Raman distributed optical fiber temperature sensing signal denoising method and device

Abstract

The invention discloses a Raman distributed optical fiber temperature sensing signal denoising method and device, and relates to the technical field of optical fiber sensing, and the method comprises the steps: inputting a received original Raman signal into a trained wavelet guide dense residual network; carrying out channel feature enhancement and feature reuse enhancement based on wavelet decomposition on the original Raman signal in parallel by adopting an expansion convolutional dense network and a standard convolutional dense network, and correspondingly generating an expansion output feature and a standard output feature; after the expansion output features and the standard output features are spliced, the spliced features are input into a cascaded fusion convolution module and a channel attention module for feature processing, and original signal noise is determined; and subtracting the original signal noise from the original Raman signal to determine a de-noised Raman signal. Based on the scheme, the processing advantages of dense connection, residual learning and attention enhancement are fused under the guidance of the explicit frequency domain provided by wavelet decomposition, and the signal denoising reliability is improved.

Description

Technical Field

[0001] This invention relates to the field of fiber optic sensing technology, and in particular to a method and apparatus for denoising Raman distributed fiber optic temperature sensing signals. Background Technology

[0002] In practical applications, Raman distributed fiber optic temperature sensing (RDTS) systems are subject to complex multi-band noise interference, among which low-frequency noise problems such as ambient temperature drift, laser power jitter, and low-frequency drift of system circuits are particularly prominent, and these noises have complex distributions in the frequency domain.

[0003] To improve Raman distributed temperature sensing signals, existing methods typically introduce deep learning models such as one-dimensional denoising convolutional neural networks (1DDCNN) and residual combined dual convolutional neural networks (RRCDNet) for signal denoising. However, these methods mainly suppress noise in the time domain and have insufficient ability to suppress low-frequency noise, resulting in low reliability of signal denoising. Summary of the Invention

[0004] This invention provides a method and apparatus for denoising Raman distributed fiber optic temperature sensing signals, which solves the technical problem that the deep learning models introduced in existing signal denoising methods have low reliability for Raman distributed temperature sensing signal denoising.

[0005] The first aspect of this invention provides a method for denoising Raman distributed fiber optic temperature sensing signals, comprising:

[0006] The received raw Raman signal is input into a trained wavelet-guided dense residual network, which includes a dilated convolutional dense network, a standard convolutional dense network, a fused convolutional module, and a channel attention module.

[0007] The original Raman signal is enhanced by channel features and feature reuse enhancement based on wavelet decomposition using the dilated convolutional dense network and the standard convolutional dense network in parallel, thereby generating dilated output features and standard output features accordingly.

[0008] After concatenating the dilated output feature with the standard output feature, the input is processed by the cascaded fusion convolution module and the channel attention module to determine the original signal noise.

[0009] The denoised Raman signal is determined by subtracting the noise from the original Raman signal.

[0010] Optionally, the dilated convolutional dense network includes a wavelet decomposition module, an initial convolution module, a dilated dense residual enhancement module, a channel attention module, and a final standard Conv1d layer. The dilated convolutional dense network further includes a wavelet decomposition module, an initial convolution module, a standard dense residual enhancement module, a channel attention module, and a final standard Conv1d layer. The method of using the dilated convolutional dense network and the standard convolutional dense network in parallel to perform wavelet decomposition-based channel feature enhancement and feature reuse enhancement on the original Raman signal, correspondingly generating dilated output features and standard output features, includes:

[0011] In the dilated convolutional dense network, the original Raman signal is subjected to wavelet transform by the wavelet decomposition module to determine the dilated multi-channel features, and then input into the initial convolution module for feature extraction to determine the first dilated intermediate features;

[0012] In the dilated convolutional dense network, multiple cascaded dilated dense residual enhancement modules are used to sequentially enhance channel features and feature reuse based on the first dilated intermediate feature, and then input into the channel attention module for feature enhancement to generate the second dilated intermediate feature.

[0013] In the dilated convolutional dense network, the second dilated intermediate feature is reduced in dimensionality based on the final standard Conv1d layer and then added element-wise to the original Raman signal to output the dilated output feature;

[0014] In the standard convolutional dense network, the original Raman signal is subjected to wavelet transform by the wavelet decomposition module to determine the standard multi-channel features, and then input into the initial convolution module for feature processing to determine the first standard intermediate features;

[0015] In the standard convolutional dense network, multiple cascaded standard dense residual enhancement modules are used to sequentially enhance channel features and feature reuse based on the first standard intermediate features, and then input into the channel attention module for channel enhancement to generate the second standard intermediate features.

