CN-121441323-B - DAS data compression method based on event segmentation labeling

CN121441323BCN 121441323 BCN121441323 BCN 121441323BCN-121441323-B

Abstract

The invention discloses a DAS data compression method based on event segmentation labeling, which belongs to the technical field of DAS data compression and comprises the steps of constructing a data set D, constructing a DAS data compression network comprising a construction coding module, a side quantization branch, a side decoding branch, a main quantization branch and a main decoding branch, constructing a loss function L, training the DAS data compression network to obtain a DAS data compression model M DAS , inputting samples in a verification set into M DAS to select an optimal event threshold, acquiring DAS data to be compressed, compressing and decompressing by using M DAS , and filtering by using the optimal event threshold to obtain a corresponding filtered sample. According to the invention, the DAS data compression and noise reduction tasks are realized at the same time, the noise reduction tasks are used for guiding the compression model to compress target events in a key way, and noise compression with low compression efficiency is avoided, so that the compression ratio of a far-beyond single compression model is realized, and the storage and transmission efficiency and the subsequent processing efficiency of DAS data are greatly improved.

Inventors

WANG HONGHUI
LIU TONG
YANG XIKE
WANG XIANG
REN JIZHOU
HE SULAN
YAO GUANGLE

Assignees

成都理工大学

Dates

Publication Date: 20260505
Application Date: 20251231

Claims (10)

