Multi-resolution convolutional neural network based on multi-span input for DAS signal denoising method

Abstract

This invention proposes a method for denoising DAS signals based on multi-scale input multi-resolution convolutional neural networks, belonging to the field of fiber optic sensing and signal processing technology. The method includes the following steps: Step S1: Acquire raw DAS data to generate multi-scale DAS data with multiple scales; Step S2: Input the multi-scale DAS data into a multi-scale convolutional neural network to achieve feature extraction, linear fusion (i.e., spatial scale alignment and weighted combination of multi-scale encoded features, followed by convolutional compression of channel number), and feature enhancement; perform multi-scale decoding and noise reconstruction on the enhanced features to obtain the denoised DAS signal. This invention can achieve DAS signal processing in complex environments, has good robustness, and can maintain consistency between the network's multi-scale modeling and the physical scale of the DAS system based on existing acquisition conditions and hardware configurations, improving the interpretability and engineering applicability of the denoising results.

Description

Technical Field

[0001] This invention belongs to the field of fiber optic sensing and signal processing technology, specifically relating to a multi-scale modeling and reconstruction method for distributed optical fiber acoustic wave sensing (DAS) data, and particularly to a method for denoising DAS signals based on a multi-scale input multi-resolution convolutional neural network. Background Technology

[0002] Distributed acoustic sensing (DAS) utilizes the characteristic of Rayleigh scattering phase variation with strain in optical fibers to achieve continuous seismic observation along the optical cable axis. It boasts advantages such as distributed deployment, low cost, and adaptability to harsh environments, and has been widely applied in scenarios such as vertical seismic profiles (VSP), active source and passive seismic monitoring. However, DAS data is often subject to various interferences, including random noise, attenuation noise, horizontal noise, checkerboard noise, and strong low-frequency noise. These noises highly overlap with the effective signal in the frequency band, and the significant difference in amplitude between transmitted and reflected waves makes weak event identification difficult. Furthermore, the high number of channels and high sampling rate of DAS further increase the engineering complexity of noise suppression.

[0003] Traditional seismic denoising methods such as filtering, sparse transform, mode decomposition, and low-rank decomposition are prone to signal distortion or insufficient denoising performance when noise and signal are mixed or noise statistical characteristics change significantly. Furthermore, they are computationally intensive and cannot meet the robust processing requirements of DAS-VSP data. Therefore, in recent years, DAS denoising research has gradually shifted towards data-driven methods such as convolutional neural networks (CNNs) to improve denoising performance by learning the nonlinear mapping between noise and effective signals. However, supervised methods are limited by the difficulty of obtaining truly noise-free data, weakly supervised methods usually require the assumption of noise independence, and self-supervised methods have high computational costs in high-noise, multi-channel scenarios and limitations in weak event recovery. To enhance the model's ability to represent complex noise, researchers have proposed various U-Net or multi-scale neural network designs, constructing mathematically significant multi-scale features through pooling, downsampling, different convolutional kernels, or parallel branches. Although these structures have achieved good performance in image and some seismic data processing, their scale originates from changes in network resolution and has no direct correspondence with the physical acquisition characteristics of DAS.

[0004] In DAS-VSP data acquisition, the gauge length determines the strain integration interval, significantly affecting spatial resolution, noise level, and effective bandwidth. Short gauge lengths offer high resolution but are noisier, while long gauge lengths have less noise but are prone to losing details; the two are physically complementary. However, existing deep learning methods generally only model single-gauge data, and the multi-scale features within the network often come only from mathematical downsampling, failing to utilize the differences in physical scale and noise statistics of different gauge lengths. Therefore, there is still a lack of a denoising method that can explicitly fuse multi-gauge DAS raw data in the network structure, fully utilize its physical multi-scale complementary information, and improve the ability to recover weak events under complex noise conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-resolution convolutional neural network (DAS) signal denoising method based on multi-scale input. By introducing multi-scale original input and constructing a multi-scale convolutional neural network, high-fidelity denoising and event reconstruction of DAS signals can be achieved, thus solving the technical problem that traditional single-scale or mathematical downsampling methods in the prior art cannot simultaneously guarantee the spatial resolution and signal-to-noise ratio of DAS signals.

