeling. The calculation process of local modeling can be expressed as:
[0068]
[0069] In the formula, F(x) represents the feature sequence output after the expansion causal convolution calculation, f(i) is the convolution kernel, and x s-d·iThe feature sequence in the corresponding convolution window, d represents the expansion coefficient, and the convolution calculation is performed at different intervals on different convolution layers. k is the size of the one-dimensional convolution kernel. The dilated causal convolution can obtain a larger receptive field by expanding the sampling interval, which is determined by the expansion coefficient d.
[0070] By modeling the global information between the encoded feature sequences, the feature sequence after global modeling (CA) is obtained. The modeling process is as follows: Figure 4 As shown in Figure 2, the specific process of global modeling is as follows:
[0071] (1) First, two one-dimensional average pooling operations are performed to extract the features of global information in the vertical and horizontal directions respectively, and then the features are concatenated into two pairs of feature vectors with spatial perception information;
[0072] The calculation process of the eigenvector is:
[0073]
[0074] In the formula, x c is the input feature vector, c is the number of channels of the input feature vector, [,] is the concatenation operation of the feature vector, f is the feature map containing spatial information encoded along the X-axis and Y-axis, F1 is the nonlinear transformation, and δ is the nonlinear activation function;
[0075] (2) Then, according to the height H and width W of the input feature vector, the pooling kernel size is set to (H, 1) and (1, W);
[0076] (3) The information of different channels is then encoded along the vertical and horizontal directions and decomposed into weight matrices in the horizontal and vertical directions;
[0077] Among them, the calculation process of the weight matrix in the horizontal and vertical directions is:
[0078]
[0079]
[0080] In the formula, F h is a 1×1 convolution operation in the horizontal direction, F w is a 1×1 convolution operation in the vertical direction, δ is a nonlinear activation function, σ is a Sigmoid activation function, is the weighted output feature in the vertical direction, is the weighted output feature in the horizontal direction, and are two independent tensors after the feature map f is decomposed;
[0081] (4) Finally, the input local feature sequence is weighted according to the weight matrices in the horizontal and vertical directions to obtain a feature sequence containing global information.
[0082] S4: Restore and decode the characteristic sequence of the target signal and the characteristic sequence of the background noise signal in step S3 through a one-dimensional deconvolution layer operation, and output the final characteristic sequence of the target signal and the characteristic sequence of the background noise signal, thereby achieving noise reduction of the gas leakage sound signal appearing in the actual working environment.
[0083] The present invention separates the gas leakage sound signal with background noise into a pure gas leakage sound signal and a background noise signal. The target signal is the pure gas leakage sound signal. Compared with the gas leakage sound signal with background noise, the noise reduction process has been achieved.
[0084] A gas leakage acoustic signal denoising system based on an hourglass model, specifically comprising: an acquisition module, a feature encoding module, a feature separation module and a feature decoding module;
[0085] An acquisition module is used to acquire gas leakage sound signals in real time in a non-contact manner to construct a training set and a test set, wherein the gas leakage sound signals include a target signal and background noise;
[0086] The feature encoding module is used to encode the original gas leakage sound signal into a feature sequence through a one-dimensional convolution layer, and activate it using a ReLU activation function to obtain a coded feature sequence, and then segment the coded feature sequence, and then input it into a subsequent feature separation module and a feature decoding module; in this module, a neural network model is used to establish an hourglass model, and the hourglass model is trained and tested through a training set and a test set to obtain a final hourglass model, which includes a one-dimensional convolution layer and a one-dimensional deconvolution layer;
[0087] A feature separation module is used to separate the segmented coding feature sequence to obtain a feature sequence of the target signal and a feature sequence of the background noise signal;
[0088] Among them, the feature separation module includes 6 groups of local modeling modules and a global modeling module. Figure 2 It is represented by SandglassetBlock, which is responsible for modeling the local features and global features of the feature sequence respectively.
[0089] The feature decoding module is mainly used to restore and decode the feature sequences that have been locally modeled and globally modeled, and output the feature sequences of the final target signal and the feature sequences of the background noise signal, thereby achieving noise reduction of the gas leakage sound signal that occurs in the actual working environment.
[0090] Among them, feature sequence restoration and feature sequence segmentation are symmetric processes.
[0091] The method and system provided by the present invention can reduce the noise of gas leakage signals occurring in different working environments, improve the signal-to-noise ratio, and reduce the difficulty of subsequent gas leakage detection.
