A Neutron Rare Tomography Method for Non-destructive Testing of Insulation Materials of Combined Electrical Switches

Abstract

This application relates to the field of nondestructive testing, specifically providing a neutron sparse tomography method for nondestructive testing of insulating materials of combined electrical switches. The method includes the following steps: S1, obtaining a sparse projection scheme through a preset neural network model; S2, performing sparse projection on the combined electrical switch under test according to the sparse projection scheme; S3, reconstructing an image of the combined electrical switch under test based on the projection data; S4, determining the damage status of the combined electrical switch under test based on the reconstructed image. This invention uses neutron sparse projection to obtain projection data for image reconstruction. By directly detecting high-precision images of solid insulating materials, the location, extent, and type of defects are obtained, rather than through indirect partial discharge detection, eliminating the need for a period of defect accumulation. Therefore, defects can be detected quickly and accurately in their early stages, thereby eliminating potential hazards as early as possible, improving the overall reliability of the power system, and reducing operation and maintenance costs.

Description

Technical Field

[0001] This application relates to the field of nondestructive testing, and more specifically, to a neutron rarefaction tomography method for nondestructive testing of insulating materials of combined electrical switches. Background Technology

[0002] Integral switchgear (GIS) is an important electrical component, and its solid insulation materials (such as electrical porcelain, fiberglass, silicone rubber, and epoxy resin) play a crucial role in ensuring its safe operation. Defects in the solid insulation materials can lead to a decrease in the performance of the GIS; therefore, defect detection of solid insulation materials in GIS is a key issue in the power industry.

[0003] Because a significant number of combined electrical switches have a closed structure, it is difficult to obtain the surface condition of the internal solid insulation material (such as surface cracks, dust particles, and oil stains) through direct observation. Defects in the solid insulation material can lead to partial discharge, which is a major manifestation of performance degradation in the solid insulation material of many combined electrical switches. Existing technologies determine the presence and type of defects in the solid insulation material by observing the occurrence and intensity of partial discharge. Partial discharge detection technologies based on various principles such as ultrasonic waves, UHF, and decomposition analysis have become a major component of solid insulation defect diagnosis. However, these methods are qualitative in nature. They can only be detected after the defects in the solid insulation material have accumulated to a certain extent and caused partial discharge. Furthermore, they depend on the sensitivity of the monitoring device to the partial discharge signal. In other words, they can determine whether partial discharge has occurred and roughly locate the defect after the performance of the solid insulation material has degraded to a certain extent, but they cannot accurately present the severity and type of the defect.

[0004] In summary, existing detection technologies can only detect defects in solid insulation materials after they have accumulated to a certain extent and caused partial discharge. They cannot detect defects quickly and accurately in their early stages, thus exacerbating potential hazards in combined electrical switches. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by providing a neutron rarefaction tomography method for non-destructive testing of insulating materials in combined electrical switches. This method solves the problem that in the prior art, defects in solid insulating materials can only be detected after they have accumulated to a certain extent and caused partial discharge, making it impossible to perform rapid and accurate detection in the early stages of defect formation, thus exacerbating the potential hazards in combined electrical switches.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] This application provides a neutron sparse tomography method for non-destructive testing of insulating materials of combined electrical switches. The method includes the following steps: S1, obtaining a sparse projection scheme through a preset neural network model; S2, performing sparse projection on the combined electrical switch under test according to the sparse projection scheme; S3, reconstructing an image of the combined electrical switch under test based on the projection data; S4, determining the damage condition of the combined electrical switch under test based on the reconstructed image.

[0008] Furthermore, the preset neural network model is a coupled neural network model, which includes a sparse projection module and an encoder-decoder module, and the sparse projection module and the encoder-decoder module are connected in series.

[0009] Furthermore, the sparse projection module is a multi-layer fully connected binary neural network, which is used to provide the sparse projection scheme.

[0010] Furthermore, the input to the binary neural network is an n-dimensional vector encoding with 1s, where n represents the maximum number of projections used for the tomography of the combined electrical switches under test. The output of the binary neural network is a sparse projection scheme, which is a projection scheme encoding.

[0011] Furthermore, the encoder-decoder module includes an encoder network and a decoder network, which are used to reconstruct the tomographic image.

[0012] Furthermore, step S1 also includes training a preset neural network model, using a dataset to perform unsupervised training on the encoder-decoder module, with the input of the encoder-decoder module being the calculated projection data, and the calculated projection data being obtained from the images in the dataset.

