An insulator flashover risk warning system

Abstract

The present invention relates to an insulator flashover risk warning system, which belongs to the technical field of insulators and solves the problem of low accuracy and efficiency of flashover risk warning in the prior art. The system includes: a current acquisition unit to acquire the current signal on the surface of the insulator; an acoustic signal acquisition unit, which is used to acquire the sound signal and send it to the processing unit; a shooting unit, which is used to acquire the surface image of the insulator and send it to the processing unit; a processing unit, which is used to perform analog-to-digital conversion on the current signal and the acoustic signal and send the converted current signal, acoustic signal and surface image to the monitoring center; a monitoring center, which is used to extract multi-dimensional feature data according to the converted current signal, acoustic signal and surface image; predict the current flashover risk of the insulator based on a trained flashover risk prediction model; when the flashover risk is greater than the preset risk threshold, send an alarm message to the alarm unit; the alarm unit is used to make an alarm. An efficient and accurate flashover risk warning for insulators is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of insulators, and in particular to an insulator flashover risk early warning system. Background Art

[0002] Insulators are devices installed between conductors at different potentials or between conductors and grounded components. Their main functions are to achieve electrical insulation and mechanical fixation. There is a special physical phenomenon in the insulation system composed of vacuum and insulators: the voltage withstand capacity of the insulator in a vacuum environment will be greatly reduced, even lower than the insulation strength of the vacuum gap of the same distance or the solid insulator itself. The root cause is that a penetrating electrical breakdown process occurs on the surface of the insulator along the vacuum-insulator interface, namely, surface flashover. The surface flashover phenomenon of vacuum insulation in high-voltage vacuum equipment not only seriously restricts the voltage withstand capacity of the equipment, affects the normal operation of related equipment, but may also cause huge economic losses.

[0003] With the increase of voltage level and transmission capacity of power system, higher requirements are put forward for the electrical insulation performance of insulators. Traditional flashover voltage measurement methods mostly start from the structural parameters or physical parameters of the insulation body, and then fit and calculate the critical flashover voltage value, that is, record the instantaneous voltage value of insulation surface breakdown, ignoring the surface charge accumulation characteristics and aging characteristics of the insulation material under long-term service. This limits the prediction of whether surface flashover will occur. It is difficult to accurately and efficiently evaluate the potential threat of sudden surface flashover under continuous high electric field strength after the insulator surface deteriorates under long-term stable operation, resulting in the failure of timely flashover warning, causing losses to power companies and bringing serious harm to the national economy. Summary of the invention

[0004] In view of the above analysis, an embodiment of the present invention aims to provide an insulator flashover risk warning system to solve the problems of low accuracy and efficiency in existing flashover risk assessment.

[0005] On the one hand, an embodiment of the present invention provides an insulator flashover risk warning system, comprising:

[0006] A current collection unit is used to collect current signals when leakage current is generated on the surface of the insulator; when the current signal exceeds a preset threshold, a trigger signal is sent to the sound signal collection unit and the shooting unit, and the collected current signal is sent to the processing unit;

[0007] an acoustic signal collecting unit, configured to collect the acoustic signal after receiving the trigger signal, and send the collected acoustic signal to the processing unit;

[0008] A shooting unit, used for collecting an image of the insulator surface after receiving the trigger signal, and sending the captured surface image to a processing unit;

[0009] a processing unit, configured to perform analog-to-digital conversion on the current signal and the acoustic signal and to send the converted current signal, acoustic signal and surface image to a monitoring center;

[0010] A monitoring center is used to extract multi-dimensional feature data according to the converted current signal, acoustic signal and surface image; predict the current flashover risk of the insulator based on the trained flashover risk prediction model; and send an alarm message to the alarm unit when the flashover risk is greater than a preset risk threshold;

[0011] An alarm unit is used to generate an alarm according to the alarm information.

