A method for training a condition detection model of a solid insulation material

Abstract

The present application relates to a kind of solid insulation material state detection model training method, belong to insulating material technical field, solve the problem of low accuracy and efficiency in prior art evaluation.It is method that includes: voltage is applied to solid insulating material until surface flashover occurs, collect the electric signal, acoustic signal and discharge image of each time when leakage current is generated in solid insulating material during pressure application process;Based on the electric signal, acoustic signal and discharge image when each time when leakage current is generated, extract multi-dimensional feature data;Multi-dimensional feature data and corresponding state label construct training sample set;The sample in the training sample set is expanded to obtain the training sample set after expansion;Multi-classification neural network model is constructed, and the multi-classification neural network model is trained based on the training sample set after expansion, and the state detection model of solid insulating material is obtained.The state evaluation of solid insulating material with high efficiency and accuracy is realized.

Description

Technical Field

[0001] This invention relates to the field of insulating materials technology, and in particular to a training method for a state detection model of solid insulating materials. Background Technology

[0002] Solid insulation materials are key factors in achieving electrical isolation and mechanical support for high-voltage transmission lines and electrical insulation equipment. Their safety is an important guarantee for the stable and efficient operation of the power system, and their good insulation performance is a prerequisite for ensuring the long-term safe and stable operation of the power system. Under long-term outdoor service environments, solid insulation materials are susceptible to the effects of concentrated electrical stress and complex and variable environmental factors. Inevitably, insulation degradation occurs on the material surface, leading to aging, moisture absorption, contamination, and even cracking, increasing the risk of surface flashover.

[0003] With the increase in voltage levels and transmission capacity of power systems, higher requirements are placed on the electrical insulation performance of solid insulating materials. Traditional methods often start from the structural or physical parameters of the insulation body, ignoring the surface charge accumulation and aging characteristics of the insulating material under long-term service. This limits the prediction of whether surface flashover will occur and makes it difficult to accurately and efficiently assess the state of solid insulating materials after surface degradation under long-term stable operation. Summary of the Invention

[0004] Based on the above analysis, the embodiments of the present invention aim to provide a training method for a state detection model of solid insulating materials, in order to solve the problems of low accuracy and efficiency in existing assessments.

[0005] On one hand, embodiments of the present invention provide a training method for a state detection model of a solid insulating material, comprising the following steps: A voltage is applied to a solid insulating material until a surface flashover occurs, and electrical signals, acoustic signals, and discharge images are collected each time the solid insulating material generates leakage current during the pressure application process. Multidimensional feature data is extracted based on the electrical signal, acoustic signal, and discharge image at each leakage current generation; a training sample set is constructed based on the multidimensional feature data and the corresponding state labels; the state labels include no obvious discharge, corona discharge, and flashover. The samples in the training sample set are expanded to obtain the expanded training sample set; A multi-class neural network model is constructed, and the model is trained based on the expanded training sample set to obtain a state detection model for solid insulating materials.

[0006] Based on a further improvement of the above method, the multi-class neural network model includes: The first feature extraction module is used to extract the first feature based on the numerical feature data in the multidimensional feature data. The second feature extraction module is used to extract the second feature based on the image feature data of the multidimensional feature data. A feature fusion module is used to fuse the first feature and the second feature to obtain a fused feature; The classification module is used to classify the state of insulators based on the fused features.

[0007] Based on a further improvement of the above method, the samples in the training sample set are expanded to obtain an expanded training sample set, including: Extract samples labeled as flashover from the training sample set, and stitch together the numerical features of the extracted samples with the corresponding discharge images to form source data; The source data is input into the trained generative adversarial network model to generate new flashover state samples. The generated new flashover state samples are then added to the training sample set to obtain the expanded training sample set.