[0006] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows:

[0007] A method for denoising DAS signals based on multi-scale input multi-resolution convolutional neural networks, the method comprising the following steps:

[0008] Step S1: Collect raw DAS data and generate multi-scale DAS data with multiple scales;

[0009] Step S2: Input the multi-scale DAS data into a multi-scale convolutional neural network to achieve feature extraction, feature fusion and feature enhancement of information at different scales, reconstruct noise and obtain the denoised DAS signal;

[0010] Specifically, step S2 includes:

[0011] Step S21: Input the multi-scale DAS data into the multi-scale coding module for feature extraction to obtain multi-scale coded features;

[0012] Step S22: Input the multi-scale encoded features into the cross-scale feature fusion module for linear fusion to obtain multi-scale fused features;

[0013] Step S23: Input the multi-scale fused features into the attention enhancement module to obtain multi-scale enhanced features;

[0014] Step S24: Input the multi-scale enhanced features into the decoding and reconstruction module, perform multi-scale decoding and noise reconstruction on the multi-scale enhanced features, and obtain the denoised DAS signal.

[0015] Further, in step S1, the acquired raw DAS data is represented as follows: ,in, Indicates fiber location The strain value at the location; Indicates the number of sampling points;

[0016] Utilizing the superposition property of differences, for any gauge length The DAS data is calculated as follows:

[0017]

[0018]

[0019] in, Indicates the minimum gauge length multiple; Indicates the minimum gauge length; Represents DAS data with arbitrary gauge length; This represents the i-th differential signal with any gauge length;

[0020] Based on the calculation method of arbitrary gauge length DAS data, three multi-scale DAS data with different gauge lengths can be obtained from the collected raw DAS data, namely the first gauge length DAS data. Second gauge length DAS data and third gauge length DAS data The gauge lengths of the multi-scale DAS data are respectively the first gauge length. Second gauge length Third gauge length ;and Corresponding to short, medium, and long gauge lengths;

[0021] Multiscale DAS data It is expressed as follows:

[0022] .

[0023] Furthermore, in step S21, the multi-scale coding module consists of three parallel coding branches, each of which comprises three convolutional layers and downsampling layers. The processing method of each coding branch is as follows:

[0024]

[0025]

[0026]

[0027] in, Represents the first in multi-scale DAS data One gauge DAS data, These represent the features of the first, second, and third convolutional layers, respectively. This indicates spatial downsampling with a stride of 2; Indicates basic convolutional layer operations;

[0028] First Gauge DAS Data Second gauge length DAS data and third gauge length DAS data The first encoded feature is obtained by inputting three parallel encoding branches respectively and performing feature extraction. Second coding features and third coding features First coding feature Second coding features and third coding features The third convolutional feature corresponding to the coding branch; the first coding feature Second coding features and third coding features It constitutes multi-scale features.

[0029] Furthermore, in step S22, linear fusion includes spatial scale alignment and weighted combination of multi-scale encoded features, followed by... The number of channels in convolution compression is expressed as follows:

[0030]

[0031]

[0032]

[0033] in, Indicates the fusion characteristics of the first channel; Indicates the fusion characteristics of the second channel; Indicates the fusion characteristics of the third channel; , , ,in, Indicates the height of the data; Indicates the data width; Indicates data dimensions; This indicates spatial scale alignment and weighted combination operations; express Stride convolution; Indicates convolution operation; first channel fusion feature Second channel fusion features Third channel fusion features It constitutes a multi-scale fusion feature.

[0034] Furthermore, in step S23, the attention enhancement module consists of three parallel feature enhancement branches, and each feature enhancement branch consists of eight groups of enhanced attention units;

[0035] Each enhanced attention unit will have a dimension of Multidimensional data compression into dimensions of The one-dimensional data is processed, and the attention output is obtained after feature extraction, as shown below:

[0036]

[0037] in, Indicates attention output, ; Indicates global average pooling; and The weight matrix is ​​trainable. It is the ReLU activation function. It is the sigmoid activation function; This indicates the fusion feature.

[0038] Simultaneously, each enhancement unit adaptively weights the attention output, thereby strengthening important features and increasing the dimension. The attention output is restored to a dimension of The enhanced features are represented as follows:

[0039]

[0040] in, Indicates enhanced features; Indicates the convolution operation;

[0041] First channel fusion characteristics Second channel fusion features Third channel fusion features The first enhanced attention unit of each of the three parallel feature enhancement branches of the attention enhancement module is input to the first enhanced attention unit. The output of the first enhanced attention unit is then used as the input of the second enhanced attention unit, and so on. The output of the last enhanced attention unit of the feature enhancement branch is the final enhanced feature output.