[0016] In the standard convolutional dense network, the second standard intermediate feature is reduced in dimensionality based on the final standard Conv1d layer and then connected to the original Raman signal residual to output the standard output feature.

[0017] Optionally, the dilated dense residual enhancement module includes a cascaded channel attention module, a dilated dense block, and a transition standard Conv1d layer, wherein the channel attention module is residually connected to the transition standard Conv1d layer.

[0018] The dilated dense block comprises multiple densely connected dilated dense layers, each of which includes cascaded dilated Conv1d units, BN units, and ReLU activation functions.

[0019] The standard dense residual enhancement module includes a cascaded channel attention module, a standard dense block, and a transition standard Conv1d layer, wherein the channel attention module is residually connected to the transition standard Conv1d layer.

[0020] The standard dense block comprises multiple densely connected standard dense layers, and the standard dense layer comprises cascaded standard Conv1d units, BN units, and ReLU activation functions.

[0021] Optionally, the channel attention module includes a global average pooling layer, a dimensionality-reduced standard Conv1d layer, a ReLU activation function, a dimensionality-upgraded standard Conv1d layer, and a Sigmoid activation function; the processing procedure of the channel attention module includes:

[0022] The attention input features are compressed using a global average pooling layer to determine the pooling features.

[0023] After learning channel features through a cascaded standard Conv1d layer for dimensionality reduction, ReLU activation function and standard Conv1d layer for dimensionality increase, the input Sigmoid activation function maps the output channel weights.

[0024] Attention output features are generated by multiplying the attention input features with the channel weights channel by channel.

[0025] Optionally, the processing procedure of the wavelet decomposition module includes:

[0026] The original Raman signal is subjected to wavelet transform using a wavelet basis to determine the low-frequency and high-frequency components.

[0027] After upsampling the low-frequency and high-frequency components, they are concatenated with the original Raman signal to determine the dilated multichannel feature or the standard multichannel feature.

[0028] Optionally, the wavelet basis is the db3 wavelet basis; the step of performing wavelet transform on the original Raman signal using the wavelet basis to determine the low-frequency and high-frequency components includes:

[0029] The original Raman signal was subjected to multi-level discrete wavelet transform using the db3 wavelet basis to determine the low-frequency and high-frequency components.

[0030] A second aspect of the present invention provides a Raman distributed fiber optic temperature sensing signal denoising device, comprising:

[0031] The signal input module is used to input the received raw Raman signal into the trained wavelet-guided dense residual network, which includes a dilated convolutional dense network, a standard convolutional dense network, a fusion convolutional module, and a channel attention module.

[0032] The feature extraction module is used to perform channel feature enhancement and feature reuse enhancement based on wavelet decomposition on the original Raman signal in parallel using the dilated convolutional dense network and the standard convolutional dense network, and generate dilated output features and standard output features accordingly.

[0033] The noise prediction module is used to concatenate the dilated output feature with the standard output feature and then input the concatenated fusion convolution module and channel attention module for feature processing to determine the original signal noise.

[0034] The signal denoising module is used to subtract the original signal noise from the original Raman signal to determine the denoised Raman signal.

[0035] A computer device provided in a third aspect of the present invention includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor performs the steps of the Raman distributed fiber optic temperature sensing signal denoising method as described in any one of the first aspects of the present invention.

[0036] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the Raman distributed optical fiber temperature sensing signal denoising method as described in any one of the first aspects of the present invention.

[0037] The fifth aspect of the present invention provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the Raman distributed optical fiber temperature sensing signal denoising method as described in any one of the first aspects of the present invention.

[0038] As can be seen from the above technical solutions, the present invention has the following advantages:

[0039] The above-described scheme of the present invention provides a method for denoising Raman distributed fiber optic temperature sensing signals, comprising: inputting the received original Raman signal into a trained wavelet-guided dense residual network, the wavelet-guided dense residual network including a dilated convolutional dense network, a standard convolutional dense network, a fusion convolutional module, and a channel attention module; using the dilated convolutional dense network and the standard convolutional dense network in parallel to perform channel feature enhancement and feature reuse enhancement based on wavelet decomposition on the original Raman signal, correspondingly generating dilated output features and standard output features; concatenating the dilated output features and the standard output features, and inputting them into a cascaded fusion convolutional module and a channel attention module for feature processing to determine the original signal noise; subtracting the original signal noise from the original Raman signal to determine the denoised Raman signal. Based on the above scheme, the processing advantages of fusion dense connections, residual learning, and attention enhancement under the explicit frequency domain guidance provided by wavelet decomposition are improved, thereby improving the reliability of signal denoising. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart of the steps of a Raman distributed optical fiber temperature sensing signal denoising method provided in Embodiment 1 of the present invention;