[0008] Based on a further improvement of the above method, the generative adversarial network model includes a first generator, a second generator, a first discriminator, and a second discriminator; The first generator is used to generate a first sample based on the input source data and to perform arc segmentation on the discharge image in the first sample; the first discriminator is used to distinguish between true and false results generated by the first generator. The second generator is used to generate a second sample based on the input first sample and to perform arc segmentation on the discharge image in the second sample; the second discriminator is used to distinguish between true and false results generated by the second generator.

[0009] Based on a further improvement of the above method, the first generator and the second generator have the same structure; the first generator includes: An encoder is used to extract shallow features from the input source data using a multi-layer downsampling structure; The decoder is used to decode the input features using a multi-layer upsampling structure and output decoded features; the encoder and decoder are connected by a skip connection, which is used to transmit the shallow features downsampled by each layer of the encoder to the corresponding upsampling layer of the encoder as the input of the corresponding upsampling layer of the decoder; The generation module is used to generate the first sample based on the decoding features output from the last upsampling layer of the decoder; The segmentation module is used to generate a contour segmentation image of the discharge image based on the decoded features.

[0010] Based on the further improvements to the above method, the loss of the generative adversarial network model is calculated using the following formula:

[0011] in, This represents the discharge image in the input data of the first generator. This represents the discharge image in the first sample generated by the first generator. This represents the arc segmentation diagram corresponding to the discharge image in the input data of the first generator. Indicates arc segmentation loss, Indicates the feature reconstruction loss. This represents the discharge image in the second sample generated by the second generator. Indicates discriminator loss. , , and Indicates the weighting coefficient. Indicates parameters, This represents the 1-norm of a matrix.

[0012] Based on further improvements to the above method, the arc segmentation loss is calculated using the following formula. :

[0013] in, This represents the arc segmentation image obtained by the first generator performing arc segmentation on the discharge image in the first sample. This represents the arc segmentation image obtained by the second generator performing arc segmentation on the discharge image in the second sample.

[0014] Based on further improvements to the above method, the collected electrical signals include the current signals generated each time leakage current occurs. Feature data is extracted based on the electrical signal at each leakage current occurrence, including: Extract the maximum value of the current signal at each leakage current occurrence as the peak value of the leakage current, and calculate the normalized standard deviation of the peak value of the leakage current. Calculate the correlation coefficient of the current signals when leakage current occurs in two consecutive instances. The characteristic data corresponding to the electrical signal include the peak value of the leakage current, the standard deviation of the peak value of the leakage current, and the correlation coefficient of the current signal.

[0015] Based on the further improvement of the above method, the correlation coefficient of the current signal when leakage current is generated in two consecutive instances is calculated using the following formula:

[0016] in, This represents the leakage current value at the j-th sampling time when the leakage current is generated for the α-th time. Indicates the total number of sampling points. This represents the average leakage current at all sampling times when the leakage current is generated for the αth time. This represents the average leakage current at all sampling times when the leakage current is generated for the (α-1)th time.

[0017] Based on a further improvement of the above method, the characteristic data corresponding to the acoustic signal includes: center frequency, spectral amplitude, and average power; the characteristic data corresponding to the acoustic signal is obtained in the following manner: The acoustic signal is subjected to a fast Fourier transform to obtain its spectrum; Extract the center frequency and amplitude corresponding to all spectral peaks in the spectrum; calculate the average power of the acoustic signal to obtain the characteristic data corresponding to the acoustic signal.

[0018] Compared with existing technologies, this invention extracts electrical signals, acoustic signals, and discharge images when leakage current is generated, and constructs samples by extracting multi-dimensional feature data. The extracted multi-dimensional features help to more accurately determine the state of the insulating material. To avoid overfitting of the model due to data imbalance and improve detection accuracy, the samples are expanded to achieve a balanced training sample set. The constructed multi-class neural network model is trained based on the expanded training sample set, so that the trained model can accurately and efficiently monitor the state of solid insulating materials, and thus accurately determine the performance of solid insulating materials. This is of great significance for the selection and maintenance of solid insulating materials, the comprehensive evaluation of electrical insulation performance, and ultimately the improvement of the operational safety and power supply reliability of power systems.