[0013] Furthermore, the calculated projection data is obtained based on the Radon transform, as shown in the following formula: Where R represents the Radon operator, f represents the target image, (x,y) represents the pixel coordinates, α represents the projection angle, s represents the distance from the far point on the scan line, and δ is the Dirac delta function.

[0014] Furthermore, in step S2, the movable neutron source performs a sparse neutron layer scan on the combined electronic switch under test according to the sparse projection scheme. The neutrons transmitted through the combined electronic switch under test are received by the n / γ conversion screen and converted into photon signals. The photon signals are then captured by the optical camera to obtain projection data.

[0015] Furthermore, in step S4, the damage status of the combined electrical switch under test is obtained based on the smoothness and position of the pixel values ​​representing the solid insulating material structure in the reconstruction results.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a method for non-destructive testing of the solid insulation material of a GIS (Gas Insulated Switchgear) without damaging its external metal shell structure. Specifically, the method of this invention is based on neutron tomography to perform tomographic scanning and reconstruction of the GIS. Neutron rays do not interact with electrons but with atomic nuclei, unlike X-rays and gamma rays. Neutrons are attenuated by some lightweight materials and can penetrate heavy materials, meaning that neutrons can easily pass through the shell of the GIS and interact with the lightweight solid insulation material. This invention utilizes the high penetrability of neutron rays to the GIS shell and their high sensitivity to elements such as carbon, hydrogen, and silicon (corresponding to stains in solid insulation materials) to accurately monitor internal defects, external defects, and stain adhesion in the solid insulation material. More specifically, the method of this invention first obtains projection data through neutron sparse projection, and then uses the projection data to reconstruct the solid insulation material structure with high precision. The resulting reconstructed image clearly reflects defects in the solid insulation material. This invention obtains the location, extent, and type of damage of defects by directly detecting high-precision images of solid insulating materials, rather than through indirect partial discharge detection. It does not require a period of defect accumulation. Therefore, defects can be detected quickly and accurately in their early stages, thereby eliminating potential hazards as early as possible, improving the overall reliability of the power system, and reducing operation and maintenance costs.

[0017] In addition, the method of the present invention utilizes sparse projection to invert the precise characteristics of solid insulation materials in GIS, thereby improving the tomography efficiency, which is to say, the accuracy of non-destructive testing of solid insulation materials. At the same time, it reduces the radiation dose absorbed by the equipment, that is, sparse projection reduces the radiation dose absorbed by the combined electrical switches, thus ensuring equipment safety. Attached Figure Description

[0018] Figure 1 A schematic diagram of a neutron rarefaction tomography method for non-destructive testing of insulating materials of combined electrical switches provided by the present invention;

[0019] Figure 2 A schematic diagram of the coupled neural network model in the neutron sparse tomography method for non-destructive testing of insulation materials of combined electrical switches provided by the present invention;

[0020] Figure 3 The flowchart corresponds to the neutron rarefaction tomography method for non-destructive testing of insulating materials of combined electrical switches provided by the present invention.

[0021] Figure 4 This invention provides a system corresponding to a neutron rarefaction tomography method for non-destructive testing of insulating materials of combined electrical switches.

[0022] Icons: 1-Mobile neutron source; 2-Combined electrical switch rotatable platform; 3-n / γ conversion screen; 4-Optical camera; 5-Data processing system. Detailed Implementation

[0023] To make the implementation process of this invention clearer, a detailed description will be provided below in conjunction with the accompanying drawings.

[0024] This invention provides a neutron rarefaction tomography method for non-destructive testing of insulating materials of combined electrical switches, such as... Figure 1 As shown, the method includes the following steps:

[0025] S1, obtain the sparse projection scheme through a preset neural network model;

[0026] An n-dimensional vector encoding with an element of 1 is input into a preset neural network model, which outputs the corresponding sparse projection scheme. Here, dimension n represents the maximum number of projections used for tomography of the combined electrical switches under test, i.e., the number of angles selected during projection, which determines the specific angles. Preferably, the maximum number of projections is less than 360, meaning the scan interval is greater than or equal to 1 degree. Specifically, the sparse projection scheme is essentially a vector encoding with the same dimension as the input vector encoding, where the elements are either 0 or 1. An element of 1 means that the corresponding projection angle needs to be used in the sparse projection, while an element of 0 means that the corresponding projection angle is not used. The preset neural network model is a coupled neural network model, which also needs to be trained. Specifically, the construction and training process of the coupled neural network model is as follows:

[0027] S11, Construct a coupled neural network model;

