Underground cable surrounding temperature field inversion method and device based on RDC-SE-Net

Abstract

The invention relates to a buried cable surrounding temperature field inversion method and device based on RDC-SE-Net, and the method comprises the steps: obtaining sound velocity time distance data and label temperature data; sound velocity time distance data is processed by the first feature operation module, the first SE module, the residual convolution cavity module, the first connection module, the second connection module, the second SE module and the second feature operation module in sequence, then simplified data is output, the simplified data and label temperature data are processed by the Loss operation module, and inversion temperature data are output; by adopting the technical scheme, the inversion precision is high.

Description

Technical Field

[0001] The present invention belongs to the technical field of temperature field detection around buried cables, and particularly relates to a method for inverting the temperature field around buried cables based on RDC-SE-Net. Background Technique

[0002] The accurate inversion of the temperature field around cables is a key link in the state assessment and fault warning of high-voltage cables. The heat generated during cable operation is conducted to the surrounding environment through the insulation layer, sheath and external medium. The temperature field distribution characteristics not only reflect the heat dissipation efficiency, current-carrying capacity margin and insulation aging degree of the cable body, but also are closely related to the thermal physical parameters of the laying environment (such as soil thermal conductivity, thermal interference of adjacent pipelines, etc.). Research shows that local temperature anomalies in the cable trench may be caused by soil cavities, water intrusion or external mechanical damage. However, traditional monitoring methods are limited to measuring the temperature of the cable surface or core wire, and it is difficult to capture such potential risks. In addition, the non-uniform distribution of the temperature field will exacerbate the thermal stress of the insulating material, leading to chain failures such as partial discharge or thermal breakdown. Due to the long laying path and complex spatial structure of the cable, directly deploying a dense temperature sensor network faces problems such as high cost, weak anti-interference ability and insufficient long-term stability. There is an urgent need for non-invasive and high-precision temperature field inversion technology.

[0003] The current temperature field inversion methods are mainly divided into two categories: the forward calculation method based on physical models and the inverse reconstruction method based on data-driven. The forward calculation method deduces the temperature distribution by establishing a multi-physical field coupling model of the cable-environment (such as the finite element heat conduction equation) and combining boundary conditions. However, its accuracy highly depends on the accuracy of material parameters and it has poor adaptability to complex working conditions such as soil stratification heterogeneity and dynamic heat source interference. The inverse reconstruction method uses sparse temperature measurement point data and adopts an optimization algorithm to inversely deduce the global temperature field. However, the existing methods have significant defects: the linear inversion based on the least squares is difficult to handle the strongly non-linear heat conduction process. In addition, most of the existing technologies monitor the cable itself and do not consider actual interference factors such as transient thermal disturbances caused by the cable surrounding environment, resulting in problems such as limited monitoring range (confined to the cable diameter, generally within 10 cm of the cable area) and low spatial resolution (the error generally exceeds ±3°C). For example, the spatial resolution of Reference 1 is ±3°C. Reference 1: Zhao H, Zhang Z, Yang Y, et al. Real-time reconstruction of temperature field for cable joints based on inverse analysis[J]. International Journal of Electrical Power & Energy Systems, 2023, 144: 108573. Therefore, developing a new high-precision inversion technology for the temperature field around buried cables has urgent engineering value. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to make up for the deficiencies of the existing technology and provide a method for inverting the temperature field around buried cables based on RDC-SE-Net.