[0012] Based on the further improvement of the above system, the current acquisition unit includes a current sensor, an amplifier and a trigger circuit;

[0013] The current sensor adopts a non-contact through-core leakage current sensor, which is used to collect the leakage current signal of the insulator;

[0014] The amplifier is used to amplify the collected leakage current signal;

[0015] The trigger circuit is used to compare the amplified leakage current signal with a preset threshold value, and output a high level signal as a trigger signal when the leakage current signal is greater than the preset threshold value.

[0016] Based on the further improvement of the above system, the flashover risk prediction model is a multi-classification neural network model; the multi-classification neural network model includes:

[0017] A first feature extraction module, used for extracting a first feature according to the converted current signal and acoustic signal;

[0018] A second feature extraction module, used to extract a second feature from the surface image;

[0019] A feature fusion module, used for fusing the first feature and the second feature to obtain a fused feature;

[0020] A classification module, used for classifying the insulator status based on the fusion features;

[0021] The risk calculation module is used to calculate the insulator flashover risk based on the insulator state classification results.

[0022] Based on the further improvement of the above system, the trained flashover risk prediction model is obtained in the following way:

[0023] Collect multi-dimensional feature data of insulators in different states; use the multi-dimensional feature data as input data of training samples and use the states of insulators as labels to construct a training sample set; the different states include no obvious discharge, corona and flashover states;

[0024] Constructing a multi-classification neural network model; training the multi-classification neural network model based on the training sample set, and using the trained multi-classification neural network model as a trained flashover risk prediction model.

[0025] Based on the further improvement of the above system, after constructing the training sample set and before training the multi-classification neural network model based on the training sample set, it also includes:

[0026] Extracting samples labeled as flashover from the training sample set, and splicing the multi-dimensional features of the extracted samples and the corresponding surface images to form source data;

[0027] The source data is input into the trained generative adversarial network model to generate new flashover state samples, and the generated new flashover state samples are added to the training sample set to expand the training sample set.

[0028] Based on the further improvement of the above system, the generative adversarial network model includes a first generator, a second generator, a first discriminator and a second discriminator;

[0029] The first generator is used to generate a first sample according to input data and perform arc segmentation on the surface image in the first sample; the first discriminator is used to perform authenticity discrimination on the result generated by the first generator;

[0030] The second generator is used to generate a second sample according to the input first sample and perform arc segmentation on the surface image in the second sample; the second discriminator is used to perform authenticity discrimination on the result generated by the second generator.

[0031] Based on further improvement of the above system, the first generator and the second generator have the same structure; the first generator includes:

[0032] An encoder, for extracting shallow features of input source data using a multi-layer downsampling structure;

[0033] A decoder is used to decode the input features and output decoded features using a multi-layer upsampling structure; a jump connection is connected between the encoder and the decoder, which is used to transmit the shallow features of each layer downsampled by the encoder to the upsampling layer corresponding to the encoder as the input of the upsampling layer corresponding to the encoder;

[0034] A generation module, used to generate a first sample according to the decoding features output by the last upsampling layer of the decoder;

[0035] The segmentation module is used to generate a contour segmentation image of the surface image according to the decoded features.

[0036] Based on the further improvement of the above system, the following formula is used to calculate the loss of the generative adversarial network model:

[0037] Loss=θ1(‖Xp-Yp‖1+β‖Xp·S-Yp·S‖1)+θ2L dice

[0038] +θ3L feat (Xp,X p ′ )+θ4L D

[0039] Wherein, Xp represents the surface image in the input data of the first generator, Yp represents the surface image in the first sample generated by the first generator, S represents the arc segmentation map corresponding to the surface image in the input data of the first generator, and L dice Indicates arc segmentation loss, L feat (·,·) represents the feature reconstruction loss, X p ′ represents the surface image in the second sample generated by the second generator, L D represents the discriminator loss, θ1, θ2, θ3 and θ4 represent weight coefficients, β represents parameters, and ‖·‖1 represents the norm of the matrix.