[0042] Figure 2 This is a schematic diagram of the architecture of the wavelet-guided dense residual network provided in Embodiment 1 of the present invention;

[0043] Figure 3 This is a schematic diagram of the structure of a dense block provided in Embodiment 1 of the present invention;

[0044] Figure 4 This is a schematic diagram of the SE module provided in Embodiment 1 of the present invention;

[0045] Figure 5 This is a structural block diagram of a Raman distributed optical fiber temperature sensing signal denoising device provided in Embodiment 2 of the present invention. Detailed Implementation

[0046] This invention provides a method and apparatus for denoising Raman distributed fiber optic temperature sensing signals, which addresses the technical problem of low reliability of deep learning models introduced in existing signal denoising methods for Raman distributed temperature sensing signals.

[0047] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0048] Please see Figure 1 The present invention provides a Raman distributed fiber optic temperature sensing signal denoising method, which is mainly applied in devices such as Raman distributed fiber optic temperature sensing (RDTS) systems that realize distributed temperature measurement based on the Raman scattering principle, including:

[0049] Step 101: Input the received raw Raman signal into the trained wavelet-guided dense residual network. The wavelet-guided dense residual network includes a dilated convolutional dense network, a standard convolutional dense network, a fusion convolutional module, and a channel attention module.

[0050] The original Raman signal refers to the distributed fiber optic temperature sensing signal acquired based on the Raman scattering principle. Due to the limitations of the acquisition equipment itself or interference from the external environment, noise may be present in the signal.

[0051] It should be noted that, in order to denoise the original Raman signal, this embodiment designs a wavelet-guided dense residual network (WGDRNet). WGDRNet integrates the processing advantages of wavelet transform, dense connections, residual learning, and attention enhancement, such as... Figure 2 As shown, the model includes a dilated convolutional dense network (WGDRNetLeft), a standard convolutional dense network (WGDRNetRight), a fused convolutional module, and a channel attention module. The wavelet-guided dense residual network is trained to determine the trained wavelet-guided dense residual network. When the original Raman signal is received, the trained wavelet-guided dense residual network is used for signal denoising.

[0052] In one specific implementation of this embodiment, the training process of the denoising model includes: using a training noisy Raman signal as input to a wavelet-guided dense residual network to determine the training prediction noise, then denoising the network, and outputting a predicted denoised Raman signal, i.e., using a residual learning paradigm. , To predict the denoised Raman signal, To train noisy Raman signals, To train the prediction noise, the loss function value is determined based on the training noisy Raman signal and the predicted denoised Raman signal. The model parameters of the wavelet-guided dense residual network to be trained are iteratively optimized according to the gradient of the loss function value until the loss function value converges, thus determining the trained wavelet-guided dense residual network.

[0053] In a more specific implementation of this embodiment, the calculation process of the loss function value includes:

[0054]

[0055] In the formula, For model parameters, The value of the loss function. For signal index, For the number of signals, For the first A training noisy Raman signal, For the first A training noisy Raman signal, For the first A predictive denoised Raman signal It is the square of the Frobenius norm.

[0056] Step 102: Use dilated convolutional dense network and standard convolutional dense network in parallel to perform channel feature enhancement and feature reuse enhancement based on wavelet decomposition on the original Raman signal, and generate dilated output features and standard output features accordingly.

[0057] It should be noted that in this embodiment, the original Raman signal is processed by two branch networks: a dilated convolutional dense network and a standard convolutional dense network. The dilated output feature and the standard output feature are the feature outputs of the left branch network and the right branch network, respectively.

[0058] In one specific implementation of this embodiment, such as Figure 2 As shown, the dilated convolutional dense network includes a wavelet decomposition module, an initial convolution module, a dilated dense residual enhancement module, a channel attention module, and a final standard Conv1d layer. The dilated convolutional dense network and the standard convolutional dense network are used in parallel to enhance the channel features and reuse features of the original Raman signal based on wavelet decomposition, generating dilated output features and standard output features accordingly.

