A DAS dynamic strain monitoring signal denoising method and device based on Dn-ResUnet
By processing DAS signals through the Dn-ResUnet neural network and utilizing a residual block encoder and decoder, the problem of various noise processing in high-noise environments is solved, achieving efficient signal denoising and significantly improving the signal-to-noise ratio.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA SHIP DEV & DESIGN CENT
- Filing Date
- 2026-03-18
- Publication Date
- 2026-07-10
Smart Images
Figure CN122364641A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of distributed fiber optic acoustic wave sensing (DAS) dynamic strain monitoring signal processing technology, specifically to a DAS signal denoising method and device based on a denoising residual neural network Dn-ResUnet. Background Technology
[0002] Distributed Optical Fiber Acoustic Sensing (DAS) is a novel sensing technology that enables distributed detection of external strain. DAS utilizes three scattering wavelengths: Rayleigh scattering, Raman scattering, and Brillouin scattering. Rayleigh and Brillouin scattering can be used for strain monitoring, while Raman scattering is used for temperature monitoring. Compared to Brillouin scattering, Rayleigh scattering has stronger scattering energy and a wider vibration detection distance and frequency range, making it more suitable for dynamic strain monitoring in DAS. During fiber fabrication, the fiber interior is not perfectly uniform, resulting in a non-uniform refractive index distribution. This generates relatively high-intensity secondary wave sources, producing backscattered light in other directions during transmission—this is Rayleigh scattering. The phase information of the backscattered Rayleigh signal is used to measure the strain caused by acoustic waves or vibrations. When the sensing fiber is subjected to external vibration in the longitudinal or axial direction, the fiber length, internal refractive index, and core diameter change due to the vibration, leading to changes in optical path difference and phase, which can be expressed as:
[0003] in, The length of the optical fiber. The transmission constant, The diameter of the optical fiber. Let be the refractive index of the fiber core. The three terms on the right-hand side of the equation correspond to the strain effect, the elasto-optic effect, and the Poisson effect, respectively.
[0004] Due to its advantages such as high sensitivity, compactness, integration of multiple sensors, and resistance to electromagnetic interference in extreme environments, Direct Atmospheric Sensors (DAS) are now widely used in various fields such as intelligent security, pipeline monitoring, and seismic exploration. However, the signals detected by DAS are complex (e.g., random noise, coupling noise, optical noise), resulting in a low signal-to-noise ratio (SNR). Therefore, denoising of the acquired DAS signals is necessary to improve signal quality. Currently, DAS signal denoising methods are divided into three categories: The first category is based on filtering methods, which first determine the noise frequency and directly use bandpass filtering to suppress the noise. This method inevitably loses the effective signal at the corresponding frequency when removing noise. The second category is sparse inversion methods based on dictionary learning. These methods use strategies such as cosine transform and dictionary learning to first obtain adaptive basis functions and then utilize the sparsity of the transform domain to achieve denoising. These methods are difficult to adjust parameters and have low computational efficiency. The third category is machine learning algorithms. These algorithms learn nonlinear mapping features based on neural networks and can achieve better denoising effects. Machine learning also has problems such as high computational complexity and a lack of labeled datasets. As a data-driven algorithm, it is highly dependent on label data.
[0005] In summary, existing noise reduction methods suffer from a sharp performance degradation in high-noise environments, struggle to handle multiple noise types simultaneously, and are difficult to balance computational efficiency with effective noise reduction. Therefore, there is an urgent need for a method that can still function effectively in high-noise environments, process mixed noise in a single pass, and achieve high efficiency. Summary of the Invention
[0006] This invention provides a method and apparatus for denoising DAS dynamic strain monitoring signals based on a denoising-residual network (Dn-ResUnet). It uses a residual block encoder and decoder and introduces residual learning to obtain the noise distribution from the noise signal. It can process multiple types of noise at once, greatly improve the denoising effect, and can successfully denoise even under high noise conditions.
[0007] In a first aspect, the present invention provides a method for denoising DAS dynamic strain monitoring signals based on Dn-ResUnet, comprising: Acquire noisy DAS signals; The noisy DAS signal is input into a pre-trained denoising neural network model, which is a Dn-ResUnet network based on the U-Net architecture and incorporating residual blocks. The noise estimation signal is output through the noise reduction neural network model. The noise-infused DAS signal is subtracted from the noise estimation signal to obtain the denoised DAS signal.