[0042] First channel fusion characteristics Second channel fusion features Third channel fusion features By inputting three parallel feature enhancement branches respectively, the first enhanced feature will be obtained. Second enhancement feature Third Enhancement Feature First Enhancement Feature Second enhancement feature Third Enhancement Feature This constitutes a multi-scale enhancement feature.

[0043] Furthermore, in step S24, the first enhanced feature Second enhancement feature Third Enhancement Feature The decoder in the input decoding and reconstruction module performs multi-scale decoding and noise reconstruction, and then performs spatial scale alignment and weighted combination, followed by... Convolution operation yields estimated noise. ; Using first gauge length DAS data The denoised DAS signal is obtained by subtracting the estimated noise. , means as follows:

[0044]

[0045]

[0046] in, This represents the denoised DAS signal; express Stepped deconvolution.

[0047] Furthermore, multi-scale DAS data with multiple gauge lengths can generate two or more gauge length combinations as needed.

[0048] Compared to existing technologies, this invention offers the following advantages: By constructing a multi-gauge physical multi-scale input based on the gauge length integration mechanism, and implementing cross-gauge feature alignment, fusion, and adaptive modulation within the network, the denoising process simultaneously utilizes the high-resolution detail information of short gauge lengths and the robust structure and low-noise characteristics of medium-to-long gauge lengths. This improves the recoverability of weak reflection events and enhances the accuracy of noise estimation. Because different gauge lengths carry complementary noise statistics and spatial resolution characteristics, this invention enables DAS signal processing in complex environments, exhibiting good robustness. Based on existing acquisition conditions and hardware configurations, it ensures that the network's multi-scale modeling remains consistent with the physical scale of the DAS system, further improving the interpretability and engineering applicability of the denoising results. This invention maintains the advantage of high spatial resolution with small gauge lengths while introducing high robustness with large gauge lengths, significantly improving the ability to identify deep and weak events and the stability of denoising. Furthermore, it adapts to continuous well site monitoring environments, improving robustness and deployability under complex conditions, and possesses significant engineering application value. Attached Figure Description

[0049] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a schematic diagram of the signal denoising method of the multi-resolution convolutional neural network DAS based on multi-scale input of the present invention.

[0051] Figure 2 This is a schematic diagram of signal acquisition in a DAS system.

[0052] Figure 3 This is a schematic diagram of the multi-scale convolutional neural network structure of the present invention.

[0053] Figure 4 This is a schematic diagram of the attention enhancement module structure of the present invention. Detailed Implementation

[0054] 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.

[0055] This invention proposes a multi-resolution convolutional neural network (DAS) signal denoising method based on multi-scale input, such as... Figure 1 As shown, the method includes the following steps:

[0056] Step S1: Collect raw DAS data and generate multi-scale DAS data with multiple scales.

[0057] Using the DAS system at the minimum gauge length Distributed strain data along the optical fiber in the wellbore were collected under the specified conditions to obtain DAS data.

[0058] minimum gauge length The raw DAS data obtained is represented as ,in, Indicates fiber location The strain value at the location; Indicates the number of sampling points, such as Figure 2 As shown.

[0059] A DAS system can acquire backscattered Rayleigh signals along an optical fiber. When the fiber is subjected to stress and other environmental noise, the intensity and phase of the acquired backscattered Rayleigh signal will change accordingly. This characteristic can be used to demodulate strain using the acquired intensity signal. Typically, the raw DAS data acquired... Due to the limitations of the system gauge length, the fiber optic location strain value on This indicates the minimum gauge length of the optical fiber. The average strain over this length is calculated as follows:

[0060]

[0061] in, This represents the backscattered Rayleigh signal; Indicates noise; For length variables The differential of represents the infinitesimal length element along the fiber axis; This represents the averaged backscattered Rayleigh signal; This represents the average noise level.

[0062] It is evident that a larger gauge length results in a wider integration interval, thereby introducing more unstable noise components and reducing spatial resolution. When the minimum gauge length... for At that time, the minimum gauge length DAS data can be obtained through differential calculation. ,in ,like Figure 3 As shown. Similarly, utilizing the superposition property of differences, for any gauge length... The DAS data is calculated as follows:

[0063]

[0064]

[0065] in, Indicates the minimum gauge length multiple; Indicates the minimum gauge length; Represents DAS data with arbitrary gauge length; This represents the i-th differential signal with any gauge length.