1. The DAS data compression method based on event segmentation labeling is characterized by comprising the following steps of: s1, constructing a data set D; Collecting DAS data comprising W×H data points, wherein part of the data points are target events, W, H are the single-channel sampling points and the total number of channels of the DAS data respectively; Normalizing each DAS data channel by channel, and then carrying out data enhancement to generate DAS data samples; Generating a pseudo-pure event sample for each DAS data sample, wherein the i-th pseudo-pure event sample of DAS data sample x i is i , forming (x i ,gt i ) sample pairs, forming a data set D by all sample pairs, and dividing a training set, a verification set and a test set; s2, constructing a DAS data compression network, which comprises the steps S21-S22; s21, presetting a first channel number C1 and a second channel number C2, wherein C1 is smaller than C2; s22, constructing an encoding module, a side quantization branch, a side decoding branch, a main quantization branch and a main decoding branch; The coding module comprises a first convolution module, a first compression module, a first noise reduction module, a second compression module and a second noise reduction module which are sequentially arranged, and is used for inputting x i , lifting the number of channels to C1 through convolution operation, and generating coding characteristics y i with the number of channels of C2 through two-round staggered compression and noise reduction operation; The side quantization branch is used for performing super-parametric coding on y i to obtain a parameter distribution characteristic z i , and then quantizing the parameter distribution characteristic z i into discrete values Arithmetic coding into a binary bit stream z string ; The side decoding branch is used for inputting z string , sequentially carrying out arithmetic decoding, inverse quantization and super-parameter decoding operation on the side decoding branch, and outputting super-parameter characteristic params for predicting y i Gaussian distribution; the main quantization branch is used for quantizing y i into discrete values And then arithmetic encoded with params into a binary bit stream y string , wherein 、 The number of channels of (2) is C2; The main decoding branch is used for arithmetic decoding of y string back according to params Will be Inverse quantization to continuous predictive coding features Re-decoding into reconstructed data corresponding to x i ; S3, constructing a loss function L of the DAS data compression network; , Wherein, L bpp is compression efficiency, L mse is noise reduction mass loss, L fea is characteristic loss based on an output structure of x i 、gt i through a first noise reduction module and a second noise reduction module, and alpha and beta are weights of L mse and L fea respectively; S4, training the DAS data compression network to be converged by using the training set and the verification set to obtain a DAS data compression model M DAS ; S5, inputting samples in the verification set into M DAS to obtain corresponding reconstruction data, and forming a first reconstruction data set D1; S6, presetting an event threshold value, filtering the samples in the step D1 to obtain filtered samples, forming a second reconstruction data set D2, and calculating the peak signal-to-noise ratio of the D2; s7, repeating the steps S5-S6 by using different event thresholds, and taking the event threshold with the lowest peak signal-to-noise ratio as an optimal event threshold; s8, compression and decompression: and in the decompression stage, M DAS obtains corresponding reconstruction data according to y string and z string , and then filters the reconstruction data through an optimal event threshold value to obtain corresponding filtering samples.
2. The DAS data compression method based on event segmentation labeling of claim 1, wherein in S1, the target event is a microseism event; When normalizing channel by channel in S1, generating a normalized value for the w data point d wh of the h channel in DAS data according to the following formula ; , Wherein, max (d h )、Min(d h ) is the maximum value and the minimum value of the data points in the H channel, W is more than or equal to 1 and less than or equal to W, H is more than or equal to 1 and less than or equal to H; The data enhancement includes randomly adding a first noise, panning.
3. The DAS data compression method based on event segmentation labeling of claim 1, wherein in S1, the pseudo-pure event sample method for generating DAS data samples is: Sa1, obtaining DAS data samples containing W multiplied by H sampling points; And Sa2, carrying out semantic segmentation labeling on the DAS data samples, reserving data points with marked categories as target events in the DAS data samples, and changing the values of the rest data points to 0.5 to obtain pseudo-pure event samples.
4. The DAS data compression method based on event segmentation annotation according to claim 1, wherein the encoding module comprises a first convolution module, a first compression module, a first noise reduction module, a second compression module and a second noise reduction module, which are sequentially arranged; The first convolution module is configured to boost the input number of x i channels to a preset value to obtain a first convolution feature ; The first compression module comprises a first bneck layer, a second bneck layer and a third bneck layer which are stacked for inputting Outputting a first compression characteristic ; The first noise reduction module comprises four expansion convolution layers with different expansion rates, a fourth bneck layer, a CBAM layer and a fifth bneck layer, Respectively performing expansion convolution on 4 expansion convolution layers, splicing the expansion convolution layers on the channel layers, sequentially sending the expansion convolution layers into a fourth bneck layer, a CBAM layer and a fifth bneck layer, and outputting a first noise reduction feature ; The second compression module structure is the same as the first compression module structure and is used for inputting Outputting a second compression characteristic The second noise reduction module has the same structure as the first noise reduction module and is used for inputting The encoding feature y i is output.
5. The DAS data compression method based on event segmentation labeling of claim 1, wherein the side quantization branch comprises a super parameter coding module, a side quantization layer and a side arithmetic coding layer; The super-parameter coding module comprises 5 stacked convolution layers, the convolution kernel size of each convolution layer is 3, the number of input channels is 96, the number of output channels is 96, the first 4 convolution layers use LeakyRelu to activate functions, the step length of the 3 rd convolution layer and the 5 th convolution layer is 2, the rest step length is 1, the super-parameter coding module is used for inputting y i , and after super-parameter coding of the 5 th convolution layer, the output parameter distribution characteristic z i is output; the side quantization layer is used for quantizing z i into discrete values by adding uniformly distributed second noise in the training stage and by quantizing function in the reasoning stage ; The side arithmetic coding layer will be based on entropy coding Encoded as a binary bit stream z string .
6. The DAS data compression method based on event segmentation labeling of claim 1, wherein the side decoding branches comprise a side arithmetic decoding layer, a side inverse quantization layer and a super-parameter decoding module; the side arithmetic decoding layer is used for decoding z string back to discrete values ; The side inverse quantization layer is used for processing Restoring to continuous value to obtain predicted parameter distribution characteristics ; The super-parameter decoding module comprises a first convolution layer, a first up-sampling layer, a second convolution layer, a second up-sampling layer and a third convolution layer which are sequentially connected and used for inputting Sequentially performing staggered convolution and up-sampling operation to obtain the super-parameter characteristic params with the channel number of 2×C2.
7. The DAS data compression method based on event segmentation labeling of claim 1, wherein the main quantization branch comprises a main quantization layer and a main arithmetic coding layer; The main quantization layer is used for quantizing y i into discrete values by adding uniformly distributed second noise in the training stage and by using quantization function in the reasoning stage ; The main arithmetic coding layer is used for dividing the data into groups according to params The arithmetic coding is a binary bit stream y string .
8. The DAS data compression method based on event segmentation annotation of claim 1, wherein the primary decoding branch comprises a primary arithmetic decoding layer, a primary inverse quantization layer and a decoder; The main arithmetic decoding layer is used for arithmetic decoding of y string back according to params ; The primary dequantization layer dequantizes the dequantized coded data into a continuous predictive coding feature ; The decoder is used for decoding Decoding into reconstructed data corresponding to x i The whole is of an 8-layer structure, the 4 bneck layers and the 4 upsampling layers are connected in sequence in a staggered mode, and the first layer of the decoder is bneck layers.
9. The DAS data compression method based on event segmentation labeling of claim 1, wherein L bpp 、L mse 、L fea is calculated according to the following equation: , , , Where Data comp,i is the sum of the number of bits of y string and z string corresponding to x i , 、 The output characteristics of x i and gt i after passing through the first noise reduction module, 、 The output characteristics of x i and gt i after passing through the second noise reduction module are respectively.
10. The DAS data compression method based on event segmentation labeling of claim 1, wherein in S6, the method of filtering the samples with an event threshold θ is; , wherein, the value of the w data point of the h channel in the reconstructed data is the value of the w data point of the h channel of the filtered sample.