[0008] In some instances, the loss function of the denoising neural network model is: ,in, For loss function, For input The corresponding output at that time For the Frobenius norm, This represents the training data, where N is the number of datasets. These are network parameters.
[0009] In some instances, the noise reduction neural network model includes a convolutional block, an encoder path, a convolutional block, a decoder path, and an output convolutional block connected in sequence. The encoder path includes multiple cascaded downsampled residual blocks, the decoder includes multiple cascaded upsampled residual blocks, and residual connections are provided between corresponding layers of the encoder and the decoder.
[0010] In some instances, the method further includes, before inputting the noisy DAS signal into a pre-trained denoising neural network model: The spatiotemporal spectrum of the noisy DAS signal is unfolded into a one-dimensional array.
[0011] In some instances, the training of the denoising neural network model includes: The training set was constructed by mixing measured DAS signals with various synthetic noises. The real noise signal in the training set is input into the noise reduction neural network model to obtain the estimated noise signal; Calculate the mean square error between the estimated noise signal and the corresponding real noise signal; The gradient descent algorithm is used to backpropagate the gradient of the loss function to each layer of the network and update the network weight parameters. The optimization aims to minimize the mean square error between the noise estimate and the actual noise value of the network output, resulting in a pre-trained noise reduction neural network model.
[0012] Secondly, the present invention provides a DAS dynamic strain monitoring signal noise reduction device based on Dn-ResUnet, comprising: Noise signal acquisition module, used to acquire DAS signals containing noise; The noise estimation module is used to input the noisy DAS signal into a pre-trained denoising neural network model and output a noise estimation signal through the denoising neural network model. The denoising neural network model is a Dn-ResUnet network based on the U-Net architecture and introducing residual blocks. The noise reduction module is used to subtract the noise estimation signal from the noisy DAS signal to obtain the noise-reduced DAS signal.
[0013] In some instances, the loss function of the denoising neural network model is: ,in, For loss function, For input The corresponding output at that time For the Frobenius norm, This represents the training data, where N is the number of datasets. These are network parameters.
[0014] In some instances, the noise reduction neural network model includes a convolutional block, an encoder path, a convolutional block, a decoder path, and an output convolutional block connected in sequence. The encoder path includes multiple cascaded downsampled residual blocks, the decoder includes multiple cascaded upsampled residual blocks, and residual connections are provided between corresponding layers of the encoder and the decoder.
[0015] In some instances, the apparatus further includes a preprocessing module for unfolding the spatiotemporal spectrum of the noisy DAS signal into a one-dimensional array.
[0016] In some instances, the apparatus further includes: a training module for constructing a training set using measured DAS signals mixed with various synthetic noises; inputting real noise signals from the training set into the denoising neural network model to obtain estimated noise signals; calculating the mean square error between the estimated noise signals and the corresponding real noise signals; backpropagating the gradient of the loss function to each layer of the network using a gradient descent algorithm to update the network weight parameters; and optimizing the network to minimize the mean square error between the noise estimates and the real noise values, thereby obtaining a pre-trained denoising neural network model.
[0017] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: (1) The present invention uses a residual block encoder and decoder and introduces residual learning to obtain the noise distribution from the noise signal, which can process multiple noises at one time, greatly improve the noise reduction effect, and can also successfully reduce noise under high noise conditions.
[0018] (2) The present invention proposes a neural network algorithm framework for denoising DAS dynamic strain monitoring signals. This framework can remove various noises and adaptively process parameters without multi-step processing, and can also achieve good denoising effect in high noise environment.
[0019] (3) The decoder in this invention uses residual connection between the residual block in the same layer and the residual block in the upper layer to combine the residual information in the same layer with the output information of the upper layer to better reconstruct the noise.
[0020] (4) The present invention adopts the DAS (Distributed Acoustic Sensing) dynamic strain monitoring signal noise reduction method of Dn-ResUnet. The results of experiments or simulations are used to verify that the SNR of the noise-reduced signal is usually 3~10 dB higher than that of the original noise signal. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.