[0066] From the collected raw DAS data, three multi-scale DAS data with different gauge lengths can be obtained, namely the first gauge length DAS data. Second gauge length DAS data and third gauge length DAS data The gauge lengths of the multi-scale DAS data are respectively the first gauge length. Second gauge length Third gauge length .

[0067] That is, any gauge length is equivalent to performing a length of on the smallest gauge length signal. The sliding accumulation (convolution) method allows for the generation of multi-scale DAS data with arbitrary scales from raw DAS data. Multi-scale DAS data It is expressed as follows:

[0068]

[0069] in, Corresponding to short, medium, and long gauge lengths. Multi-scale DAS data. It provides high-resolution but noisy short gauge length information (first gauge length DAS data) Low-noise but lacking-detail long gauge length information (third gauge length DAS data) ), and the intermediate scale between the two (second gauge length DAS data) This constitutes a multi-scale input source with obvious physical complementarity, which can be used as the multi-scale input for subsequent neural networks.

[0070] Step S2: Input the multi-scale DAS data into the multi-scale convolutional neural network to achieve feature extraction, feature fusion and feature enhancement of information at different scales, reconstruct noise and obtain the denoised DAS signal.

[0071] A multi-scale convolutional neural network consists of a multi-scale encoding module, a cross-scale feature fusion module, an attention enhancement module, and a decoding and reconstruction module.

[0072] Step S21: Input the multi-scale DAS data into the multi-scale coding module for feature extraction to obtain multi-scale coded features.

[0073] The multi-scale coding module consists of three parallel coding branches, each consisting of three convolutional layers and downsampling layers, used to extract high-resolution features, structural features or steady-state low-frequency features at the corresponding scale.

[0074] The processing method for each coding branch is represented as follows:

[0075]

[0076]

[0077]

[0078] in, Represents the first in multi-scale DAS data One gauge DAS data, These represent the features of the first, second, and third convolutional layers, respectively. Indicates the height of the data; Indicates the data width; Indicates data dimensions; This indicates spatial downsampling with a stride of 2; This represents the basic convolutional layer operation.

[0079] This invention limits the convolution kernel size to 3×3 and the activation function to ReLU. This structure decouples the physical characteristics of different gauge lengths during the encoding stage. Detailed features primarily originate from the short gauge branch, while the long gauge branch provides a smoother, more stable, and lower-noise description.

[0080] First Gauge DAS Data Second gauge length DAS data and third gauge length DAS data The first encoded feature is obtained by inputting three parallel encoding branches respectively and performing feature extraction. Second coding features and third coding features First coding feature Second coding features and third coding features The third convolutional feature corresponding to the coding branch; the first coding feature Second coding features and third coding features It constitutes multi-scale features.

[0081] Step S22: Input the multi-scale encoded features into the cross-scale feature fusion module for linear fusion to obtain multi-scale fused features.

[0082] To fully utilize multi-scale encoded features, the cross-scale feature fusion module is divided into three channels. The first channel receives the first encoded feature. The second channel inputs the first encoded feature. Second coding features The third channel inputs the first encoded feature. Second coding features Third coding feature The three channels are linearly fused to obtain the multi-scale fused features.

[0083] The cross-scale feature fusion module enables the complementary use of multi-scale encoded features within the network. It linearly fuses these features to obtain multi-scale fused features. Linear fusion includes spatial scale alignment and weighted combination of the multi-scale encoded features, followed by... The number of channels compressed by strided convolution is expressed as follows:

[0084]

[0085]

[0086]

[0087] in, Indicates the fusion characteristics of the first channel; Indicates the fusion characteristics of the second channel; This indicates the fusion characteristics of the third channel. , , ,in, This indicates spatial scale alignment and weighted combination operations; express Stride convolution; This represents the convolution operation. The first channel fuses features. Second channel fusion features Third channel fusion features It constitutes a multi-scale fusion feature.

[0088] The cross-scale feature fusion module can utilize the coding features of three physical scales on each channel to form a "real multi-scale" representation consistent with the physical gauge length within the network, thereby enhancing the network's ability to identify weak reflections, attenuated waves, and low-frequency background noise.

[0089] Step S23: Input the multi-scale fused features into the attention enhancement module to obtain multi-scale enhanced features.