[0040] Based on the further improvement of the above system, the arc segmentation loss L is calculated using the following formula dice :

[0041]

[0042] Among them, S Yp represents the arc segmentation image obtained by the first generator performing arc segmentation on the surface image in the first sample, It represents the arc segmentation image obtained by the second generator performing arc segmentation on the surface image in the second sample.

[0043] Based on the further improvement of the above system, the risk calculation module calculates the flashover risk probability of the insulator based on the insulator state classification results using the following formula:

[0044] Risk=ω1p1+ω2p2+ω3p3

[0045] Among them, p1 represents the probability that the insulator is in the flashover state predicted by the model, p2 represents the probability that the insulator is in the corona state predicted by the model, p3 represents the probability that the insulator is in the no obvious discharge state predicted by the model, ω1, ω2 and ω3 represent weights, where ω1>ω2>ω3.

[0046] Compared with the prior art, the present invention collects leakage current through a current collection unit, collects acoustic signals through an acoustic signal collection unit, collects images when leakage current is generated through a shooting unit, and extracts electrical, acoustic and optical multi-dimensional features through current signals, acoustic signals and surface images, thereby accurately and efficiently obtaining the risk of flashover of insulators based on the electrical, acoustic and optical multi-dimensional features. When the risk exceeds a threshold, an alarm is promptly issued through an alarm unit, which is convenient for staff to perform maintenance and repair in a timely manner, thereby improving the operational safety of the power system, which is of great significance to power supply reliability.

[0047] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0049] Figure 1 The block diagram of the insulator flashover risk warning system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a 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 used to limit the scope of the present invention.

[0051] A specific embodiment of the present invention discloses an insulator flashover risk warning system, such as Figure 1 As shown, including:

[0052] A current collection unit is used to collect current signals when leakage current is generated on the surface of the insulator; when the current signal exceeds a preset threshold, a trigger signal is sent to the sound signal collection unit and the shooting unit, and the collected current signal is sent to the processing unit;

[0053] an acoustic signal collecting unit, configured to collect the acoustic signal after receiving the trigger signal, and send the collected acoustic signal to the processing unit;

[0054] A shooting unit, used for collecting an image of the insulator surface after receiving the trigger signal, and sending the captured surface image to a processing unit;

[0055] a processing unit, configured to perform analog-to-digital conversion on the current signal and the acoustic signal and to send the converted current signal, acoustic signal and surface image to a monitoring center;

[0056] A monitoring center is used to extract multi-dimensional feature data according to the converted current signal, acoustic signal and surface image; predict the current flashover risk of the insulator based on the trained flashover risk prediction model; and send an alarm message to the alarm unit when the flashover risk is greater than a preset risk threshold;

[0057] An alarm unit is used to generate an alarm according to the alarm information.

[0058] 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 insulator near the test electrode is excited, and dark purple corona discharge occurs; at the same time, the collision of charges inside the corona and the vibration of the surrounding air during the gas ionization process will cause obvious acoustic signals until the electrodes are completely broken down to form a bright flashover channel.

[0059] Compared with the prior art, the insulator flashover risk warning system provided in the present embodiment collects leakage current through a current collection unit, collects acoustic signals through an acoustic signal collection unit, collects images when leakage current is generated through a shooting unit, and extracts electrical, acoustic, and optical multi-dimensional features through current signals, acoustic signals, and surface images, thereby accurately and efficiently obtaining the risk of insulator flashover based on the electrical, acoustic, and optical multi-dimensional features. When the risk exceeds a threshold, an alarm is promptly sounded through an alarm unit, thereby realizing fault prediction and health management of insulators, facilitating timely maintenance and inspection by staff, thereby improving the operational safety of the power system, which is of great significance to power supply reliability.

[0060] During implementation, the monitoring center can use an industrial cloud platform or an industrial Internet platform.