[0019] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0020] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a flowchart of a training method for a state detection model of solid insulating materials according to an embodiment of the present invention. Detailed Implementation

[0021] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0022] A specific embodiment of the present invention discloses a training method for a state detection model of solid insulating materials, such as... Figure 1As shown, it includes the following steps: S1. Apply voltage to the solid insulating material until surface flashover occurs, and collect electrical signals, acoustic signals and discharge images when the solid insulating material generates leakage current each time during the pressure application process; S2. Extract multi-dimensional feature data based on the electrical signal, acoustic signal, and discharge image at each leakage current generation; construct a training sample set based on the multi-dimensional feature data and the corresponding state labels; the state labels include no obvious discharge, corona discharge, and flashover. S3. Expand the samples in the training sample set to obtain an expanded training sample set; S4. Construct a multi-class neural network model, and train the multi-class neural network model based on the expanded training sample set to obtain a state detection model for solid insulating materials.

[0023] During implementation, solid insulating material is pressed under the finger electrodes of the flashover voltage testing platform. Voltage is applied to the solid insulating material in stages until surface flashover occurs. For example, a voltage of 0.5 kV is applied initially, stabilized for a period, then increased to 1 kV, stabilized for a period, then increased to 1.5 kV, and so on, until surface flashover occurs. Electrical signals, acoustic signals, and discharge images are collected each time leakage current is generated by the solid insulating material during the pressure application process. Electrical sensors, such as current sensors, are used to collect current; acoustic acquisition devices are used to collect acoustic signals; and cameras are used to capture images.

[0024] It should be noted that the states of solid insulating materials include no obvious discharge state, corona state, and flashover state. The generation of leakage current does not mean that the solid insulating material is in a discharge state.

[0025] The development of surface flashover is accompanied by strong electrical, acoustic, and optical phenomena. During the application of voltage, the gas on the surface of the solid insulating material near the test electrode is excited, resulting in a dark purple corona discharge. At the same time, the collision of charges inside the corona and the vibration of the surrounding air during gas ionization will cause obvious acoustic signals until the electrodes are completely broken down to form a bright flashover channel.

[0026] It should be noted that a multi-class neural network model is constructed to predict the type of solid insulating material, that is, to determine whether the state of the solid insulating material is no obvious discharge, corona, or flashover.

[0027] Compared with existing technologies, the training method for the state detection model of solid insulation materials provided in this embodiment extracts electrical signals, acoustic signals, and discharge images when leakage current is generated, and constructs samples by extracting multi-dimensional feature data. The extracted multi-dimensional features help to more accurately judge the state of the insulation material. In order to avoid overfitting of the model due to data imbalance and improve detection accuracy, the samples are expanded to achieve a balanced training sample set. The multi-class neural network model is trained based on the expanded training sample set, so that the trained model can accurately and efficiently monitor the state of solid insulation materials, and thus accurately judge the performance of solid insulation materials. This is of great significance for the fault prediction and health management of solid insulation materials, the selection and maintenance of solid insulation materials, the comprehensive evaluation of electrical insulation performance, and the improvement of the operational safety and power supply reliability of power systems.

[0028] During the pressure application process, the leakage current of the solid insulation material is monitored. The rising edge of the current signal is used as a trigger signal, and the electrical signal, acoustic signal, and discharge image are acquired at this time. That is, for each trigger, an image at the trigger time, as well as electrical and acoustic signals for a period of time, are acquired. In practice, the trigger threshold current is 20 μA.

[0029] Based on the collected electrical signals, acoustic signals, and discharge images, corresponding feature data are extracted to form multidimensional feature data.