[0022] Figure 1 This is a schematic diagram of the DAS dynamic strain monitoring system provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the method flow provided in an embodiment of the present invention; Figure 3 This is a computational layer diagram of the residual block provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the internal structure of the residual block provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the device provided in an embodiment of the present invention. Detailed Implementation
[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0024] In the following description, specific embodiments of the invention will be illustrated with reference to steps and symbols performed by one or more computers, unless otherwise stated. Therefore, these steps and operations will be referred to several times as being performed by a computer, and computer execution as referred to herein includes operations by a computer processing unit representing electronic signals of data in a structured format. This operation transforms the data or maintains it at a location in the computer's memory system, which can be reconfigured or otherwise alter the operation of the computer in a manner well known to those skilled in the art. The data structure maintained by the data is the physical location of the memory, which has specific characteristics defined by the data format. However, the principles of the invention described above are not intended to be limiting, and those skilled in the art will understand that many of the following steps and operations can also be implemented in hardware.
[0025] The terms "module" or "unit" as used herein can be considered as software objects executing on the computing system. Different components, modules, engines, and services described herein can be considered as implementations on the computing system. The apparatus and methods described herein are preferably implemented in software, but can also be implemented in hardware, both of which are within the scope of this invention.
[0026] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.
[0027] In this embodiment of the invention, a DAS signal denoising method based on a convolutional neural network is provided, comprising the following steps: Step 1: Obtain noisy DAS data.
[0028] Step 2: Expand the spatiotemporal spectra of the DAS data obtained in Step 1 into a one-dimensional array in a specific manner.
[0029] Step 3: Construct a new intelligent denoising framework, Dn-ResUnet, as a basic tool for deep learning. This neural network is obtained by making the following improvements to the U_Net convolutional neural network: (1) Introducing residual blocks to solve the degradation problem caused by the increase of network depth in the ordinary U_Net convolutional neural network. (2) Introducing residual learning to solve the gradient vanishing problem, which is used to achieve better backpropagation of errors.
[0030] Step 4: Obtain a training set or an existing test set to perform initial training on the neural network, resulting in a trained model. Considering both training speed and time overhead, determine the batch size and number of iterations for training the neural network.
[0031] Step 5: Substitute the noisy DAS signal processed in Step 2 into the neural network for training to obtain the noise spectrum in the DAS. Subtract the obtained noise spectrum from the original spectrum to obtain the denoised result.
[0032] The goal of establishing Dn-ResUnet in step 3 is to determine the relationship between the output noise data and the desired noise data as a loss function, which can be expressed as:
[0033] in, For loss function, For input The corresponding output at that time For the Frobenius norm, The training data represents the number of datasets (N), and the training data directly determines the neural network's ability to denoise noisy DAS signals. Therefore, different datasets need to be created for noise under different influencing conditions. To ensure the accuracy of the loss function, gradient descent is used to update the network parameters. .
[0034] The process of DAS signal noise reduction method is as follows: Figure 2 As shown, the first step is to process the obtained noisy DAS data into a one-dimensional array format that can be input into the neural network. The second step is to input the processed noisy DAS signal into the trained neural network to obtain the results.
[0035] The specific structure of the constructed noise reduction neural network is as follows: First, the processed DAS data is passed through a convolutional block (Convblock), then through an encoder, i.e., a 5-layer down-sampling residual block (Down-Resblock). The output of the Resblock group is input into the bottom convolutional block (Convblock). The bottom layer data is then transmitted layer by layer to the decoder, i.e., the up-sampling residual block (Up_Resblock). Finally, it is output through the output convolutional block (Out_Convblock). The output result is used to recover the noise spectrum of DAS using the method corresponding to step 2. Noise reduction is achieved by subtracting the noise spectrum from the obtained noisy DAS signal.