[0005] To solve the above technical problems, the technical solution of the present invention is as follows: A method for inverting the temperature field around buried cables based on RDC-SE-Net, comprising the following steps: Step S1: Determine a straight line L parallel to the buried cable, and the distance between the straight line L and the buried cable is not greater than 1 m; Step S2: Determine C1 equally spaced speaker moving placement points on the straight line L. When the speaker moves to each moving placement point, it emits sound waves, and a distributed fiber optic acoustic sensing system is used to collect a set of acoustic velocity time-distance data with dimensions of M×N, and a total of acoustic velocity time-distance data with dimensions of C1×M×N is obtained. The number of time data collected = M, and the number of distance data collected = N; Step S3: In the plane formed by the straight line L and the buried cable, determine M×N temperature acquisition points according to the determinant, where the number of rows along the axial direction of the buried cable is M, and the number of columns perpendicular to the axial direction of the buried cable is N. Use temperature sensors to collect 1 temperature value at each temperature acquisition point as the labeled temperature, and obtain labeled temperature data with a dimension of 1×M×N in total; Step S4: Input the sound speed time distance data obtained in Step S2 into the first feature operation module. The data is successively subjected to operations with a 3×3×32 convolution kernel, a 3×3×64 convolution kernel, batch normalization, a 3×3×64 convolution kernel, and a 3×3×32 convolution kernel, and high-level feature data is output; Step S5: Use the first SE module to perform compression processing on the high-level feature data obtained in Step S4, and output compressed data; Step S6: Respectively use the first dilated convolution sub-module, the second dilated convolution sub-module, and the third dilated convolution sub-module to perform operations on the compressed data obtained in Step S5. The dilation rates of these 3 sub-modules are all 2, the strides are all 1, and the padding rates are all 1; the first dilated convolution sub-module outputs high-level feature data with a dilated convolution kernel of 3, the second dilated convolution sub-module outputs high-level feature data with a dilated convolution kernel of 5, and the third dilated convolution sub-module outputs high-level feature data with a dilated convolution kernel of 8; use the first connection module to splice the three groups of high-level feature data with dilated convolution kernels of 3, 5, and 8 and output the first spliced data, perform residual operation on the compressed data through the residual operation sub-module and output the residual operation data, use the second connection module to connect the residual operation data and the first spliced data and output the second spliced data, use the second SE module to perform compression processing on the second spliced data and output the second compressed data, use the second feature operation module to perform regularization operation on the second compressed data, and then calculate the regularized data using a 3×3×1 convolution with a padding rate of 1 and a stride of 1, and output the simplified data; Step S7: Input the simplified data obtained in Step S6 and the labeled temperature data obtained in Step S3 into the Loss operation module. Use the mean square error loss function to calculate the Loss value between the simplified data and the labeled temperature data. Use the Adam algorithm optimizer to set the number of training epochs. When Loss≤0.002, the training epoch ends, and the trained simplified data is output. The trained simplified data is the inversion temperature data.

[0006] Further, in Step S2, the frequency of the sound wave emitted by the speaker is 1 - 200 Hz, and the amplitude ≥ 5Vpp; C1≥10, M≥200, N≥200.

[0007] Further, in step S4, the sound speed time distance data obtained in step S2 is input into the first feature operation module. First, a 3×3×32 convolution kernel is used for operation and primary feature data is output. The size of the primary feature data is C2×(M - 1)×(N - 1), where C2 = 32. A 3×3×64 convolution kernel is used to operate on the primary feature data and secondary feature operation data is output. The size of the secondary feature operation data is C3×(M - 2)×(N - 2), where C3 = 64. Batch normalization operation is performed on the secondary feature operation data and normalized padding data is output. The padding size is 4, and the size of the normalized padding data is C3×(M + 2)×(N + 2). A 3×3×64 convolution kernel is used to operate on the normalized padding data and sub - senior feature data is output. The size of the sub - senior feature data is C3×(M + 1)×(N + 1). A 3×3×32 convolution kernel is used to operate on the sub - senior feature data and senior feature data is output. The size of the senior feature data is C2×M×N.

[0008] Further, in step S5, the size of the compressed data is C2×M×N.

[0009] Further, in step S6, the first dilated convolution sub - module is used to operate on the compressed data obtained in step S5. A 3×3×16 convolution kernel is used for operation to output data with a conventional convolution kernel of 3, and the size is C4×(M - 1)×(N - 1), where C4 = 16. Then, it is operated through a 3×3×16 convolution kernel, with a dilation rate of 2, a stride of 1, and a padding rate of 1 to output data with a dilated convolution kernel of 3, and the size is C4×(M - 4)×(N - 4). The data with a dilated convolution kernel of 3 undergoes upsampling padding operation, with a padding size of 4, to output high - level feature data with a dilated convolution kernel of 3. The size of the high - level feature data with a dilated convolution kernel of 3 is C4×M×N.

[0010] Further, in step S6, the second dilated convolution sub - module is used to operate on the compressed data obtained in step S5. A 5×5×16 convolution kernel is used for operation to output data with a conventional convolution kernel of 5, and the size is C4×(M - 3)×(N - 3), where C4 = 16. Then, it is operated through a 5×5×16 convolution kernel, with a dilation rate of 2, a stride of 1, and a padding rate of 1 to output data with a dilated convolution kernel of 5, and the size is C4×(M - 10)×(N - 10). The data with a dilated convolution kernel of 5 undergoes upsampling padding operation, with a padding size of 10, to output high - level feature data with a dilated convolution kernel of 5. The size of the high - level feature data with a dilated convolution kernel of 5 is C4×M×N.