[0030] Specifically, the collected electrical signals include the current signal and the applied voltage each time the solid insulating material generates leakage current; Feature data is extracted based on the electrical signal at each leakage current occurrence, including: Extract the maximum value of the current signal at each leakage current occurrence as the peak value of the leakage current, and calculate the normalized standard deviation of the peak value of the leakage current. Calculate the correlation coefficient of the current signals when leakage current occurs in two consecutive instances. The characteristics corresponding to the electrical signal include the peak value of the leakage current, the standard deviation of the peak value of the leakage current, and the correlation coefficient of the current signal.

[0031] During implementation, for the αth trigger, i.e., when the leakage current is generated for the αth time, an electrical signal is collected for a period of time, along with the voltage applied to the solid insulating material at the time of triggering. The collection duration is set to... When the sampling frequency is At that time, the total number of sampling points for the α-th trigger acquisition is The collected electrical signals are represented as , .

[0032] Then, the maximum value in the current signal when the leakage current is generated for the αth time is calculated as the peak value of the leakage current when the leakage current is generated for the αth time, denoted as: .

[0033] Calculate the normalized standard deviation of the peak leakage current. .

[0034] in, This represents the r-th peak value of the leakage current when the α-th leakage current is generated. This represents the average of all leakage current peak values. This represents the number of peak values ​​of the leakage current when the leakage current is generated for the αth time.

[0035] During the transition from corona to flashover in solid materials, leakage current can dynamically reflect the discharge status of the insulating material surface. By calculating the correlation coefficient between leakage currents that exceed the trigger threshold current each time, the temporal variation trend of leakage current over the pressurization time can be quantitatively identified, and the discharge status of the insulating material surface can be dynamically evaluated.

[0036] Specifically, the Pearson correlation coefficient of the current signals at two consecutive leakage current occurrences is calculated using the following formula:

[0037] in, This represents the leakage current value at the j-th sampling time when the leakage current is generated for the α-th time. Indicates the total number of sampling points. This represents the average leakage current at all sampling times when the leakage current is generated for the αth time. This represents the average leakage current at all sampling times when the leakage current is generated for the (α-1)th time.

[0038] The acoustic signal during the voltage boosting process is caused by the vibration of the surrounding air due to the collision and ionization of charges generated by the excitation of gas near the test electrode on the surface of the solid insulating material. The electrical signal and the acoustic signal are acquired synchronously. During the αth trigger, the acoustic signal is acquired through the acoustic acquisition module.

[0039] Specifically, the characteristic data corresponding to the acoustic signal includes: center frequency, spectral amplitude, and average power; the characteristic data corresponding to the acoustic signal is obtained in the following manner: The acoustic signal is subjected to a fast Fourier transform to obtain its spectrum; Extract the center frequency and amplitude corresponding to all spectral peaks in the spectrum; calculate the average power of the acoustic signal to obtain the characteristic data corresponding to the acoustic signal.

[0040] Set the collection time to When the sampling frequency is At that time, the total number of sampling points for the acoustic signal acquired during the α-th trigger is The collected acoustic signals are represented as , . This represents the signal at the j-th sampling point of the acoustic signal acquired during the α-th trigger.

[0041] First, the collected acoustic signal is subjected to a fast Fourier transform to obtain the spectrum of the acoustic signal.

[0042] Extract the center frequency corresponding to all spectral peaks in the spectrum. and amplitude This serves as a basis for assessing the changes in acoustic signal characteristics caused by the radiation waves generated by charge collisions and the accompanying air vibrations during the reciprocating migration of ions in space before and during flashover. , This indicates the number of frequencies in the spectrum.

[0043] The energy of the acoustic signal varies under different conditions. The energy of the acoustic signal changes with the amplitude of the applied voltage and the intensity of the discharge phenomenon during the test. Therefore, the average power of the acoustic signal is calculated to quantitatively reflect the energy of the acoustic spectrum.

[0044] in, This represents the total number of sampling points for the acoustic signal acquired during the α-th trigger. This represents the signal at the j-th sampling point of the acoustic signal acquired during the α-th trigger.