[0028] A reconstruction model of a solid insulating material structure and a neutron sparse projection scheme are jointly constructed using a deeply coupled neural network consisting of a sparse projection module and an encoder-decoder module. The coupled neural network includes both a sparse projection module and an encoder-decoder module. The sparse projection scheme is obtained from the sparse projection module, which is input as an n-dimensional vector encoding with elements equal to 1, and outputs the corresponding sparse projection scheme. Specifically, the sparse projection module consists of a multi-layer fully connected binary neural network, more specifically, three fully connected network layers plus an activation function, which can be either a binary activation function or a continuous activation function. The encoder-decoder module consists of an encoding network that performs low-dimensional manifold mapping of the projection information and a decoding network that reconstructs the low-dimensional encoding into a tomographic image. More specifically, the encoding network consists of three sets of dimensionality reduction modules, each containing a downsampling layer, a convolutional layer, a pooling layer, and a normalization layer; the decoding network consists of three sets of dimensionality upsampling modules, each containing an upsampling layer, a convolutional layer, a pooling layer, and a normalization layer. Figure 2 As shown, the sparse projection module is connected in series with the encoder-decoder module. The projection scheme code output by the sparse projection module is used as one of the inputs and sent to the encoder-decoder module along with the projection data. The encoder-decoder module is responsible for reconstructing the internal fault structure of the combined electrical switch based on the sparse projection information and outputting the fault reconstruction image. This improves the accuracy of the reconstructed fault structure. In this way, the preset neural network model can simultaneously output the sparse projection scheme and the reconstruction result, thereby improving the accuracy of non-destructive testing.

[0029] In summary, this invention constructs a scheme for simultaneously forming sparse tomography and a method for reconstructing the tomographic structure of combined electrical switches. The sparse tomography scheme is formed in a sparse projection module, and the tomographic image is reconstructed in an encoding-decoding module, which is serially connected to both the sparse projection module and the encoding-decoding module. This allows for the simultaneous output of the sparse projection scheme and the tomographic reconstructed image, improving tomography efficiency and increasing the accuracy of nondestructive testing, while also reducing the dose absorbed by the equipment.

[0030] The activation function in the sparse projection module can be either a continuous activation function or a 0-1 activation function. Specifically, if a variable dose is required for each projection, a continuous activation function is chosen; if a constant dose is required, a 0-1 activation function is chosen. When the variable dose option is selected, the neutron dose for each projection during neutron tomography is different, requiring the neutron source to dynamically adjust the neutron flux for each projection. When a 0-1 activation function is selected, the backpropagation process is skipped directly, avoiding the failure of differentiation in the backpropagation process due to the discontinuity of the 0-1 activation function. If differentiation fails, the backpropagation training of the sparse projection module cannot proceed. A continuous activation function does not require skipping backpropagation and will not cause differentiation failure.

[0031] By irradiating the combined electrical switch from a few angles—specifically, the few angles are determined by the sparse projection scheme output by the sparse projection module—the deep learning process requires rapid tomography with dose constraints, focusing primarily on the local insulating material structure. The sparse projection scheme is ultimately obtained through gradient descent optimization. This improves the efficiency of neutron tomography, reduces costs, and obtains tomographic imaging with fewer projections, while also reducing the radiation dose received by the combined electrical switch.

[0032] S12, Training the coupled neural network model.

[0033] The coupled neural network model constructed in step S11 is trained by combining an unsupervised training method based on large-scale unlabeled tomographic images with a supervised training method based on a small amount of labeled neutron tomographic projection data. For the encoder-decoder module, large-scale unsupervised training is required to enable the encoder network to possess general feature extraction capabilities. Specifically, a massive unsupervised dataset is constructed using publicly available and easily accessible images; specifically, the images are CT tomographic images, preferably with more than 10,000 images in the dataset. Radon transform is performed on each image in the dataset to obtain the calculated projection data. The expression for the Radon transform is: Where R represents the Radon operator, f represents the target image, (x, y) represents the pixel coordinates, α represents the projection angle, s represents the distance from the far point on the scan line, and δ is the Dirac delta function. Specifically, the integration is performed along the line x cosα + y sinα = s, and the scan can be along any normal direction α and any line at any distance s from the origin. The Radon transform result of the images in the above dataset is used as an input to the encoder-decoder module, that is, the calculated projection data obtained by the Radon transform is input to the encoder network to realize the training process. The binary neural network and the encoder network are subjected to the above unsupervised training, and the decoder network is subjected to supervised training.