[0036] In step 3, each downsampling residual block in the denoising neural network structure corresponds to an upsampling residual block. Figure 3 The residual blocks shown have 5 layers, and the corresponding residual blocks in each layer are connected by residual connections. This prevents the influence of subsequent layers from being diminished due to a small gradient in one layer. The combined features contain both high-level features and fine-grained information, which helps to better reconstruct noise. The internal structure of the convolutional block is as follows: Figure 4As shown on the left, the input data is processed by convolutional layers (Conv), batch normalization layers (BN), and rectified linear units (ReLU) before output. In the convolutional layers, the input is multiplied by the feature matrix to extract important information from the data. The batch normalization layer makes the training data distribution more stable, effectively avoiding the gradient vanishing problem and further facilitating the network training process. The rectified linear unit, as the activation function, can increase the speed of gradient descent operations. The expression for the activation function of the convolutional layer is:
[0037] The internal structure of the residual block is as follows Figure 4 As shown, the result from the convolutional block or the residual block of the previous level is input into the residual block. The residual block contains two parallel branches. The first branch is the same as the convolutional block, passing through the convolutional layer, batch normalization layer, and rectified linear unit in series. The second branch first passes through the batch normalization layer and rectified linear unit before entering the convolutional layer. This process is repeated. The results from both the first and second branches are connected to the Add layer, and the result is added and output to the next layer. The upsampling residual block is the opposite operation of the residual block, except that an upsampling layer is added to increase the size of the feature map. The last part of the denoising neural network is the output convolutional block, whose internal structure is the opposite of the convolutional block. However, the activation function used is changed to the tanh function, and its expression is:
[0038] The tanh function makes the output less than 1 and greater than -1, which can better simulate noise and realize end-to-end transmission of DAS signals. Completing the above process is one iteration process. The result is retained and output. The specific neural network parameters are given in Table 1.
[0039] Table 1. Parameters of the noise reduction neural network
[0040] In step 4, during network training, 80% of the data is used to train the network, and 20% of the data is used to validate the network. An end-to-end training method is used to obtain the loss function. To improve training efficiency, only a small amount of data is used to optimize the network during each training iteration. When building the network, the batch size is 32, meaning 32 samples are used to optimize the denoising neural network each time. The Adam optimizer is used during training to balance training speed and memory usage. To determine the number of iterations, Mean Squared Error (MSE) and Peak Signal-to-Noise Ratio (PSNR) are used to measure the denoising performance during network training. MSE and PSNR are calculated as follows:
[0041]
[0042] in and These are the effective MRS signal and the signal obtained after denoising, respectively, where M is the signal length. During network training, if performance does not significantly improve after 5 iterations, this number of iterations is considered the optimal number, and the model is saved. This denoising neural network uses 27 iterations.
[0043] DAS dynamic strain monitoring system, such as Figure 1 As shown, the system includes a narrow-linewidth laser, an acousto-optic modulator, an erbium-doped fiber amplifier 1, a circulator, an erbium-doped fiber amplifier 2, a filter, a coupler 1, and a coupler 2, which are connected in sequence using sensing optical fibers. The circulator is connected to the signal acquisition optical fiber laid in the application scenario. The coupler 2 is connected to photodetector 1, photodetector 2, and photodetector 3, respectively. The three photodetectors are connected to the data acquisition card using data cables.
[0044] In another embodiment of the present invention, to facilitate better implementation of the method provided in the embodiments of the present invention, an apparatus based on the above method is also provided. The meanings of the terms used are the same as in the above method, and specific implementation details can be found in the description of the method embodiments.
[0045] Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of a device provided in an embodiment of the present invention, wherein the device may include a noise signal acquisition module 501, a noise estimation module 502, and a noise reduction module 503, wherein: The noise signal acquisition module 501 is used to acquire the DAS signal containing noise; The noise estimation module 502 is used to input the noisy DAS signal into the pre-trained denoising neural network model and output the noise estimation signal through the denoising neural network model. The denoising neural network model is a Dn-ResUnet network based on the U-Net architecture and introducing residual blocks. The noise reduction module 503 is used to subtract the noise estimation signal from the noisy DAS signal to obtain the noise-reduced DAS signal.
[0046] In some instances, the loss function of the denoising neural network model is: ,in, For loss function, For input The corresponding output at that time For the Frobenius norm, This represents the training data, where N is the number of datasets. These are network parameters.
[0047] In some instances, the denoising neural network model includes a convolutional block, an encoder path, a convolutional block, a decoder path, and an output convolutional block connected in sequence. The encoder path includes multiple cascaded downsampled residual blocks, and the decoder includes multiple cascaded upsampled residual blocks. Residual connections are provided between corresponding layers of the encoder and decoder.
[0048] In some instances, the apparatus also includes a preprocessing module for unfolding the spatiotemporal spectrum of the noisy DAS signal into a one-dimensional array.