[0059] In the dilated convolutional dense network, the original Raman signal is subjected to wavelet transform by the wavelet decomposition module to determine the dilated multi-channel features, and then input into the initial convolution module for feature extraction to determine the first dilated intermediate features;

[0060] In dilated convolutional dense networks, multiple cascaded dilated dense residual enhancement modules are used to sequentially enhance channel features and feature reuse based on the first dilated intermediate feature. Then, the input is processed by the channel attention module to enhance the features and generate the second dilated intermediate feature.

[0061] In dilated convolutional dense networks, the second dilated intermediate feature is reduced in dimensionality based on the final standard Conv1d layer and then added element-wise to the original Raman signal to output the dilated output feature.

[0062] In a standard convolutional dense network, the original Raman signal is subjected to wavelet transform by a wavelet decomposition module to determine standard multi-channel features, which are then input into an initial convolution module for feature processing to determine the first standard intermediate features.

[0063] In a standard convolutional dense network, multiple cascaded standard dense residual enhancement modules are used to sequentially enhance channel features and feature reuse based on the first standard intermediate features. Then, the input is processed by the channel attention module to enhance the channels and generate the second standard intermediate features.

[0064] In a standard convolutional dense network, the intermediate features of the second standard are reduced in dimensionality based on the final standard Conv1d layer and then connected to the residual of the original Raman signal to output the standard output features.

[0065] It should be noted that the dilated convolutional dense network in this embodiment has a similar network structure to the standard convolutional dense network. The main difference is that the standard dense residual enhancement module of the standard convolutional dense network uses a standard convolution with a dilation rate of 1, while the corresponding structure in the dilated convolutional dense network uses a dilated convolution with a dilation rate greater than 1. Dilated convolution can provide a larger receptive field. In the dual-branch network, by performing wavelet transform on the original Raman signal to decompose the signal, explicit multi-resolution frequency domain knowledge can be provided to enhance the frequency domain prior advantage. Dense connections are used to achieve efficient feature reuse, residual connections are used to effectively alleviate gradient vanishing, and channel feature enhancement is used to adaptively optimize feature channels, thereby achieving end-to-end optimization of frequency domain prior and time domain feature learning.

[0066] In a more specific embodiment of this example, the processing procedure of the wavelet decomposition module includes:

[0067] Wavelet transform of the original Raman signal is performed using a wavelet basis to determine the low-frequency and high-frequency components.

[0068] After upsampling the low-frequency and high-frequency components, they are spliced ​​with the original Raman signal to determine the dilated multichannel feature or the standard multichannel feature.

[0069] Low-frequency components refer to the slowly changing parts of a signal.

[0070] High-frequency components refer to the rapidly changing parts of a signal.

[0071] It should be noted that by performing wavelet transform on the original Raman signal, components of different frequency scales are decomposed from the signal, thereby obtaining low-frequency components and high-frequency components. After upsampling to the length of the original Raman signal, the components are concatenated with the original Raman signal in the channel dimension to obtain the dilated multi-channel features output by the original Raman signal in the dilated convolutional dense network and the standard multi-channel features output by the standard convolutional dense network.

[0072] In a preferred embodiment of this example, the wavelet basis is the db3 wavelet basis; the wavelet transform of the original Raman signal using the wavelet basis to determine the low-frequency and high-frequency components includes: performing a multi-level discrete wavelet transform of the original Raman signal using the db3 wavelet basis to determine the low-frequency and high-frequency components.

[0073] It should be noted that, in specific implementations, the wavelet transform performed by the wavelet decomposition module (WDFDM) can be either discrete or continuous. Discrete wavelet transforms can use wavelet bases such as Daubechies wavelet, Haar wavelet, and Biorthogonal wavelet, while continuous wavelet transforms can use wavelet bases such as Morlet wavelet. For specific selection of the appropriate wavelet base, refer to existing technologies. In a preferred implementation, the Daubechies wavelet base is used for multi-level discrete wavelet transform. The number of decomposition levels of the discrete wavelet transform can be designed as 2, 3, or 4 levels. Although more levels may provide more detailed frequency domain information, they will increase computational complexity. Therefore, more preferably, the Daubechies 3 (db3) wavelet base can better balance reconstruction accuracy and computational efficiency.