[0092] The method and system disclosed in the present invention can perform transfer learning and retrain the hourglass model after changing different application scenarios to adapt to changes in different application scenarios, so that the model can cover different application scenarios and improve the generalization of the denoising model.
[0093] The beneficial effects of the present invention are:
[0094] 1. The present invention uses a one-dimensional convolutional neural network instead of a short-time Fourier transform to extract the time-frequency domain features of the signal to obtain a coding feature sequence, which can simultaneously retain phase information and amplitude information, improve the feature expression ability of the coding sequence, and thus obtain a better noise reduction effect; then, a local modeling module and a global modeling module are used to locally and globally model the signal, and the use of an expanded causal convolution for local modeling can better extract the contextual information of the time series and improve the modeling ability of the sequence within the segment; finally, the modeled feature sequence is decoded to separate the target signal (i.e., a pure gas leakage sound signal), that is, the background noise component is reduced, the signal-to-noise ratio (the signal-to-noise ratio refers to the target signal / background noise signal) is improved, and the accuracy of the subsequent recognition network is improved.
[0095] 2. The gas leakage acoustic signal denoising method provided by the present invention does not require preprocessing, occupies less memory and has a short inference time. By mixing background noises with different signal-to-noise ratios, the hourglass model can achieve good denoising effects when denoising weak leakage signals.
[0096] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for reducing noise of gas leakage sound signal based on hourglass model, characterized in that: include: S1: collecting gas leakage sound signals in real time in a non-contact manner to construct a training set and a test set, wherein the gas leakage sound signals include a target signal and background noise; S2: using a neural network model to establish an hourglass model, training and testing the hourglass model through a training set and a test set to obtain a final hourglass model, wherein the final hourglass model includes a one-dimensional convolution layer and a one-dimensional deconvolution layer, and performing feature encoding and segmentation on the acoustic signal through the one-dimensional convolution layer to obtain a coded feature sequence; S3: Separate the coding feature sequence to obtain the feature sequence of the target signal and the feature sequence of the background noise signal; In step S3, local modeling and global modeling are used to separate the coding feature sequences; In this step, the dilated causal convolution with dilation coefficients of 1, 2, and 4 is used to model the local information in the encoded feature sequence to obtain the feature sequence after local modeling. The calculation process of local modeling can be expressed as: In the formula, F(x) represents the feature sequence output after the expansion causal convolution calculation, f(i) is the convolution kernel, and x k-d·i is the feature sequence in the convolution window corresponding to f(i), d represents the expansion coefficient, i represents the position of the convolution window corresponding to f(i), and k is the size of the one-dimensional convolution kernel; By modeling the global information between the encoded feature sequences, the feature sequence after global modeling is obtained. The specific process of global modeling is: (1) First, two one-dimensional average pooling operations are performed to extract the features in the vertical and horizontal directions respectively, and then the two feature vectors with spatial perception information are concatenated. (2) Then, according to the height H and width W of the input feature vector, the pooling kernel size is set to (H, 1) and (1, W); (3) The information of different channels is then encoded along the vertical and horizontal directions and decomposed into weight matrices in the horizontal and vertical directions; (4) Finally, the input local modeling feature sequence is weighted to obtain a feature sequence containing global information; S4: Finally, the characteristic sequence of the target signal and the characteristic sequence of the background noise signal obtained in step S3 are restored and decoded through a one-dimensional deconvolution layer, and the final characteristic sequence of the target signal and the characteristic sequence of the background noise signal are output, thereby achieving noise reduction of the gas leakage sound signal appearing in the actual working environment.
2. A method for reducing noise of gas leakage sound signals based on an hourglass model as claimed in claim 1, characterized in that: In step S2, the process of feature encoding and segmentation is as follows: the acoustic signal is encoded using a one-dimensional convolutional layer and activated using a ReLU activation function to obtain a coded feature sequence, and then the feature sequence is segmented to facilitate subsequent feature separation.
3. The method for reducing noise of gas leakage sound signal based on hourglass model according to claim 1, characterized in that: In step (1), the calculation process of the feature vector is: In the formula, x c is the input feature vector, c is the number of channels of the input feature vector, [,] is the concatenation operation of the feature vector, f is the feature map containing spatial information encoded along the X-axis and Y-axis, F1 is a nonlinear transformation, δ is a nonlinear activation function, H and W represent the height H and width W of the feature vector respectively, i represents the convolution window position corresponding to f in the X-axis direction, j represents the convolution window position corresponding to f in the Y-axis direction, h represents the height of the corresponding convolution window, and w represents the width of the corresponding convolution window.