[0090] To automatically identify the importance of multi-scale fused features, this invention introduces an EA attention mechanism on these features. The attention enhancement module consists of three parallel feature enhancement branches, each composed of eight Enhanced Attention (EA) units. Figure 3 As shown, the enhanced attention unit is based on a "compression-excitation-recalibration" mechanism, which adaptively adjusts the weights of different channels.

[0091] Each enhanced attention unit will have a dimension of Multidimensional data compression into dimensions of One-dimensional data, and after feature extraction, the attention output is obtained, such as Figure 4 As shown, it represents the following:

[0092]

[0093] in, Indicates attention output, ; Indicates global average pooling; and The weight matrix is ​​trainable. It is the ReLU activation function. It is the sigmoid activation function; This indicates the fusion feature.

[0094] Simultaneously, each enhancement unit adaptively weights the attention output, thereby strengthening important features and increasing the dimension. The attention output is restored to a dimension of The enhanced features are represented as follows:

[0095]

[0096] in, Indicates enhanced features; This indicates a convolution operation.

[0097] First channel fusion characteristics Second channel fusion features Third channel fusion features The first enhanced attention unit of each of the three parallel feature enhancement branches of the attention enhancement module is input to the first enhanced attention unit. The output of the first enhanced attention unit is then processed by the enhanced attention unit and used as the input of the second enhanced attention unit. This process continues until the output of the last enhanced attention unit of the feature enhancement branch is the final enhanced feature output.

[0098] First channel fusion characteristics Second channel fusion features Third channel fusion features By inputting three parallel feature enhancement branches respectively, the first enhanced feature will be obtained. Second enhancement feature Third Enhancement Feature First Enhancement Feature Second enhancement feature Third Enhancement Feature This constitutes a multi-scale enhancement feature.

[0099] Step S24: Input the multi-scale enhanced features into the decoding and reconstruction module, perform multi-scale decoding and noise reconstruction on the multi-scale enhanced features, and obtain the denoised DAS signal.

[0100] First Enhancement Feature Second enhancement feature Third Enhancement Feature Passing through After strided deconvolution, the input is given to the decoder for multi-scale decoding and noise reconstruction. Through spatial scale alignment and weighted combination, the estimated noise is obtained. Using first gauge length DAS data The denoised DAS signal is obtained by subtracting the estimated noise. , means as follows:

[0101]

[0102]

[0103] in, This represents the denoised DAS signal; express Stepped deconvolution.

[0104] This invention proposes a multi-resolution convolutional neural network (DAS) signal denoising method based on multi-scale input. It utilizes the sliding accumulation of the minimum scale difference signal to generate DAS data at different scales, enabling the network input to directly contain real physical scale information, rather than relying on mathematical downsampling within the network. A multi-branch convolutional coding structure corresponding one-to-one with the physical scale is constructed. High-resolution, high-noise, and medium-scale structures, as well as low-noise and robust low-frequency features, are extracted for short, medium, and long scales respectively, achieving "physical scale-driven multi-scale coding" for the first time. A cross-scale feature alignment and layer-by-layer fusion mechanism is proposed. Scale alignment and fusion of features at different scales are performed at each depth layer of the network, preserving and enhancing complementary information between physical scales in the feature space, forming a consistent "physical scale representation" within the network. An EA attention modulation mechanism suitable for physical multi-scale features is introduced. Based on the multi-scale coding module, the response intensity of features at different scales is adaptively adjusted, strengthening weak reflection events and suppressing low-frequency noise components, achieving dynamic selection of cross-scale features. A noise construction and signal reconstruction method based on physical scale fusion is also proposed. The noise term is estimated by joint modeling of deep and shallow features, and the final denoising reconstruction is completed by combining the original signal with short gauge length, so that the noise estimation process itself contains real physical scale information.

[0105] Based on the method implementation, the multi-resolution convolutional neural network based on multi-gauge input proposed in this invention can be flexibly configured according to different well site conditions, data quality, and computing resources. The following optional embodiments are all equivalent forms of this invention and do not constitute a limitation on the scope of protection of this invention.