[0074] The original Raman signal is decomposed into approximate coefficients as low-frequency components by performing a three-level discrete wavelet transform on the db3 wavelet basis. and detail coefficients as high-frequency components The process is represented as: , This is the wavelet decomposition layer index. The original Raman signal, It is a db3 wavelet basis. for, These are approximate coefficients. These are the detail coefficients; the decomposed signal is upsampled and then concatenated with the original Raman signal, i.e. , These are the approximate coefficients after upsampling. This is the upsampled result. This provides the network with explicit frequency domain prior knowledge by providing multi-channel features (i.e., expanded multi-channel features or standard multi-channel features).

[0075] In a more specific implementation of this embodiment, such as Figure 2 and Figure 3 As shown, the dilated dense residual enhancement module includes a cascaded channel attention module, a dilated dense block, and a transition standard Conv1d layer, wherein the channel attention module is residually connected to the transition standard Conv1d layer.

[0076] The dilated dense block consists of multiple densely connected dilated dense layers, each of which includes cascaded dilated Conv1d units, BN units, and ReLU activation functions.

[0077] The standard dense residual enhancement module includes a cascaded channel attention module, a standard dense block, and a transition standard Conv1d layer, wherein the channel attention module is residually connected to the transition standard Conv1d layer;

[0078] A standard dense block consists of multiple densely connected standard dense layers, which include cascaded standard Conv1d units, BN units, and ReLU activation functions.

[0079] It should be noted that the dilated dense residual enhancement module has a similar structural composition to the standard dense residual enhancement module. The main difference is that the standard dense residual enhancement module uses dense blocks based on standard Conv1d units, while the dilated dense residual enhancement module uses dense blocks based on dilated Conv1d units. For example, the dilation rate of the dilated Conv1d unit is set to 2. It can be understood that the dense block adopts a densely connected structure. , For the hierarchical index of dense layers, For the first Output of dense layers For the first In a dense layer, each layer in the dense block is directly connected to all the preceding layers to achieve feature reuse. In specific implementation, the growth rate can be set to 16 to control the growth of the number of channels.

[0080] Furthermore, in one implementation, there are multiple inflated dense blocks and standard dense blocks, and a dense transition layer is set between adjacent inflated dense blocks or standard dense blocks for channel compression.

[0081] Step 103: After concatenating the dilated output features with the standard output features, input them into the cascaded fusion convolution module and channel attention module for feature processing to determine the original signal noise.

[0082] Original signal noise refers to the noise predicted from the original Raman signal.

[0083] It should be noted that after the features output by the dual-branch network are concatenated, they undergo dimensionality reduction and nonlinear mapping in the fusion convolution module, and are then processed by the final channel attention module to predict output noise.

[0084] In one specific implementation of this embodiment, such as Figure 2 As shown, the fused convolutional module consists of cascaded standard Conv1d layers and Tanh activation functions.

[0085] In some preferred embodiments of this example, such as Figure 2 and Figure 4 As shown, the channel attention module includes a global average pooling layer, a dimensionality-reduced standard Conv1d layer, a ReLU activation function, a dimensionality-upgraded standard Conv1d layer, and a Sigmoid activation function; the processing procedure of the channel attention module includes:

[0086] The attention input features are compressed using a global average pooling layer to determine the pooling features.

[0087] After learning channel features through a cascaded standard Conv1d layer for dimensionality reduction, ReLU activation function and standard Conv1d layer for dimensionality increase, the input Sigmoid activation function maps the output channel weights.

[0088] Attention output features are generated by multiplying the attention input features with the channel weights channel by channel.

[0089] It should be noted that modules capable of channel feature enhancement include, but are not limited to, the SE (Squeeze-and-Excitation) module, the ECA (Efficient Channel Attention) module, and the CBAM (Convolutional Block Attention) module. In this embodiment, the processing of using the SE module as the channel attention module is as follows:

[0090] First, channel statistics are extracted using global average pooling:

[0091]

[0092] In the formula, For channel index, For the first The average pooling vector of each channel For the first The attention input feature vector for each channel, It is a global average pooling layer. The height of the feature map, The width of the feature map. For the position index in the height direction, The position index is in the width direction; the vectors of all channels after global average pooling constitute the pooling features;

[0093] Then, the channel weights are learned using a two-layer fully connected network consisting of a standard Conv1d layer and a ReLU activation function. In a specific implementation, the compression ratio can be set to r=16, where the first layer uses ReLU activation and the second layer uses Sigmoid activation. In the formula, For channel weights, For pooling features, This is the weight matrix. This is the first weight matrix. This is the second weight matrix. It is the ReLU activation function. It is the Sigmoid activation function.