[0045] When solid insulating materials are in a state without obvious discharge, the images taken show almost no significant changes due to the absence of corona discharge. When corona discharge occurs, the images show obvious diffuse corona. When dry flashover occurs, the flashover discharge image shows a through-conductive channel, with an increased number of arc channels and a thicker shape. When wet flashover occurs, the arc distribution is bent under the influence of droplets. The images differ under different conditions. Therefore, by acquiring discharge images, the state of solid insulating materials can be further accurately monitored, providing a basis for predicting flashover risks.

[0046] Specifically, the features corresponding to the discharge image include discharge brightness and electrical damage area; the features corresponding to the discharge image are obtained in the following manner: The discharge brightness is obtained by calculating the standard deviation of gray levels in the region of the discharge image where the gray level is greater than the first threshold. The electrical damage area is obtained by calculating the percentage of the region with a gray level greater than the first threshold in the discharge image.

[0047] It should be noted that the features corresponding to the discharge image also include the discharge image itself.

[0048] Solid insulating materials exhibit different discharge brightness and electrical damage area under different conditions. To accurately detect the electrical damage area, the discharge image is converted into a grayscale image. The standard deviation of the grayscale value of the region in the discharge image with a grayscale value greater than a first threshold is calculated to obtain the discharge brightness. The percentage of the region with a grayscale value greater than the first threshold in the discharge image is then calculated to obtain the electrical damage area.

[0049] A sample set is constructed based on the multi-dimensional characteristic data corresponding to each triggering of leakage current and the state of the solid insulating material. Characteristic data of different insulating materials can be collected to form big data, which can then be stored and managed, providing data support for evaluating the surface withstand voltage performance and long-term aging insulation performance of solid insulating materials.

[0050] Because the number of samples in the flashover state is relatively small compared to other states in the collected data, there is a problem of imbalanced samples. Directly training the model would lead to overfitting and affect the accuracy of the model. Therefore, the samples are first expanded to improve the accuracy of the subsequent classification model.

[0051] Specifically, the training sample set is expanded to obtain an expanded training sample set, including: Extract samples labeled as flashover from the training sample set, and stitch together the numerical features of the extracted samples with the corresponding discharge images to form source data; The source data is input into the trained generative adversarial network model to generate new flashover state samples. The generated new flashover state samples are then added to the training sample set to obtain the expanded training sample set.

[0052] To generate samples similar to real samples, the numerical features of samples labeled as flashover in the training sample set and the corresponding discharge images are stitched together to form source data. Samples are then generated based on the source data, making the generated samples more realistic.

[0053] In practice, numerical features are converted into one-dimensional vectors, and the converted one-dimensional vectors are then converted into LxW form and stitched together with the discharge image as source data, where W is the width of the discharge image.

[0054] By stitching together other features and discharge images to form a complete sample, a new sample is generated. This process explores the relationship between other features and discharge images, making the generated sample more realistic and more similar to real samples.

[0055] Specifically, the constructed generative adversarial network model includes a first generator, a second generator, a first discriminator, and a second discriminator; The first generator is used to generate a first sample based on the input source data and to perform arc segmentation on the discharge image in the first sample; the first discriminator is used to distinguish between true and false results generated by the first generator. The second generator is used to generate a second sample based on the input first sample and to perform arc segmentation on the discharge image in the second sample; the second discriminator is used to distinguish between true and false results generated by the second generator.

[0056] By combining the arc segmentation task with the sample generation task, the two tasks complement each other, thereby improving the efficiency of sample generation. The generated samples have a high similarity to real samples, thus providing a data foundation for subsequent training of classification models.