[0061] Specifically, the current acquisition unit includes a current sensor, an amplifier and a trigger circuit;

[0062] The current sensor adopts a non-contact through-core leakage current sensor, which is used to collect the leakage current signal of the insulator;

[0063] The amplifier is used to amplify the collected leakage current signal;

[0064] The trigger circuit is used to compare the amplified leakage current signal with a preset threshold value, and output a high level signal as a trigger signal when the leakage current signal is greater than the preset threshold value.

[0065] During implementation, the current collection unit uses an electrical sensor. Since the through-core current sensor has installation positions on several connecting fittings of the insulator and the tower, it is convenient to realize live operation, so the current sensor uses a through-core leakage current sensor. The through-core leakage current sensor is set on the surface of the insulator to intercept the leakage current on the surface of the insulator. In order to make the copper ring and the insulator fully contact, conductive glue can be applied during installation, which also plays a fixing role. The collected leakage current is drained to the amplifier through the shielded twisted wire.

[0066] Since the current signal may be relatively weak, the signal is amplified by an amplifier for subsequent use.

[0067] During implementation, the amplified current is compared with the preset threshold through the trigger circuit. The tiny current can be ignored. The sound signal and image are collected only when the current exceeds the threshold. During implementation, the trigger circuit can use a comparator. The preset threshold can be set to 20μA ~ 100μA.

[0068] During implementation, the acoustic signal collection unit can be installed on the hardware, and the shooting unit can be placed outside around the insulator.

[0069] The data collected by the current acquisition unit, the sound signal acquisition unit and the shooting unit are transmitted to the processing unit. Since the current signal and the sound signal are both analog data, they need to be converted into digital signals first. The processing unit transmits the converted sound signal data, electrical signal data and surface image to the monitoring center. During implementation, the transmission can be carried out by wireless transmission.

[0070] The monitoring center extracts multi-dimensional feature data based on the converted current signal, acoustic signal and surface image, and inputs the extracted multi-dimensional feature data into the trained flashover risk prediction model to obtain the risk probability of the insulator; if the calculated risk probability is greater than the preset risk threshold, an alarm is issued through the alarm unit. During implementation, the alarm information sent by the monitoring center to the alarm unit includes the location of the insulator, risk level and other information, so that the staff can perform timely maintenance according to the importance and urgency.

[0071] Specifically, the characteristic data extracted by the monitoring center according to the converted current signal include the leakage current peak value, the standard deviation of the leakage current peak value and the correlation coefficient of the current signal;

[0072] The leakage current peak value, the standard deviation of the leakage current peak value and the correlation coefficient of the current signal are obtained by collecting the following methods:

[0073] Extracting the maximum value in the converted current signal as the leakage current peak value, and calculating the standard deviation normalized value of the leakage current peak value;

[0074] The correlation coefficient between the current signal when the leakage current is currently generated and the current signal when the leakage current was last generated is calculated.

[0075] During implementation, for the αth trigger, that is, the current signal obtained when the αth leakage current is generated is represented as i α,j , j=1,…,K1, K1 represents the signal length, that is, the number of sampling points.

[0076] The maximum value of the current signal when the leakage current is generated for the αth time is calculated as the leakage current peak value when the leakage current is generated for the αth time, expressed as I α .

[0077] Calculate the normalized value of the standard deviation of the leakage current peak

[0078] Among them, I α,r represents the rth leakage current peak value of the leakage current when the leakage current is generated for the αth time, EI represents the average of all leakage current peak values, and n1 represents the number of leakage current peak values ​​when the leakage current is generated for the αth time.

[0079] In the process of insulator development from corona to flashover, the leakage current can dynamically reflect the discharge condition on the surface of the insulating material. By calculating the correlation coefficient between two adjacent leakage currents, the time series change trend of the leakage current can be quantitatively identified and the discharge condition on the surface of the insulating material can be dynamically evaluated.

[0080] Specifically, the Pearson correlation coefficient of the current signal is calculated using the following formula:

[0081]

[0082] Among them, i α,j represents the leakage current value at the jth sampling moment when the leakage current is generated for the αth time, K1 represents the total number of sampling points, and Ei α It represents the mean value of leakage current at all sampling moments when leakage current is generated for the αth time, Ei α-1 It represents the mean value of the leakage current at all sampling moments when the leakage current is generated for the α-1th time.