[0106] (1) Optional configuration of the number of gauge lengths. In different application scenarios, the multi-gauge input of this invention can adopt two or more combinations of gauge lengths. Including but not limited to the following methods:

[0107] Dual gauge length configuration: You can select either a short gauge length or a long gauge length to reduce the number of network branches and reduce computational load;

[0108] Triple gauge length configuration: The combination of short, medium and long gauge lengths can achieve a better balance between resolution and noise robustness, and is the typical recommended configuration.

[0109] Four or more gauge length configurations: Four or more gauge length inputs can be generated based on the acquisition parameters to enhance the depth model’s coverage of the physical scale.

[0110] In the above configuration, the gauge length can be equal interval, multiple interval, or adaptively determined according to the well field coupling situation, all of which are within the scope of this invention.

[0111] (2) Optional configuration of encoder branch structure. Depending on the gauge difference, data quality, and model size requirements, the multi-branch convolutional encoder of this invention can adopt the following alternative structures:

[0112] Completely independent multi-branch structure: Each gauge corresponds to an independent convolutional encoder, which is used to preserve feature differences at different physical scales to the greatest extent.

[0113] Partially shared multi-branch structure: shallow convolutional kernels can be shared for inputs with similar calibration distances, while branch structures are set in deeper layers to balance information differences and the number of parameters;

[0114] A fully shared backbone network: multi-caliber data is used as multi-channel input, features are extracted by a single-path encoder, and a caliber channel attention module is set in the intermediate layer.

[0115] Whether to share convolutional layers can be freely chosen based on actual engineering needs.

[0116] (3) Optional configuration of cross-gauge fusion strategy. The cross-gauge fusion of the present invention can be carried out in any of the following ways:

[0117] Channel splicing and fusion: Multi-calibration features are directly spliced ​​in the channel dimension, and then compressed and mapped through convolution, which is the default solution of this invention;

[0118] Weighted summation fusion: Learnable weights are assigned to features with different gauge lengths to achieve a lighter fusion structure;

[0119] Attention-guided fusion: Introducing spatial or channel attention on top of splicing or weighted fusion to enhance key scale information.

[0120] The choice of integration method can be flexibly adjusted based on data characteristics and resource conditions.

[0121] (4) Optional configuration of the attention module. In addition to the EA module, the present invention can also employ other attention forms, including but not limited to: CA (Channel Attention) and SA (Spatial Attention). All of the above modules can be embedded in the cross-scale fused features to enhance physical multi-scale information.

[0122] (5) Optional configuration of convolution type and dimension. Based on the organization of DAS data, the convolution calculation of this invention can adopt: one-dimensional convolution (1D Conv): extracting sequence features along the temporal or spatial direction; two-dimensional convolution (2D Conv): treating the temporal-channel as a two-dimensional structure; depthwise separable convolution: reducing the number of network parameters and computational cost; dilated convolution: expanding the receptive field for structural recovery of weak events. The convolution type does not change the technical features of this invention and can be used as an equivalent replacement.

[0123] (6) Optional configuration of noise construction and output header. The noise estimation header structure of this invention can be adjusted according to model complexity and real-time requirements, including: a single convolutional output layer; a multi-layer convolutional stacked output header; a noise reconstruction module with residual structure; and an attention-based noise decoupling module; the core function of this module is to output noise terms. The specific structure does not affect the essence of the invention.

[0124] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for denoising DAS signals based on multi-scale input multi-resolution convolutional neural networks, characterized in that, The method includes the following steps: Step S1: Collect raw DAS data and generate multi-scale DAS data with multiple scales; Step S2: Input the multi-scale DAS data into a multi-scale convolutional neural network to achieve feature extraction, feature fusion and feature enhancement of information at different scales, reconstruct noise and obtain the denoised DAS signal; Specifically, step S2 includes: Step S21: Input the multi-scale DAS data into the multi-scale coding module for feature extraction to obtain multi-scale coded features; Step S22: Input the multi-scale encoded features into the cross-scale feature fusion module for linear fusion to obtain multi-scale fused features; Step S23: Input the multi-scale fused features into the attention enhancement module to obtain multi-scale enhanced features; Step S24: Input the multi-scale enhanced features into the decoding and reconstruction module, perform multi-scale decoding and noise reconstruction on the multi-scale enhanced features, and obtain the denoised DAS signal; In step S1, the raw DAS data obtained is represented as follows: ,in, Indicates fiber location The strain value at the location; Indicates the number of sampling points; Using the superposition property of differences, for any gauge length The DAS data is calculated as follows: in, Indicates the minimum gauge length multiple; Indicates the minimum gauge length; Represents DAS data with arbitrary gauge length; This represents the i-th differential signal with any gauge length; Based on the calculation method of arbitrary gauge length DAS data, three multi-scale DAS data with different gauge lengths can be obtained from the collected raw DAS data, namely the first gauge length DAS data. Second gauge length DAS data and third gauge length DAS data The gauge lengths of the multi-scale DAS data are respectively the first gauge length. Second gauge length Third gauge length ;and Corresponding to short, medium, and long gauge lengths; Multiscale DAS data It is expressed as follows: In step S22, linear fusion includes spatial scale alignment and weighted combination of multi-scale encoded features, followed by... The number of channels compressed by strided convolution is expressed as follows: in, Indicates the fusion characteristics of the first channel; Indicates the fusion characteristics of the second channel; This indicates the fusion characteristics of the third channel; among which, This indicates spatial scale alignment and weighted combination operations; express Stride convolution; Indicates convolution operation; first channel fusion feature Second channel fusion features Third channel fusion features This constitutes a multi-scale fusion feature; Indicates the first coding feature, Indicates the second coding feature and This represents the third coding feature.