[0011] Further, in step S6, the third dilated convolution sub-module is used to operate on the compressed data obtained in step S5. A convolution kernel of 8×8×16 is used for the operation to output data with a conventional convolution kernel of 8, and the size is C4×(M - 6)×(N - 6), where C4 = 16. Then, an operation is performed through a convolution kernel of 8×8×16, a dilation rate of 2, a stride of 1, and a padding rate of 1 to output data with a dilated convolution kernel of 5, and the size is C4×(M - 19)×(N - 19). The data with a dilated convolution kernel of 8 is subjected to upsampling filling operation with a filling size of 19 to output high-level feature data with a dilated convolution kernel of 8. The size of the high-level feature data with a dilated convolution kernel of 5 is C4×M×N.

[0012] Further, in step S6, the size of the first concatenated data is C5×M×N, where C5 = 48; the size of the residual operation data is C2×M×N, and W3 ∈ (0, 1); the size of the second concatenated data is C7×M×N, where C7 = 80; the size of the second compressed data is C6×M×N; in the regularization operation, the convolution kernel is 3×3×1, the padding rate is 1, the stride is 1, and the size of the simplified data is 1×M×N.

[0013] Further, in step S7, where Loss is the loss value, n is the number of parameter operation data, X 输出 is the simplified data, and X 标签 is the label temperature data.

[0014] An underground cable surrounding temperature field inversion device based on RDC-SE-Net includes a first feature operation module, a first SE module, a residual convolution dilated module, a first connection module, a second connection module, a second SE module, a second feature operation module, and a Loss operation module; The first feature operation module: receives the sound velocity time distance data with a dimension of C1×M×N, and sequentially passes the data through convolution operations with a 3×3×32 convolution kernel, a 3×3×64 convolution kernel, batch normalization operation, a 3×3×64 convolution kernel operation, and a 3×3×32 convolution kernel operation to output high-level feature data; The first SE module: receives the high-level feature data output by the first feature operation module, and performs compression processing on the high-level feature data to output compressed data; Residual convolution hole module: includes the first hole convolution submodule, the second hole convolution submodule, the third hole convolution submodule and the residual operation submodule; the first hole convolution submodule, the second hole convolution submodule and the third hole convolution submodule respectively receive the compressed data output by the first SE module and perform operations, and output high-level feature data with hole convolution kernels of 3, 5, and 8 respectively. The hole rate, step size and filling rate of the first hole convolution submodule, the second hole convolution submodule and the third hole convolution submodule are all 2, 1, and 1 respectively; the residual operation submodule receives the compressed data output by the first SE module and performs residual operation, outputting residual operation data; The first connection module receives the high-level feature data with dilated convolution kernels of 3, 5, and 8 output by the first, second, and third dilated convolution submodules, concatenates them, and outputs the concatenated data. The second connection module receives the primary splicing data output by the first connection module and the residual operation data output by the residual operation submodule, performs splicing on the data, and outputs secondary splicing data; The second SE module receives the secondary splicing data output by the second connection module and the compressed data output by the first SE module, performs compression processing on the data, and outputs secondary compressed data. The second feature operation module receives the secondary compressed data output by the second SE module, performs regularization operation, and outputs simplified data; Loss operation module: receives the label temperature data with a dimension of 1×M×N and the simplified data output by the second feature operation module, uses the mean square error loss function to calculate the Loss value between the simplified data and the label temperature data, and uses the Adam algorithm optimizer to set the number of training rounds. When Loss ≤ 0.002, the number of training rounds is terminated and the simplified data after training is output. The simplified data after training is the inverted temperature data.

[0015] The beneficial effects that can be achieved by the present invention are: (1) By integrating the spatiotemporal characteristics of the acoustic field time-distance map and combining the residual dilated convolution module (RDC) with the SE mechanism, an efficient modeling of the acoustic-thermal coupling effect in complex media is achieved, breaking through the constraints of traditional physical models on physical media, and achieving an inversion accuracy of the temperature field around the optoelectronic composite cable to ±0.7℃ in the soil scenario.

[0016] (2) The lightweight RDC-SE-Net architecture designed in this invention compresses the calculation time of a single temperature field inversion to less than 500ms through the parallel extraction mechanism of spatiotemporal features and the adaptive gradient clipping technology; combined with the kilometer-level distributed sensing capability of the DAS system, it can simultaneously realize the two-dimensional temperature field dynamic imaging of the entire cable channel (up to 10km), meeting the stringent requirements of real-time power grid control for high refresh rate and low latency.