[0094] Finally, adaptive channel weight adjustment is implemented: In the formula, For the first The attention output feature vector for each channel. For the first The channel weight vector of each channel. For the first The attention input feature vector for each channel, This is the feature scaling function.

[0095] Step 104: Subtract the original signal noise from the original Raman signal to determine the denoised Raman signal.

[0096] A denoised Raman signal refers to the signal obtained after denoising the original Raman signal.

[0097] It should be noted that the denoising result is obtained by subtracting the original signal noise from the original Raman signal.

[0098] To better illustrate the effectiveness of the above method, refer to Tables 1 and 2, which show a comparison of experimental results between this embodiment and existing denoising algorithms:

[0099] Table 1 Evaluation metrics for processing Raman distributed fiber optic sensing signals under different noise amplitudes

[0100]

[0101] Table 2 Evaluation Indicators of Processing Temperature Curves under Different Noise Amplitudes

[0102]

[0103] In Tables 1 and 2, WD3 represents a three-layer wavelet denoising algorithm, WD4 represents a four-layer wavelet denoising algorithm, RRCDNet is an existing neural network model, and WGDRNet represents the method of this embodiment. The evaluation metrics in the tables show that compared to traditional wavelet denoising algorithms, the SSIM exponent of this embodiment is significantly improved, avoiding the subjectivity of threshold selection and signal distortion problems, and achieving end-to-end optimization of frequency domain analysis and deep learning. Compared to the RRCDNet model, the smoothness and peak-to-peak value of the temperature curve are significantly improved.

[0104] In this embodiment of the invention, the multi-resolution analysis capability of wavelet transform is combined with deep learning to provide explicit frequency domain guidance. Dense connections reduce information loss and repetitive learning, multi-path residual connections effectively alleviate gradient vanishing, and channel feature enhancement is used to perform adaptive selection optimization in feature channels. Overall, end-to-end optimization of frequency domain prior and time domain feature learning is achieved, which helps to improve the low-frequency noise suppression capability and thus improve the reliability of signal denoising.

[0105] Please see Figure 5 The present invention provides a Raman distributed optical fiber temperature sensing signal denoising device according to Embodiment 2, comprising:

[0106] The signal input module 501 is used to input the received raw Raman signal into the trained wavelet-guided dense residual network. The wavelet-guided dense residual network includes a dilated convolutional dense network, a standard convolutional dense network, a fusion convolutional module, and a channel attention module.

[0107] The feature extraction module 502 is used to perform channel feature enhancement and feature reuse enhancement based on wavelet decomposition on the original Raman signal in parallel using a dilated convolutional dense network and a standard convolutional dense network, and generate dilated output features and standard output features accordingly.

[0108] The noise prediction module 503 is used to concatenate the dilated output features with the standard output features and then input them into the cascaded fusion convolution module and channel attention module for feature processing to determine the original signal noise.

[0109] The signal denoising module 504 is used to subtract the original signal noise from the original Raman signal to determine the denoised Raman signal.

[0110] Furthermore, the dilated convolutional dense network includes a wavelet decomposition module, an initial convolutional module, a dilated dense residual enhancement module, a channel attention module, and a final standard Conv1d layer. The dilated convolutional dense network and the standard convolutional dense network are used in parallel to perform channel feature enhancement and feature reuse enhancement based on wavelet decomposition on the original Raman signal, correspondingly generating dilated output features and standard output features, including:

[0111] In the dilated convolutional dense network, the original Raman signal is subjected to wavelet transform by the wavelet decomposition module to determine the dilated multi-channel features, and then input into the initial convolution module for feature extraction to determine the first dilated intermediate features;

[0112] In dilated convolutional dense networks, multiple cascaded dilated dense residual enhancement modules are used to sequentially enhance channel features and feature reuse based on the first dilated intermediate feature. Then, the input is processed by the channel attention module to enhance the features and generate the second dilated intermediate feature.

[0113] In dilated convolutional dense networks, the second dilated intermediate feature is reduced in dimensionality based on the final standard Conv1d layer and then added element-wise to the original Raman signal to output the dilated output feature.

[0114] In a standard convolutional dense network, the original Raman signal is subjected to wavelet transform by a wavelet decomposition module to determine standard multi-channel features, which are then input into an initial convolution module for feature processing to determine the first standard intermediate features.

[0115] In a standard convolutional dense network, multiple cascaded standard dense residual enhancement modules are used to sequentially enhance channel features and feature reuse based on the first standard intermediate features. Then, the input is processed by the channel attention module to enhance the channels and generate the second standard intermediate features.