[0034] Specifically, the unsupervised training described above is conducted as follows: A portion of the projection data in the tomography dataset is randomly deleted, and the prediction error of the deleted projections is gradually reduced using gradient descent. This allows the trained network to learn the low-dimensional representation of general projections and the prediction method for accurate projection angles, thereby improving the accuracy of the sparse projection scheme and the reconstructed structure, and ultimately enhancing the accuracy of nondestructive testing. The accuracy of the coupled neural network's prediction of the deleted projections and the total dose from neutron tomography are used as objective functions. Backpropagation training of the coupled neural network is performed using stochastic gradient descent. Specifically, the expression for the loss function is: L = ||B||1, where B represents the total dose of the sparse projection, and ||1 represents the L1 norm. That is, the total dose from neutron tomography of the solid insulating material is included as a term in the loss function during the backpropagation training process, minimizing the dose absorbed by the combined electrical switches. Training ends when the loss function converges to a stable value. At the end of training, the sparse projection module and the encoder-decoder module will be able to simultaneously output the sparse projection scheme and the tomography reconstruction results.

[0035] S2, Perform sparse projection on the combined electrical switch under test according to the sparse projection scheme;

[0036] Using the sparse projection scheme obtained from the coupled neural network model trained in step S1, sparse projection is performed on the combined electrical switch under test to obtain projection data. That is, neutron tomography is performed on the combined electrical switch. By irradiating the combined electrical switch from a few angles, its sparse projection data is obtained. Sparse projection inverts the precise characteristics of the solid insulation material in the combined electrical switch, improving tomography efficiency and reducing the radiation dose absorbed by the equipment while improving the accuracy of non-destructive testing of solid insulation materials.

[0037] Specifically, the electrical switch under test is placed on a rotatable platform. A movable neutron source performs sparse neutron tomography scanning around the electrical switch under test according to a sparse projection scheme, i.e., projection. Specifically, it is a transmission type, that is, the movable neutron source and the n / γ conversion screen are distributed on opposite sides of the electrical switch under test. The neutron signal is received by the n / γ conversion screen and converted into a photon signal, which is received by an optical camera to obtain the projection data for a single projection. After all projections are completed, all projection data are obtained.

[0038] S3, Reconstruct the image of the combined electrical switch under test based on the projection data;

[0039] Using the encoder-decoder module in the trained coupled neural network, all the projection data obtained in step S2 is input into the encoder-decoder module, which outputs the reconstruction result to achieve tomographic reconstruction of the combined electrical switch under test. The encoder-decoder module is the encoder-decoder module in the coupled neural network constructed and trained in step S1. The projection scheme encoding output by the sparse projection module is sent to the encoder-decoder module along with the projection data. The encoder-decoder module is responsible for reconstructing the internal tomographic structure of the combined electrical switch based on the sparse projection information and outputting the tomographic reconstruction image. This improves the accuracy of the reconstructed tomographic structure. In this way, the preset neural network model can simultaneously output the sparse projection scheme and the reconstruction result, improving the accuracy of non-destructive testing. Specifically, the encoder-decoder module consists of an encoding network that performs low-dimensional manifold mapping on the projection information and a decoding network that reconstructs the low-dimensional encoding into a tomographic image. The encoder-decoder module and the sparse projection module are connected in series. The flowchart of the method of this invention is as follows: Figure 3 As shown.

[0040] The encoder-decoder module consists of an encoder network and a decoder network. The encoder network comprises three sets of dimensionality reduction modules, each containing a downsampling layer, a convolutional layer, a pooling layer, and a normalization layer, used to perform low-dimensional manifold mapping on the projected information. The decoder network consists of three sets of dimensionality upsampling modules, each containing an upsampling layer, a convolutional layer, a pooling layer, and a normalization layer; used to reconstruct the tomographic image from the low-dimensional encoding. During training, the encoder network undergoes supervised training, while the decoder network undergoes unsupervised training.

[0041] S4. Determine the damage status of the electrical switch combination under test by reconstructing the image.

[0042] The reconstruction results obtained in step S3 can determine the damage status of the combined electrical switch under test. Based on the smoothness of the pixel values ​​representing the solid insulation material structure in the reconstruction results, the defect type and severity can be determined. The precise location of the defect can be determined based on its position in the tomographic reconstruction results. In other words, the method of this invention can determine the defect type, severity, and location of solid insulation materials, such as the adhesion of external metal / non-metal particles, surface stains, internal defects caused by charge accumulation or molecular fragmentation, and external defects. These defects reduce the insulation performance of solid insulation materials, thereby increasing potential hazards. Using the non-destructive testing method of this application to inspect the solid insulation material in combined electrical switches can quickly and accurately determine the defect situation, thereby detecting potential hazards early and avoiding greater economic losses.