2. The signal denoising method based on multi-scale input multi-resolution convolutional neural network (DAS) according to claim 1, characterized in that, In step S21, the multi-scale coding module consists of three parallel coding branches. Each coding branch consists of three convolutional layers and downsampling layers. The processing method of each coding branch is as follows: in, Represents the first in multi-scale DAS data One gauge DAS data, These represent the features of the first, second, and third convolutional layers, respectively. Indicates the height of the data; Indicates the data width; Indicates data dimensions; This indicates spatial downsampling with a stride of 2; Indicates basic convolutional layer operations; First gauge length DAS data Second gauge length DAS data and third gauge length DAS data The first encoded feature is obtained by inputting three parallel encoding branches respectively and performing feature extraction. Second coding features and third coding features First coding feature Second coding features and third coding features The third convolutional feature corresponding to the coding branch; the first coding feature Second coding features and third coding features It constitutes multi-scale features.

3. The signal denoising method based on multi-scale input multi-resolution convolutional neural network (DAS) according to claim 2, characterized in that, In step S23, the attention enhancement module consists of three parallel feature enhancement branches, and each feature enhancement branch consists of eight groups of enhanced attention units. Each enhanced attention unit will have a dimension of Multidimensional data compression into dimensions of The one-dimensional data is processed, and the attention output is obtained after feature extraction, as shown below: in, Indicates attention output, ; Indicates global average pooling; and The weight matrix is ​​trainable. It is the ReLU activation function. It is the sigmoid activation function; Indicates fusion characteristics; Simultaneously, each enhancement unit adaptively weights the attention output, thereby strengthening important features and increasing the dimension. The attention output is restored to a dimension of The enhanced features are represented as follows: in, Indicates enhanced features; Indicates the convolution operation; First channel fusion characteristics Second channel fusion features Third channel fusion features The first enhanced attention unit of each of the three parallel feature enhancement branches of the attention enhancement module is input to the first enhanced attention unit. The output of the first enhanced attention unit is then used as the input of the second enhanced attention unit, and so on. The output of the last enhanced attention unit of the feature enhancement branch is the final enhanced feature output. First channel fusion characteristics Second channel fusion features Third channel fusion features By inputting three parallel feature enhancement branches respectively, the first enhanced feature will be obtained. Second enhancement feature Third Enhancement Feature First Enhancement Feature Second enhancement feature Third Enhancement Feature This constitutes a multi-scale enhancement feature.

4. The signal denoising method based on multi-scale input multi-resolution convolutional neural network (DAS) according to claim 3, characterized in that, In step S24, the first enhanced feature Second enhancement feature Third Enhancement Feature The decoder in the input decoding and reconstruction module performs multi-scale decoding and noise reconstruction, and then performs spatial scale alignment and weighted combination, followed by... Convolution operation yields estimated noise. ; Using first gauge length DAS data The denoised DAS signal is obtained by subtracting the estimated noise. , means as follows: in, This represents the denoised DAS signal; express Stepped deconvolution.

5. The signal denoising method based on multi-scale input multi-resolution convolutional neural network (DAS) according to claim 1, characterized in that, Multi-scale DAS data with multiple gauge lengths can generate two or more combinations of gauge lengths as needed.

Images