[0116] In a standard convolutional dense network, the intermediate features of the second standard are reduced in dimensionality based on the final standard Conv1d layer and then connected to the residual of the original Raman signal to output the standard output features.

[0117] Furthermore, the dilated dense residual enhancement module includes a cascaded channel attention module, a dilated dense block, and a transition standard Conv1d layer, wherein the channel attention module is residually connected to the transition standard Conv1d layer.

[0118] The dilated dense block consists of multiple densely connected dilated dense layers, each of which includes cascaded dilated Conv1d units, BN units, and ReLU activation functions.

[0119] The standard dense residual enhancement module includes a cascaded channel attention module, a standard dense block, and a transition standard Conv1d layer, wherein the channel attention module is residually connected to the transition standard Conv1d layer;

[0120] A standard dense block consists of multiple densely connected standard dense layers, which include cascaded standard Conv1d units, BN units, and ReLU activation functions.

[0121] Furthermore, the channel attention module includes a global average pooling layer, a dimensionality-reduced standard Conv1d layer, a ReLU activation function, a dimensionality-upgraded standard Conv1d layer, and a Sigmoid activation function; the processing procedure of the channel attention module includes:

[0122] The attention input features are compressed using a global average pooling layer to determine the pooling features.

[0123] After learning channel features through a cascaded standard Conv1d layer for dimensionality reduction, ReLU activation function and standard Conv1d layer for dimensionality increase, the input Sigmoid activation function maps the output channel weights.

[0124] Attention output features are generated by multiplying the attention input features with the channel weights channel by channel.

[0125] Furthermore, the wavelet decomposition module's processing steps include:

[0126] Wavelet transform of the original Raman signal is performed using a wavelet basis to determine the low-frequency and high-frequency components.

[0127] After upsampling the low-frequency and high-frequency components, they are spliced ​​with the original Raman signal to determine the dilated multichannel feature or the standard multichannel feature.

[0128] Furthermore, the wavelet basis is the db3 wavelet basis; wavelet transform is performed on the original Raman signal using the wavelet basis to determine the low-frequency and high-frequency components, including:

[0129] The original Raman signal was subjected to multi-level discrete wavelet transform using the db3 wavelet basis to determine the low-frequency and high-frequency components.

[0130] This invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program; when the computer program is executed by the processor, the processor performs the steps of the Raman distributed fiber optic temperature sensing signal denoising method as described in any of the above embodiments.

[0131] This invention also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the Raman distributed fiber optic temperature sensing signal denoising method as described in any of the above embodiments.

[0132] This invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the Raman distributed fiber optic temperature sensing signal denoising method as described in any of the above embodiments.

[0133] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described device and module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0134] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0135] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0136] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0137] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0138] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for denoising a Raman distributed fiber temperature sensing signal, the method comprising: The application relates to a method for denoising a received original Raman signal. ​ The received original Raman signal is input into a trained wavelet-guided dense residual network, which comprises an expanded convolution dense network, a standard convolution dense network, a fusion convolution module and a channel attention module. The expanded convolution dense network and the standard convolution dense network are used to perform wavelet decomposition-based channel feature enhancement and feature reuse strengthening on the original Raman signal in parallel, so as to generate expanded output features and standard output features. After the expanded output features and the standard output features are spliced, the features are processed by the cascaded fusion convolution module and the channel attention module to determine the original signal noise. The original signal noise is subtracted from the original Raman signal to determine a denoised Raman signal.

2. The Raman distributed fiber temperature sensing signal denoising method of claim 1, wherein, The expanded convolution dense network comprises a wavelet decomposition module, an initial convolution module, an expanded dense residual enhancement module, a channel attention module and a final standard Conv1d layer, and the standard convolution dense network comprises a wavelet decomposition module, an initial convolution module, a standard dense residual enhancement module, a channel attention module and a final standard Conv1d layer. In the expanded convolution dense network, the original Raman signal is subjected to wavelet transformation by the wavelet decomposition module to determine expanded multi-channel features, and the features are input into the initial convolution module for feature extraction to determine first expanded intermediate features. In the expanded convolution dense network, the first expanded intermediate features are sequentially subjected to channel feature enhancement and feature reuse strengthening by the cascaded multiple expanded dense residual enhancement modules, and then input into the channel attention module for feature enhancement to generate second expanded intermediate features. In the expanded convolution dense network, the second expanded intermediate features are dimensionally reduced based on the final standard Conv1d layer, and then added element by element with the original Raman signal to output expanded output features. In the standard convolution dense network, the original Raman signal is subjected to wavelet transformation by the wavelet decomposition module to determine standard multi-channel features, and the features are input into the initial convolution module for feature processing to determine first standard intermediate features. In the standard convolution dense network, the first standard intermediate features are sequentially subjected to channel feature enhancement and feature reuse strengthening by the cascaded multiple standard dense residual enhancement modules, and then input into the channel attention module for channel enhancement to generate second standard intermediate features. In the standard convolution dense network, the second standard intermediate features are dimensionally reduced based on the final standard Conv1d layer, and then connected in residual with the original Raman signal to output standard output features.