[0057] Specifically, the first generator and the second generator have the same structure; the first generator includes: An encoder is used to extract shallow features from the input source data using a multi-layer downsampling structure; Decoder; used to decode input features using a multi-layer upsampling structure and output decoded features; the encoder and decoder are connected by a skip connection, used to transmit shallow features downsampled by each layer of the encoder to the corresponding upsampling layer of the encoder as input to the corresponding upsampling layer of the decoder; The generation module is used to generate the first sample based on the decoding features output from the last upsampling layer of the decoder; The segmentation module is used to generate a contour segmentation image of the discharge image based on the decoded features.

[0058] In implementation, the first generator and the second generator have the same structure. The first generator and the second generator can adopt a symmetric structure model, that is, a model that includes an encoder and a decoder, such as Unet.

[0059] During implementation, a preprocessing layer can be added before the encoder to preprocess the input source data, extract initial features, and send the initial features to the encoder.

[0060] The encoder employs a multi-layer downsampling structure to extract shallow features from the input source data. In implementation, the encoder includes multiple convolutional modules and downsampling layers, with each convolutional module followed by a downsampling layer, progressively extracting shallow features and reducing the size of the feature map. Each convolutional module includes two instance normalization layers, two convolutional layers with a stride of 1, and two Leaky ReLU activation functions. Each downsampling layer contains a 2D convolutional layer.

[0061] The decoder consists of multiple convolutional modules and upsampling layers. Each convolutional module is followed by an upsampling layer, used to decode the input features using a multi-layer upsampling structure to output decoded features. The decoder's convolutional modules have the same structure as the encoder's convolutional modules, using deconvolution as the upsampling layer. The decoder decodes the features progressively and gradually increases the size of the feature map.

[0062] To improve the quality of the generated image, a skip connection is connected between the encoder and decoder. This skip connection transmits the shallow features downsampled from each layer of the encoder to the corresponding upsampled layer of the encoder, serving as the input to the corresponding upsampled layer of the decoder. The skip connection fuses features at corresponding locations in the encoder and decoder, thereby improving the quality of generation and segmentation.

[0063] After the decoder extracts the decoding features, a sample generation module and an arc segmentation module are added to realize the generation of samples and the segmentation of arcs.

[0064] In implementation, the sample generation module includes a 2D convolutional layer and a LeakyReLU activation layer, and the arc segmentation module has the same structure as the sample generation module.

[0065] In implementation, both the first and second discriminators can use PatchGAN. During implementation, the discharge images of samples in flashover state are annotated with arc contours, marking the arc contour by labeling the entire arc portion as 1 and other portions as 0.

[0066] During implementation, the loss of the generative adversarial network model is calculated using the following formula:

[0067] in, This represents the discharge image in the input data of the first generator. This represents the discharge image in the first sample generated by the first generator. This represents the arc segmentation diagram corresponding to the discharge image in the input data of the first generator. Indicates arc segmentation loss, Indicates the feature reconstruction loss. This represents the discharge image in the second sample generated by the second generator. Indicates discriminator loss. , , and Indicates the weighting coefficient. Indicates parameters, This represents the 1-norm of a matrix.

[0068] During implementation, the arc segmentation loss is calculated using the following formula. :

[0069] in, This represents the arc segmentation image obtained by the first generator performing arc segmentation on the discharge image in the first sample. This represents the arc segmentation image obtained by the second generator performing arc segmentation on the discharge image in the second sample.

[0070] During implementation, the feature reconstruction loss is calculated using the following formula. :

[0071] in, This indicates the number of feature extraction layers in the pre-trained neural network. This represents the feature map output by the j-th layer of the pre-trained neural network. This represents the number of channels in the feature map output by the j-th layer of the pre-trained neural network. This represents the height of the feature map output by the j-th layer of the pre-trained neural network. This represents the width of the feature map output by the j-th layer of the pre-trained neural network. This represents the 2-norm of a matrix.