Description

DAS data compression method based on event segmentation labeling Technical Field The invention relates to the technical field of DAS data compression, in particular to a DAS data compression method based on event segmentation labeling. Background The distributed optical fiber acoustic sensing technology (Distributed acoustic sensing, DAS) is a monitoring technology using an optical fiber as a monitoring unit and a signal transmission medium, and monitors environmental strain information by utilizing backward rayleigh scattered light generated by elastic collision of photons and impurity particles in the optical fiber. When external sound waves are transmitted to the sensing optical fiber, the local refractive index of the optical fiber changes to cause the phase change of backward Rayleigh scattered light, and the strain information of the action and the optical fiber can be obtained by demodulating the phase. The DAS has the advantages of full distributed sensing, long monitoring distance, high sensitivity, electromagnetic interference resistance, low requirement on layout environment and the like, and is widely applied to the fields of earthquake monitoring, oil and gas pipeline leakage monitoring, hydraulic fracturing monitoring, submarine detection, railway and geological disaster monitoring and the like. With the widespread use of DAS technology in numerous fields, the massive data and complex environmental noise generated during operation present serious challenges to the system's practicality. The DAS system senses external vibration signals through the phase change of Rayleigh scattered light in optical fibers, the sampling rate is generally up to thousands of hertz, a single optical fiber can generate several TB-level data every day, the storage cost is increased sharply, the real-time transmission efficiency is limited, and particularly in remote areas or long-term monitoring scenes, the high redundancy and low information density of original data are needed to achieve feature extraction and data reduction through an intelligent compression algorithm. Meanwhile, DAS signals are susceptible to environmental temperature drift, mechanical coupling noise, electromagnetic interference, optical fiber intrinsic polarization drift and other factors, noise levels are often overlapped with target signals and even buried in weak anomalies, for example, early features of pipeline micro leakage or geological deformation can be submerged, so that false alarm or detection omission risks are increased rapidly. Because of the poor statistical properties of noise data, more bits are required to compress the noise data, thereby seriously affecting the efficiency of DAS data compression. The existing DAS data compression algorithm comprises traditional Huffman coding, arithmetic coding, discrete Cosine Transform (DCT), discrete Wavelet Transform (DWT) and other deep learning methods based on convolutional neural compression network (CAE) and cyclic neural compression network (RAE) of a self-encoder, and the like, and can compress DAS data to a certain extent, however, the existing DAS compression algorithm cannot distinguish events and noise in the DAS data, the same compression weight is executed on data points at all positions, most of the DAS data is noise, the duty ratio of DAS events is relatively small, and the noise is usually compressed by more bits due to weak statistical characteristics, so that the compression ratio is lower. Noun interpretation: The bneck layer is a core module in the existing convolutional neural network MobileNet, and the bneck layer completes the flow of dimension increasing, DW convolution (depth separable convolution), SE attention reducing, dimension reducing and residual error, namely, a whole feature refining pipeline of dimension increasing manufacturing expression space, DW convolution local feature extracting, channel calibration, feature dimension reducing, redundancy discarding and residual error remaining effective information. Layer CBAM (Convolutional Block Attention Module) is a mixed attention mechanism that enhances the expressive power of the feature map by combining channel attention and spatial attention. Disclosure of Invention The invention aims to provide the DAS data compression method based on event segmentation labeling, which can solve the problems and realize DAS data compression and noise reduction tasks simultaneously through a deep learning neural network, so that a compression model can identify DAS events and noise, and remove most of noise through the noise reduction tasks, and a compression module only compresses DAS event parts, thereby greatly improving DAS compression efficiency. In order to achieve the purpose, the technical scheme adopted by the invention is that the DAS data compression method based on event segmentation labeling comprises the following steps: s1, constructing a data set D; Collecting DAS data comprising W×H data points, wherein