[0017] (3) The residual convolutional dilated module calculates the feature data using different convolutions with the same dilation. The weights of different convolutions are different, and the dilation increases the diversity of calculations. Compared with the prior art where different dilations are used with the same convolution kernel, it avoids feature redundancy and reduces unnecessary calculations. Description of the Drawings

[0018] Figure 1 It is a schematic diagram of the moving placement point of the speaker in the embodiment of the present invention.

[0019] Figure 2 It is the sound velocity-time-distance data collected by the distributed fiber optic acoustic sensing system in the embodiment of the present invention.

[0020] Figure 3 It is the label temperature data in the embodiment of the present invention.

[0021] Figure 4 It is the data processing flow chart in the embodiment of the present invention.

[0022] Figure 5 It is the loss value of the RDC-SE-Net network in different epochs in the embodiment of the present invention.

[0023] Figure 6 It is the display of the time-consuming of data processing in the embodiment of the present invention.

[0024] Figure 7 It is the inverted temperature data output in the embodiment of the present invention.

[0025] Figure 8 It is the comparison error curve between the inverted temperature data and the label temperature data in the embodiment of the present invention.

[0026] In the figure: 1 - signal generator, 2 - speaker, 3 - buried cable, 4 - optical fiber, 5 - DAS system. Detailed Embodiments

[0027] The present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0028] The buried cable in this embodiment is an optical and electrical composite cable, and the optical fiber of the distributed fiber optic acoustic sensing system (DAS system) is wrapped in the buried cable.

[0029] There is a corresponding relationship between the sound velocity and the temperature. , where is the sound velocity, is the temperature acoustic coefficient, and T is the temperature. The sound velocity is an intermediate quantity. By using the DAS system to collect the sound velocity-time-distance data, there is a one-to-one correspondence between the sound wave and the temperature. The purpose of this embodiment is to use the weights of deep learning to train the mapping relationship between the sound velocity-time-distance data and the label temperature data.

[0030] A brief introduction to the distributed fiber optic acoustic sensing system (DAS system) is as follows: The DAS system uses optical fiber as an acoustic wave sensor and can achieve distributed acoustic wave monitoring, that is, each section of distance L1 on the optical fiber serves as an independent acoustic wave sensor. Limited by hardware parameters, based on the sampling time of the acquisition card and the distance that light travels in the optical fiber, the DAS system can obtain that an optical fiber acoustic wave sensor in the optical and electrical composite cable is formed every L1m along the direction of the optical cable. The sampling rate f1 of the acquisition card of the DAS system is 1 G / s, that is, it samples once every 1 ns. The distance that light travels in the optical fiber in 1 ns is 0.1 m, so L1 = 0.1 m for the DAS system: , where c is the speed of light and n is the refractive index of the optical fiber; the hardware of the DAS system emits rectangular pulses into the optical fiber to obtain external acoustic wave information. The emission frequency of the rectangular pulse is f2 = 20 kHz, which is determined by the emission frequency f2 in the time dimension. For the acquisition time t, the number of time points M1 collected by each L1m optical fiber acoustic wave sensor in the time domain is: .

[0031] Example: A method for inverting the temperature field around a buried cable based on RDC-SE-Net includes the following steps: Step S1: Determine a straight line L parallel to the buried cable, and the distance between the straight line L and the buried cable is not greater than 1 m, as Figure 1 shown; the selected test length of the buried cable is 200 m.

[0032] Step S2: Determine C1 equally spaced speaker moving placement points on the straight line L. In this embodiment, C1 = 10; when the speaker moves to each moving placement point, it emits an acoustic wave. The frequency of the acoustic wave emitted by the speaker is 1 - 200 Hz, and the amplitude ≥ 5 Vpp; use the distributed fiber optic acoustic sensing system to collect a set of sound speed time distance data with dimensions M × N, as Figure 2 shown; a total of sound speed time distance data with dimensions C1 × M × N is obtained. The number of collected time data = M, and the number of collected distance data = N; in this embodiment, M = 200 and N = 200.

[0033] Step S3: In the plane formed by the straight line L and the buried cable, determine M × N temperature acquisition points according to the determinant. Among them, the number of rows along the axial direction of the buried cable is M, and the number of columns perpendicular to the axial direction of the buried cable is N. Use a pin-type temperature sensor to collect 1 temperature value at each temperature acquisition point as the labeled temperature, and a total of labeled temperature data with dimensions 1 × M × N is obtained, as Figure 4 shown.