[0072] In practice, the feature reconstruction loss is obtained by taking the discharge image from the input data of the first generator. Discharge image in the second sample generated by the second generator The pre-trained neural network model is input to calculate the feature reconstruction loss. The pre-trained neural network model can be a VGG network model. The pre-trained neural network model is used as a feature extractor to perform distance measurement in high-dimensional space. The feature reconstruction loss is used to constrain the model in high-dimensional space, thereby improving the generalization of the model.

[0073] During implementation, the discriminator loss is calculated using the following formula. :

[0074] Where E[·] represents expectation, This represents the discrimination result of the first discriminator on the first sample generated by the first generator. This represents the discrimination result of the second discriminator on the second sample generated by the second generator.

[0075] By adding arc segmentation constraints and feature reconstruction loss constraints, the model can more accurately capture the arc features of flashover discharge images, making the generated samples more realistic and reliable.

[0076] Once the training loss reaches the preset accuracy requirement or the training iterations are completed, training of the generative adversarial network (GAN) model is stopped, resulting in a trained GAN model. The source data corresponding to the real flashover samples is then input into the trained GAN model to obtain the first sample, which is then added to the training sample set to expand the training sample set.

[0077] Based on the expanded training sample set, a multi-class neural network model for detecting the state of solid insulating materials is trained to obtain a trained state detection model. In practice, the multi-class neural network model for detecting the state of solid insulating materials includes: The first feature extraction module is used to extract the first feature based on the numerical feature data in the multidimensional feature data. The second feature extraction module is used to extract the second feature based on the image feature data of the multidimensional feature data. A feature fusion module is used to fuse the first feature and the second feature to obtain a fused feature; The classification module is used to classify the state of insulators based on the fused features.

[0078] Since the features corresponding to the current signal and sound signal, as well as some features of the surface image, are numerical, while the surface image is image data, features need to be extracted separately. In implementation, the first feature extraction module can use existing machine learning models or deep learning models, such as a feedforward neural network structure, while the second feature extraction module can use a convolutional network structure.

[0079] The feature fusion module fuses the extracted first and second features. In practice, the fusion can be performed by splicing.

[0080] The classification module classifies the insulator state based on the fused features, that is, it predicts the probability that the insulator belongs to a state without obvious discharge, corona and flashover.

[0081] Based on the trained state assessment model, the electrical signal data, acoustic signal data, and image data of the solid insulating material to be evaluated when leakage current is generated are obtained and the corresponding multi-dimensional features are input into the state detection model to obtain the state detection result of the solid insulating material to be evaluated.

[0082] During long-term service, solid insulating materials gradually transition from a state without significant discharge to a corona state, and eventually develop into flashover. Therefore, by judging the current state of solid insulating materials, their performance and aging degree can be assessed.

[0083] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0084] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A training method for a state detection model of a solid insulating material, characterized in that, Includes the following steps: A voltage is applied to a solid insulating material until a surface flashover occurs, and electrical signals, acoustic signals, and discharge images are collected each time the solid insulating material generates leakage current during the pressure application process. Multidimensional feature data are extracted based on the electrical signals, acoustic signals, and discharge images generated each time a leakage current is generated; A training sample set is constructed based on the multidimensional feature data and the corresponding state labels; the state labels include no obvious discharge, corona discharge, and flashover. The samples in the training sample set are expanded to obtain the expanded training sample set; A multi-class neural network model is constructed, and the multi-class neural network model is trained based on the expanded training sample set to obtain a state detection model for solid insulating materials; The training sample set is augmented to obtain an augmented training sample set, including: Extract samples labeled as flashover from the training sample set, and stitch together the numerical features of the extracted samples with the corresponding discharge images to form source data; The source data is input into the trained generative adversarial network model to generate new flashover state samples. The generated new flashover state samples are then added to the training sample set to obtain the expanded training sample set.

2. The training method for the state detection model of solid insulating materials according to claim 1, characterized in that, The multi-class neural network model includes: The first feature extraction module is used to extract the first feature based on the numerical feature data in the multidimensional feature data. The second feature extraction module is used to extract the second feature based on the image feature data of the multidimensional feature data. A feature fusion module is used to fuse the first feature and the second feature to obtain a fused feature; The classification module is used to classify the state of insulators based on the fused features.