[0049] In some instances, the apparatus further includes: a training module for constructing a training set using measured DAS signals mixed with various synthetic noises; inputting real noise signals from the training set into a denoising neural network model to obtain estimated noise signals; calculating the mean square error between the estimated noise signals and the corresponding real noise signals; using a gradient descent algorithm to backpropagate the gradient of the loss function to each layer of the network and update the network weight parameters; and optimizing the network to minimize the mean square error between the noise estimates and the real noise values, thereby obtaining a pre-trained denoising neural network model.
[0050] The above provides a detailed description of a method and apparatus for denoising DAS dynamic strain monitoring signals based on Dn-ResUnet, as provided in the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A method for denoising DAS dynamic strain monitoring signals based on Dn-ResUnet, characterized in that, include: Acquire noisy DAS signals; The noisy DAS signal is input into a pre-trained denoising neural network model, which is a Dn-ResUnet network based on the U-Net architecture and incorporating residual blocks. The noise estimation signal is output through the noise reduction neural network model. The noise-infused DAS signal is subtracted from the noise estimation signal to obtain the denoised DAS signal.
2. The method according to claim 1, characterized in that, The loss function of the denoising neural network model is: ,in, For loss function, For input The corresponding output at that time For the Frobenius norm, This represents the training data, where N is the number of datasets. These are network parameters.
3. The method according to claim 2, characterized in that, The noise reduction neural network model includes a convolutional block, an encoder path, a convolutional block, a decoder path, and an output convolutional block connected in sequence. The encoder path includes multiple cascaded downsampled residual blocks, and the decoder includes multiple cascaded upsampled residual blocks. Residual connections are provided between corresponding layers of the encoder and the decoder.
4. The method according to claim 1, characterized in that, Before inputting the noisy DAS signal into the pre-trained denoising neural network model, the method further includes: The spatiotemporal spectrum of the noisy DAS signal is unfolded into a one-dimensional array.
5. The method according to claim 1, characterized in that, The training of the noise reduction neural network model includes: The training set was constructed by mixing measured DAS signals with various synthetic noises. The real noise signal in the training set is input into the noise reduction neural network model to obtain the estimated noise signal; Calculate the mean square error between the estimated noise signal and the corresponding real noise signal; The gradient descent algorithm is used to backpropagate the gradient of the loss function to each layer of the network and update the network weight parameters. The optimization aims to minimize the mean square error between the noise estimate and the actual noise value of the network output, resulting in a pre-trained noise reduction neural network model.
6. A noise reduction device for DAS dynamic strain monitoring signals based on Dn-ResUnet, characterized in that, include: Noise signal acquisition module, used to acquire DAS signals containing noise; The noise estimation module is used to input the noisy DAS signal into a pre-trained denoising neural network model and output a noise estimation signal through the denoising neural network model. The denoising neural network model is a Dn-ResUnet network based on the U-Net architecture and introducing residual blocks. The noise reduction module is used to subtract the noise estimation signal from the noisy DAS signal to obtain the noise-reduced DAS signal.
7. The apparatus according to claim 6, characterized in that, The loss function of the denoising neural network model is: ,in, For loss function, For input The corresponding output at that time For the Frobenius norm, This represents the training data, where N is the number of datasets. These are network parameters.
8. The apparatus according to claim 7, characterized in that, The noise reduction neural network model includes a convolutional block, an encoder path, a convolutional block, a decoder path, and an output convolutional block connected in sequence. The encoder path includes multiple cascaded downsampled residual blocks, and the decoder includes multiple cascaded upsampled residual blocks. Residual connections are provided between corresponding layers of the encoder and the decoder.
9. The apparatus according to claim 6, characterized in that, The device further includes a preprocessing module for unfolding the spatiotemporal spectrum of the noisy DAS signal into a one-dimensional array.
10. The apparatus according to claim 6, characterized in that, The device further includes: a training module for constructing a training set by mixing measured DAS signals with various synthetic noises; inputting real noise signals from the training set into the denoising neural network model to obtain estimated noise signals; calculating the mean square error between the estimated noise signals and the corresponding real noise signals; using a gradient descent algorithm to backpropagate the gradient of the loss function to each layer of the network and update the network weight parameters; and optimizing the network to minimize the mean square error between the noise estimate and the real noise value, thereby obtaining a pre-trained denoising neural network model.