[0034] Step S4: Input the sound speed time distance data obtained in Step S2 into the first feature operation module. First, perform operations using a 3×3×32 convolution kernel and output primary feature data. The size of the primary feature data is C2×(M - 1)×(N - 1), where C2 = 32. The convolution operation size calculation formula: (1) Perform operations on the primary feature data using a 3×3×64 convolution kernel and output secondary feature operation data. The size of the secondary feature operation data is C3×(M - 2)×(N - 2), where C3 = 64. Calculate the average value and variance of the secondary feature operation data (batch normalization operation) and output normalized padding data. The padding size is 4, and the size of the normalized padding data is C3×(M + 2)×(N + 2). Perform operations on the normalized padding data using a 3×3×64 convolution kernel and output sub - senior feature data. The size of the sub - senior feature data is C3×(M + 1)×(N + 1). Perform operations on the sub - senior feature data using a 3×3×32 convolution kernel and output high - level feature data. The size of the high - level feature data is C2×M×N.

[0035] Step S5: Use the first SE module to perform compression processing on the high - level feature data obtained in Step S4 and output compressed data , the compressed data has a size of C2×M×N.

[0036] Step S6: Use the first dilated convolution sub - module to perform operations on the compressed data obtained in Step S5 Perform operations using a 3×3×16 convolution kernel and output data with a conventional convolution kernel of 3. The size is C4×(M - 1)×(N - 1), where C4 = 16. Then perform operations using a 3×3×16 convolution kernel, a dilation rate of 2, a stride of 1, and a padding rate of 1 and output data with a dilated convolution kernel of 3. The size is C4×(M - 4)×(N - 4). The data with a dilated convolution kernel of 3 undergoes upsampling padding operation, with a padding size of 4, and outputs high - level feature data with a dilated convolution kernel of 3. The size of the high - level feature data with a dilated convolution kernel of 3 is C4×M×N.

[0037] Use the second dilated convolution sub - module to perform operations on the compressed data obtained in Step S5 Perform operations. Use a 5×5×16 convolutional kernel to perform operations and output data with a conventional convolutional kernel of 5, with a size of C4×(M - 3)×(N - 3), where C4 = 16; then perform operations using a 5×5×16 convolutional kernel, a dilation rate of 2, a stride of 1, and a padding rate of 1, and output data with a dilated convolutional kernel of 5, with a size of C4×(M - 10)×(N - 10); the data with a dilated convolutional kernel of 5 undergoes upsampling padding operations with a padding size of 10 to output high-level feature data with a dilated convolutional kernel of 5, and the size of the high-level feature data with a dilated convolutional kernel of 5 is C4×M×N.

[0038] Use the third dilated convolutional sub-module to process the compressed data obtained in step S5 Perform operations. Use an 8×8×16 convolutional kernel to perform operations and output data with a conventional convolutional kernel of 8, with a size of C4×(M - 6)×(N - 6), where C4 = 16; then perform operations using an 8×8×16 convolutional kernel, a dilation rate of 2, a stride of 1, and a padding rate of 1, and output data with a dilated convolutional kernel of 5, with a size of C4×(M - 19)×(N - 19); the data with a dilated convolutional kernel of 8 undergoes upsampling padding operations with a padding size of 19 to output high-level feature data with a dilated convolutional kernel of 8, and the size of the high-level feature data with a dilated convolutional kernel of 5 is C4×M×N.

[0039] Use the first connection module to splice three groups of high-level feature data with dilated convolutional kernels of 3, 5, and 8 and output the first spliced data, and the size of the first spliced data is C5×M×N, where C5 = 48.

[0040] Perform residual operations on the compressed data through the residual operation sub-module Perform residual operations and output residual operation data, and the size of the residual operation data is C2×M×N, where W3 ∈ (0, 1).

[0041] Use the second connection module to connect the residual operation data and the first spliced data and output the second spliced data, and the size of the second spliced data is C7×M×N, where C7 = 80.

[0042] Use the second SE module to compress the second spliced data and output the second compressed data, and the size of the second compressed data is C6×M×N.

[0043] Use the second feature operation module to perform regularization operations on the second compressed data, then use a 3×3×1 convolution, a padding rate of 1, and a stride of 1 to calculate the regularized data, and output the simplified data, and the size of the simplified data is 1×M×N.