3. The Raman distributed fiber temperature sensing signal denoising method of claim 2, wherein, The expanded dense residual enhancement module comprises a cascaded channel attention module, an expanded dense block and a transition standard Conv1d layer, wherein the channel attention module and the transition standard Conv1d layer are connected in residual. The expansion dense block comprises a plurality of densely connected expansion dense layers, and the expansion dense layer comprises a cascaded expansion Conv1d unit, a BN unit and a ReLU activation function; The standard dense residual enhancement module comprises a cascaded channel attention module, a standard dense block and a transition standard Conv1d layer, wherein the channel attention module is in residual connection with the transition standard Conv1d layer; The standard dense block comprises a plurality of densely connected standard dense layers, and the standard dense layer comprises a cascaded standard Conv1d unit, a BN unit and a ReLU activation function.

4. The Raman distributed fiber temperature sensing signal denoising method according to claim 1, 2 or 3, characterized in that, The channel attention module comprises a global average pooling layer, a dimension-reduced standard Conv1d layer, a ReLU activation function, a dimension-increased standard Conv1d layer and a Sigmoid activation function; the processing process of the channel attention module comprises: The global average pooling layer is used to compress the features of the attention input features to determine the pooled features; After the channel feature learning of the cascaded dimension-reduced standard Conv1d layer, the ReLU activation function and the dimension-increased standard Conv1d layer, the Sigmoid activation function is input to map the output channel weight; The attention input features and the channel weight are multiplied channel by channel to generate the attention output features.

5. The Raman distributed fiber temperature sensing signal denoising method of claim 2, wherein, The processing process of the wavelet decomposition module comprises: The wavelet basis is used to perform wavelet transform on the original Raman signal to determine the low-frequency component and the high-frequency component; After the low-frequency component and the high-frequency component are up-sampled, they are channel-spliced with the original Raman signal to determine the expansion multi-channel feature or the standard multi-channel feature.

6. The Raman distributed fiber temperature sensing signal denoising method of claim 5, wherein, The wavelet basis is a db3 wavelet basis; the wavelet transform of the original Raman signal by the wavelet basis to determine the low-frequency component and the high-frequency component comprises: The db3 wavelet basis is used to perform multi-layer discrete wavelet transform on the original Raman signal to determine the low-frequency component and the high-frequency component.

7. A Raman distributed fiber temperature sensing signal denoising device, characterized in that, It comprises: A signal input module is configured to input the received original Raman signal into a trained wavelet-guided dense residual network, wherein the wavelet-guided dense residual network comprises an expansion convolution dense network, a standard convolution dense network, a fusion convolution module and a channel attention module; A feature extraction module is configured to perform channel feature enhancement and feature reuse enhancement based on wavelet decomposition on the original Raman signal by using the expansion convolution dense network and the standard convolution dense network in parallel, to correspondingly generate expansion output features and standard output features; A noise prediction module is configured to input the expansion output features and the standard output features after splicing to perform feature processing on the fusion convolution module and the channel attention module in cascade to determine original signal noise; A signal denoising module is configured to subtract the original signal noise from the original Raman signal to determine a denoised Raman signal.

8. A computer device, comprising: It comprises a memory and a processor, wherein the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the Raman distributed fiber temperature sensing signal denoising method according to any one of claims 1-6.

9. A computer readable storage medium having stored thereon computer programs / instructions, characterized in that, The computer program / instructions, when executed by the processor, implement the steps of the Raman distributed fiber temperature sensing signal denoising method according to any one of claims 1-6.

10. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions, when executed by the processor, implement the steps of the Raman distributed fiber temperature sensing signal denoising method according to any one of claims 1-6.

Images