[0043] This invention also proposes a system corresponding to a neutron sparse tomography method for non-destructive testing of electrical switch insulation materials, used to realize neutron tomography scanning and image reconstruction of GIS. For example... Figure 4 As shown, the system includes a movable neutron source 1, a rotatable platform 2 for the combined electrical switch 2, an n / γ conversion screen 3, an optical camera 4, and a data processing system 5. Specifically, the rotatable platform 2 is used to place the combined electrical switch under test. The movable neutron source 1 and the n / γ conversion screen 3 are respectively positioned on opposite sides of the combined electrical switch under test, and the movable neutron source 1 can rotate around the combined electrical switch. The optical camera 4 is positioned on the side of the n / γ conversion screen 3 away from the combined electrical switch under test. The n / γ conversion screen 3 is used to convert the received neutron signal into a photon signal, which is then received by the optical camera 4. The data processing system 5 includes a sparse projection module and an encoding-decoding module, which are the same as those in the aforementioned neutron sparse tomography method for non-destructive testing of insulating materials of combined electrical switches. The data processing system 5 is connected to the optical camera 4 and the movable neutron source 1. The data processing system 5 includes a computer and necessary data interfaces.

[0044] In data processing system 5, the coupled neural network is trained according to the training steps described above. The sparse projection scheme output by the sparse projection module in the coupled neural network model is sent to the mobile neutron source 1. During operation, the mobile neutron source 1 generates a neutron beam and performs sparse projection on the combined electrical switch under test according to the sparse projection scheme. Through the action of the n / γ conversion screen 3, a photon signal is formed and received by the optical camera 4 to obtain projection data. Each projection data is received and stored by data processing system 5. After all projections are completed, data processing system 5 outputs the reconstruction result through the encoder-decoder module in the coupled neural network model. Based on the reconstruction result, defects in the combined electrical switch can be determined.

[0045] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method of combined electrical switch insulation material non-destructive testing by neutron sparse tomography, characterized in that, The method comprises the following steps: S1, obtaining a sparse projection scheme by a preset neural network model; S2, performing sparse projection on a to-be-tested combined electrical switch according to the sparse projection scheme; S3, reconstructing an image of the to-be-tested combined electrical switch according to the projection data; S4, determining a damage condition of the to-be-tested combined electrical switch by the reconstructed image; The preset neural network model is a coupled neural network model, which comprises a sparse projection module and an encoding-decoding module, and the sparse projection module and the encoding-decoding module are connected in series; the sparse projection module is a multi-layer fully connected binary neural network, which is used to give the sparse projection scheme; the encoding-decoding module comprises an encoding network and a decoding network, which are used to reconstruct a tomographic image; In the step S1, the preset neural network model is trained, a data set is used, and the encoding-decoding module is unsupervisedly trained; the input of the encoding-decoding module is the calculated projection data, and the calculated projection data is obtained according to the image in the data set; In the step S2, a movable neutron source performs sparse neutron layer scanning on the to-be-tested combined electrical switch according to the sparse projection scheme; the neutrons transmitted through the to-be-tested combined electrical switch are received by an n / gamma conversion screen and converted into a photon signal, and the photon signal is obtained by an optical camera to obtain the projection data.

2. The method of claim 1, wherein the method is a neutron sparse tomography method for non-destructive testing of electrical switch insulation materials in combination. The input of the binary neural network is an n-dimensional vector code with an element of 1, wherein n represents the maximum number of projections for to-be-tested combined electrical switch tomography; the binary neural network outputs the sparse projection scheme, and the projection scheme is a projection scheme code.

3. The method of claim 2, wherein the method is a neutron sparse tomography method for non-destructive testing of electrical switch insulation materials in combination. The projection data calculated is obtained according to a radon transform, and the formula is as follows: wherein, represents a radon operator, represents a target image, represents a pixel coordinate, represents a projection angle, represents a distance of a far point on a scanning line, is a Dirac function.

4. The method of claim 3, wherein the method is a neutron sparse tomography method for non-destructive testing of electrical switch insulation materials in combination. In the step S4, the damage condition of the to-be-tested combined electrical switch is obtained according to the smoothness and position of the pixel value representing the structure of the solid insulating material in the reconstruction result.

Images