4. The method for reducing noise of gas leakage sound signal based on hourglass model according to claim 1, characterized in that: In step (3), the calculation process of the weight matrix in the horizontal and vertical directions is: In the formula, F h is a 1×1 convolution operation in the horizontal direction, F w is a 1×1 convolution operation in the vertical direction, δ is a nonlinear activation function, σ is a Sigmoid activation function, is the weighted output feature in the vertical direction, is the weighted output feature in the horizontal direction, and are two independent tensors after the feature map f is decomposed.
5. A gas leakage acoustic signal denoising system based on an hourglass model, characterized in that: include: Acquisition module, feature encoding module, feature separation module and feature decoding module; An acquisition module is used to acquire gas leakage sound signals in real time in a non-contact manner to construct a training set and a test set, wherein the gas leakage sound signals include a target signal and background noise; The feature encoding module is used to encode the original sound signal through a one-dimensional convolution layer and activate it using a ReLU activation function to obtain a coded feature sequence, and then divide the coded feature sequence and input it into a feature separation module and a feature decoding module; in this module, a neural network model is used to establish an hourglass model, and the hourglass model is trained and tested through a training set and a test set to obtain a final hourglass model, which includes a one-dimensional convolution layer and a one-dimensional deconvolution layer; A feature separation module is used to separate the segmented coding feature sequence to obtain a feature sequence of the target signal and a feature sequence of the background noise signal; In the feature separation module, local modeling and global modeling are used to separate the coding feature sequence; This module uses dilated causal convolution with dilation coefficients of 1, 2, and 4 to model the local information in the encoded feature sequence and obtain the feature sequence after local modeling. The calculation process of local modeling can be expressed as: In the formula, F(x) represents the feature sequence output after the expansion causal convolution calculation, f(i) is the convolution kernel, and x k-d·i is the feature sequence in the convolution window corresponding to f(i), d represents the expansion coefficient, i represents the position of the convolution window corresponding to f(i), and k is the size of the one-dimensional convolution kernel; By modeling the global information between the encoded feature sequences, the feature sequence after global modeling is obtained. The specific process of global modeling is: (1) First, two one-dimensional average pooling operations are performed to extract the features in the vertical and horizontal directions respectively, and then the two feature vectors with spatial perception information are concatenated. (2) Then, according to the height H and width W of the input feature vector, the pooling kernel size is set to (H, 1) and (1, W); (3) The information of different channels is then encoded along the vertical and horizontal directions and decomposed into weight matrices in the horizontal and vertical directions; (4) Finally, the input local modeling feature sequence is weighted to obtain a feature sequence containing global information; The feature decoding module is used to restore and decode the feature sequence of the target signal and the feature sequence of the background noise signal through a one-dimensional deconvolution layer, and output the final feature sequence of the target signal and the feature sequence of the background noise signal, thereby achieving noise reduction of the gas leakage sound signal that appears in the actual working environment.
6. A gas leakage sound signal denoising system based on an hourglass model as claimed in claim 5, characterized in that: In the feature encoding module, the feature encoding and segmentation process is as follows: use a one-dimensional convolutional layer to encode the acoustic signal and use the ReLU activation function to activate it to obtain the encoded feature sequence, and then segment the feature sequence to facilitate subsequent feature separation.
7. A gas leakage sound signal denoising system based on an hourglass model as claimed in claim 5, characterized in that: In step (1), the calculation process of the feature vector is: In the formula, x c is the input feature vector, c is the number of channels of the input feature vector, [,] is the concatenation operation of the feature vector, f is the feature map containing spatial information encoded along the X-axis and Y-axis, F1 is a nonlinear transformation, δ is a nonlinear activation function, H and W represent the height H and width W of the feature vector respectively, i represents the convolution window position corresponding to f in the X-axis direction, j represents the convolution window position corresponding to f in the Y-axis direction, h represents the height of the corresponding convolution window, and w represents the width of the corresponding convolution window.
8. A gas leakage sound signal denoising system based on an hourglass model as claimed in claim 5, characterized in that: In step (3), the calculation process of the weight matrix in the horizontal and vertical directions is: In the formula, F h is a 1×1 convolution operation in the horizontal direction, F w is a 1×1 convolution operation in the vertical direction, δ is a nonlinear activation function, σ is a Sigmoid activation function, is the weighted output feature in the vertical direction, is the weighted output feature in the horizontal direction, and are two independent tensors after the feature map f is decomposed.