[0044] The size of the dilated convolutional kernel after dilation = dilation rate * (original convolutional kernel size - 1) + 1 (2) Step S7: Input the simplified data obtained in step S6 and the labeled temperature data obtained in step S3 into the Loss operation module. Calculate the Loss value between the simplified data and the labeled temperature data using the mean squared error loss function. Set the number of training epochs using the Adam algorithm optimizer. When Loss ≤ 0.002, the training epochs terminate. As Figure 5 shown, output the trained simplified data, and the trained simplified data is the inverted temperature data, as Figure 7 shown.

[0045] (3) where Loss is the loss value, n is the number of parameter operation data, X 输出 is the simplified data, and X 标签 is the labeled temperature data.

[0046] Figure 8 is the comparison error curve between the inverted temperature data and the labeled temperature data. The maximum error is 16.5℃ - 15.8℃ = 0.7℃. Compared with the traditional method, the inversion accuracy of the temperature field has been greatly improved.

[0047] A temperature field inversion device for buried cables based on RDC-SE-Net includes a first feature operation module, a first SE module, a residual convolutional dilation module, a first connection module, a second connection module, a second SE module, a second feature operation module, and a Loss operation module.

[0048] The first feature operation module: Receives the sound velocity time distance data with dimensions of C1×M×N, and sequentially performs operations with a 3×3×32 convolutional kernel, a 3×3×64 convolutional kernel, batch normalization, a 3×3×64 convolutional kernel, and a 3×3×32 convolutional kernel on the data, and outputs high-level feature data.

[0049] The first SE module: Receives the high-level feature data output by the first feature operation module, and performs compression processing on the high-level feature data, and outputs compressed data.

[0050] The residual convolutional dilation module: Includes a first dilation convolutional sub-module, a second dilation convolutional sub-module, a third dilation convolutional sub-module, and a residual operation sub-module; the first dilation convolutional sub-module, the second dilation convolutional sub-module, and the third dilation convolutional sub-module respectively receive the compressed data output by the first SE module and perform operations, and respectively output high-level feature data with dilation kernels of 3, 5, and 8. The dilation rates of the first dilation convolutional sub-module, the second dilation convolutional sub-module, and the third dilation convolutional sub-module are all 2, the strides are all 1, and the padding rates are all 1; the residual operation sub-module receives the compressed data output by the first SE module and performs residual operations, and outputs residual operation data.

[0051] The first connection module: receives the high-level feature data with dilated convolution kernels of 3, 5, and 8 output by the first dilated convolution sub-module, the second dilated convolution sub-module, and the third dilated convolution sub-module, and performs splicing to output the first spliced data.

[0052] The second connection module: receives the first spliced data output by the first connection module and the residual operation data output by the residual operation sub-module, and performs splicing to output the second spliced data.

[0053] The second SE module: receives the second spliced data output by the second connection module and the compressed data output by the first SE module, and performs compression processing to output the second compressed data.

[0054] The second feature operation module: receives the second compressed data output by the second SE module, performs regularization operations, then calculates using a 3×3×1 convolution with a padding rate of 1 and a stride of 1 on the regularized data, and outputs the simplified data.

[0055] The Loss operation module: receives the label temperature data with a dimension of 1×M×N and the simplified data output by the second feature operation module, calculates the Loss value between the simplified data and the label temperature data using the mean squared error loss function, sets the number of training epochs using the Adam algorithm optimizer, terminates the training epochs when Loss ≤ 0.002, and outputs the trained simplified data, which is the inverted temperature data.