3. The training method for the state detection model of solid insulating materials according to claim 1, characterized in that, The generative adversarial network model includes a first generator, a second generator, a first discriminator, and a second discriminator; The first generator is used to generate a first sample based on the input source data and to perform arc segmentation on the discharge image in the first sample; The first discriminator is used to determine the authenticity of the results generated by the first generator; The second generator is used to generate a second sample based on the input first sample and to perform arc segmentation on the discharge image in the second sample; the second discriminator is used to distinguish between true and false results generated by the second generator.

4. The training method for the state detection model of solid insulating materials according to claim 3, characterized in that, The first generator and the second generator have the same structure; the first generator includes: An encoder is used to extract shallow features from the input source data using a multi-layer downsampling structure; The decoder is used to decode the input features using a multi-layer upsampling structure and output decoded features; the encoder and decoder are connected by a skip connection, which is used to transmit the shallow features downsampled by each layer of the encoder to the corresponding upsampling layer of the encoder as the input of the corresponding upsampling layer of the decoder; The generation module is used to generate the first sample based on the decoding features output from the last upsampling layer of the decoder; The segmentation module is used to generate a contour segmentation image of the discharge image based on the decoded features.

5. The training method for the state detection model of solid insulating materials according to claim 3, characterized in that, The loss of the generative adversarial network model is calculated using the following formula: in, This represents the discharge image in the input data of the first generator. This represents the discharge image in the first sample generated by the first generator. This represents the arc segmentation diagram corresponding to the discharge image in the input data of the first generator. Indicates arc segmentation loss, Indicates the feature reconstruction loss. This represents the discharge image in the second sample generated by the second generator. Indicates discriminator loss. , , and Indicates the weighting coefficient. Indicates parameters, This represents the 1-norm of a matrix.

6. The training method for the state detection model of solid insulating materials according to claim 5, characterized in that, The arc segmentation loss is calculated using the following formula. : in, This represents the arc segmentation image obtained by the first generator performing arc segmentation on the discharge image in the first sample. This represents the arc segmentation image obtained by the second generator performing arc segmentation on the discharge image in the second sample.

7. The training method for the state detection model of solid insulating materials according to claim 1, characterized in that, The collected electrical signals include the current signals generated each time leakage current occurs; Feature data is extracted based on the electrical signal at each leakage current occurrence, including: Extract the maximum value of the current signal at each leakage current occurrence as the peak value of the leakage current, and calculate the normalized standard deviation of the peak value of the leakage current. Calculate the correlation coefficient of the current signals when leakage current occurs in two consecutive instances. The characteristic data corresponding to the electrical signal include the peak value of the leakage current, the standard deviation of the peak value of the leakage current, and the correlation coefficient of the current signal.

8. The training method for the state detection model of solid insulating materials according to claim 7, characterized in that, The correlation coefficient of the current signals when leakage current occurs in two consecutive instances is calculated using the following formula: in, This represents the leakage current value at the j-th sampling time when the leakage current is generated for the α-th time. Indicates the total number of sampling points. This represents the average leakage current at all sampling times when the leakage current is generated for the αth time. This represents the average leakage current at all sampling times when the leakage current is generated for the (α-1)th time.

9. The training method for the state detection model of solid insulating materials according to claim 1, characterized in that, The characteristic data corresponding to the acoustic signal includes: center frequency, spectral amplitude, and average power; the characteristic data corresponding to the acoustic signal is obtained in the following manner: The acoustic signal is subjected to a fast Fourier transform to obtain its spectrum; Extract the center frequency and amplitude corresponding to all spectral peaks in the spectrum; calculate the average power of the acoustic signal to obtain the characteristic data corresponding to the acoustic signal.

Images