Claims

1. A method for inverting the temperature field around buried cables based on RDC-SE-Net, characterized by: It includes the following steps: Step S1: Determine a straight line L parallel to the buried cable, and the distance between the straight line L and the buried cable is not greater than 1 m; Step S2: Determine C1 equally spaced speaker moving placement points on the straight line L. When the speaker moves to each moving placement point, it emits sound waves. Use a distributed fiber optic acoustic sensing system to collect a set of acoustic velocity-time-distance data with a dimension of M×N, and a total of acoustic velocity-time-distance data with a dimension of C1×M×N is obtained. The number of time data collected = M, and the number of distance data collected = N; Step S3: In the plane formed by the straight line L and the buried cable, determine M×N temperature acquisition points according to the determinant. Among them, the number of rows along the axial direction of the buried cable is M, and the number of columns perpendicular to the axial direction of the buried cable is N. Use a temperature sensor to collect 1 temperature value at each temperature acquisition point as the labeled temperature, and a total of labeled temperature data with a dimension of 1×M×N is obtained; Step S4: Input the acoustic velocity-time-distance data obtained in Step S2 into the first feature operation module. The data is successively operated by a 3×3×32 convolution kernel, a 3×3×64 convolution kernel, batch normalization, a 3×3×64 convolution kernel, and a 3×3×32 convolution kernel, and high-level feature data is output; Step S5: Use the first SE module to perform compression processing on the high-level feature data obtained in Step S4, and output compressed data; Step S6: Respectively use the first dilated convolution sub-module, the second dilated convolution sub-module, and the third dilated convolution sub-module to operate on the compressed data obtained in Step S5. The dilation rates of these 3 sub-modules are all 2, the strides are all 1, and the padding rates are all 1; The first dilated convolution sub-module outputs high-level feature data with a dilated convolution kernel of 3, the second dilated convolution sub-module outputs high-level feature data with a dilated convolution kernel of 5, and the third dilated convolution sub-module outputs high-level feature data with a dilated convolution kernel of 8; Use the first connection module to splice the three groups of high-level feature data with dilated convolution kernels of 3, 5, and 8 and output the first spliced data. Perform residual operation on the compressed data through the residual operation sub-module and output the residual operation data. Use the second connection module to connect the residual operation data and the first spliced data and output the second spliced data. Use the second SE module to perform compression processing on the second spliced data and output the second compressed data. Use the second feature operation module to perform regularization operation on the second compressed data, and then calculate it using a 3×3×1 convolution, a padding rate of 1, and a stride of 1, and output the simplified data; Step S7: Input the simplified data obtained in Step S6 and the labeled temperature data obtained in Step S3 into the Loss operation module. Use the mean square error loss function to calculate the Loss value between the simplified data and the labeled temperature data. Use the Adam algorithm optimizer to set the number of training epochs. When Loss≤0.002, the training epoch ends, and the trained simplified data is output. The trained simplified data is the inversion temperature data.

2. The method for inverting the temperature field around the buried cable based on RDC-SE-Net according to claim 1, characterized in that: In Step S2, the frequency of the sound wave emitted by the speaker is 1~200 Hz, and the amplitude ≥5 Vpp; C1≥10, M≥200, N≥200.

3. The method for inverting the temperature field around the buried cable based on RDC-SE-Net according to claim 1, wherein: Step S4 In it, the sound velocity time-distance data obtained in step S2 is input into the first feature operation module. First, a 3×3×32 convolution kernel is used for operation and primary feature data is output. The size of the primary feature data is C2×(M - 1)×(N - 1), where C2 = 32. A 3×3×64 convolution kernel is used to operate on the primary feature data and secondary feature operation data is output. The size of the secondary feature operation data is C3×(M - 2)×(N - 2), where C3 = 64. Batch normalization operation is performed on the secondary feature operation data and normalized padding data is output. The padding size is 4, and the size of the normalized padding data is C3×(M + 2)×(N + 2). A 3×3×64 convolution kernel is used to operate on the normalized padding data and sub-senior feature data is output. The size of the sub-senior feature data is C3×(M + 1)×(N + 1). A 3×3×32 convolution kernel is used to operate on the sub-senior feature data and senior feature data is output. The size of the senior feature data is C2×M×N.

4. The method for inverting the temperature field around the buried cable based on RDC-SE-Net according to claim 3, characterized in that: In step S5, the size of the compressed data is C2×M×N.

5. The method for inverting the temperature field around the buried cable based on RDC-SE-Net according to claim 4, characterized in that: In step S6, the first dilated convolution sub-module is used to operate on the compressed data obtained in step S5. A 3×3×16 convolution kernel is used for operation to output data with a conventional convolution kernel of 3, and the size is C4×(M - 1)×(N - 1), where C4 = 16. Then, it is operated through a 3×3×16 convolution kernel, with a dilation rate of 2, a stride of 1, and a padding rate of 1 to output data with a dilated convolution kernel of 3, and the size is C4×(M - 4)×(N - 4). The data with a dilated convolution kernel of 3 undergoes upsampling padding operation, with a padding size of 4, to output high-level feature data with a dilated convolution kernel of 3. The size of the high-level feature data with a dilated convolution kernel of 3 is C4×M×N.

6. The method for inverting the temperature field around the buried cable based on RDC-SE-Net according to claim 4, characterized in that: In step S6, the second dilated convolution sub-module is used to operate on the compressed data obtained in step S5. A 5×5×16 convolution kernel is used for operation to output data with a conventional convolution kernel of 5, and the size is C4×(M - 3)×(N - 3), where C4 = 16. Then, it is operated through a 5×5×16 convolution kernel, with a dilation rate of 2, a stride of 1, and a padding rate of 1 to output data with a dilated convolution kernel of 5, and the size is C4×(M - 10)×(N - 10). The data with a dilated convolution kernel of 5 undergoes upsampling padding operation, with a padding size of 10, to output high-level feature data with a dilated convolution kernel of 5. The size of the high-level feature data with a dilated convolution kernel of 5 is C4×M×N.

7. The method for inverting the temperature field around the buried cable based on RDC-SE-Net according to claim 4, characterized in that: In step S6, the third dilated convolution sub-module is used to operate on the compressed data obtained in step S5. A convolution kernel of 8×8×16 is used for the operation, and the output has a conventional convolution kernel of 8, with a size of C4×(M - 6)×(N - 6), where C4 = 16. Then, operations are performed using a convolution kernel of 8×8×16, a dilation rate of 2, a stride of 1, and a padding rate of 1, and the output has a dilated convolution kernel of 5, with a size of C4×(M - 19)×(N - 19). The data with a dilated convolution kernel of 8 undergoes upsampling padding operation with a padding size of 19, and the output is the high-level feature data with a dilated convolution kernel of 8. The size of the high-level feature data with a dilated convolution kernel of 5 is C4×M×N.

8. The method for inverting the temperature field around the buried cable based on RDC-SE-Net according to claim 4, characterized in that: In step S6, the size of the first concatenated data is C5×M×N, where C5 = 48; the size of the residual operation data is C2×M×N, and W3 ∈ (0, 1); the size of the second concatenated data is C7×M×N, where C7 = 80; the size of the second compressed data is C6×M×N.

9. The method for inverting the temperature field around the buried cable based on RDC-SE-Net according to claim 1, characterized in that: In step S7, Among them, Loss is the loss value, n is the number of parameter operation data, X 输出 To simplify the data, X 标签 is the label temperature data.

10. An underground cable surrounding temperature field inversion device based on RDC-SE-Net, characterized in that: It includes a first feature operation module, a first SE module, a residual convolutional dilation module, a first connection module, a second connection module, a second SE module, a second feature operation module, and a Loss operation module; The first feature operation module: receives the sound velocity time-distance data with a dimension of C1×M×N, and sequentially passes the data through operations using a 3×3×32 convolution kernel, a 3×3×64 convolution kernel, batch normalization operation, a 3×3×64 convolution kernel operation, and a 3×3×32 convolution kernel operation, and outputs high-level feature data; The first SE module: receives the high-level feature data output by the first feature operation module, and performs compression processing on the high-level feature data, and outputs compressed data; The residual convolutional dilation module: includes a first dilated convolution sub-module, a second dilated convolution sub-module, a third dilated convolution sub-module, and a residual operation sub-module; the first dilated convolution sub-module, the second dilated convolution sub-module, and the third dilated convolution sub-module respectively receive the compressed data output by the first SE module and perform operations, and respectively output high-level feature data with dilated convolution kernels of 3, 5, and 8. The dilation rates of the first dilated convolution sub-module, the second dilated convolution sub-module, and the third dilated convolution sub-module are all 2, the strides are all 1, and the padding rates are all 1; the residual operation sub-module receives the compressed data output by the first SE module and performs a residual operation, and outputs residual operation data; The first connection module: receives the high-level feature data with dilated convolution kernels of 3, 5, and 8 output by the first dilated convolution sub-module, the second dilated convolution sub-module, and the third dilated convolution sub-module, and performs concatenation, and outputs the first concatenated data; The second connection module: receives the first concatenated data output by the first connection module and the residual operation data output by the residual operation sub-module, and performs concatenation, and outputs the second concatenated data; The second SE module: receives the second concatenated data output by the second connection module and the compressed data output by the first SE module, and performs compression processing, and outputs the second compressed data; The second feature operation module: receives the second compressed data output by the second SE module, and performs regularization operation, and outputs simplified data; Loss operation module: Receives the labeled temperature data with a dimension of 1×M×N and the simplified data output by the second feature operation module, calculates the Loss value between the simplified data and the labeled temperature data using the mean squared error loss function, sets the number of training epochs using the Adam algorithm optimizer, terminates the number of training epochs when Loss ≤ 0.002, and outputs the trained simplified data, which is the inverted temperature data.

Images