[0083] The acoustic signal is caused by the collision of charges and the ionization process of the gas near the test electrode on the surface of the insulator, which causes the surrounding air to vibrate. The electrical signal and the acoustic signal are collected synchronously. When the αth trigger is triggered, the acoustic signal is collected through the acoustic signal collection unit.

[0084] Specifically, the characteristic data extracted from the converted acoustic signal include center frequency, spectrum amplitude and average power; the acoustic signal data is obtained in the following manner:

[0085] Performing a fast Fourier transform on the acoustic signal to obtain a frequency spectrum of the acoustic signal;

[0086] Extract the center frequency and amplitude corresponding to all the spectrum peaks in the spectrum; calculate the average power of the acoustic signal.

[0087] For example, the acoustic signal is represented by S α,j , j = 1,…,K2. S α,j It represents the signal of the jth sampling point of the acoustic signal collected by the αth trigger, and the number of sampling points is K2.

[0088] First, the sound signal is subjected to fast Fourier transform to obtain the spectrum of the sound signal.

[0089] Extract the center frequency F corresponding to all spectral peaks in the spectrum α,j and amplitude A α,j , as the basis for evaluating 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, j = 1,…,n2, n2 represents the number of frequencies in the spectrum.

[0090] The energy of the acoustic signal is different in different states. The energy of the acoustic signal changes with the applied voltage amplitude and the intensity of the discharge phenomenon during the test process. Therefore, the average power of the acoustic signal is calculated to quantitatively reflect the energy of the acoustic spectrum:

[0091]

[0092] Among them, K2 represents the total number of sampling points of the acoustic signal collected by the αth trigger, S α,j Represents the signal of the jth sampling point of the acoustic signal collected by the αth trigger.

[0093] When the insulator is in a state of no obvious discharge, there is no corona discharge and the captured image has almost no obvious change; when corona discharge occurs, the image will have obvious diffuse corona; when dry flash occurs, the flashover discharge image penetrates the conductive channel, the number of arc channels increases, and the shape becomes thicker; when wet flash occurs, the arc distribution bends under the influence of droplets. The images are different in different states. Therefore, by collecting surface images, the state of the insulator can be further accurately monitored, providing a basis for predicting flashover risks.

[0094] Specifically, the feature data extracted from the surface image include the surface image, discharge brightness and electrical damage area; the discharge brightness and electrical damage area are obtained in the following manner:

[0095] Calculating the grayscale standard deviation of the area in the surface image whose grayscale is greater than the first threshold to obtain the discharge brightness;

[0096] The electrical damage area is obtained by calculating the percentage of the area in the surface image whose grayscale is greater than the first threshold in the surface image.

[0097] The discharge brightness and electrical damage area of ​​the insulator are different under different states. In order to accurately detect the electrical damage area and then accurately predict the flashover risk, the surface image is converted into a grayscale image, and the grayscale standard deviation of the area in the surface image with a grayscale greater than the first threshold is calculated to obtain the discharge brightness, and the percentage of the area with a grayscale greater than the first threshold in the surface image is calculated to obtain the electrical damage area.

[0098] The multi-dimensional feature data when the insulator generates leakage current is input into the trained flashover risk prediction model to obtain the current flashover risk of the insulator. In implementation, for example, the features corresponding to the electrical signal, the features corresponding to the acoustic signal, and the features corresponding to the surface image obtained when the leakage current is generated for the αth time are input into the trained flashover risk prediction model, and the flashover risk assessment result of the insulator when the leakage current is generated for the αth time is obtained.

[0099] Specifically, the flashover risk prediction model is a multi-classification neural network model; the multi-classification neural network model includes:

[0100] A first feature extraction module, used for extracting a first feature according to the converted current signal and acoustic signal;

[0101] A second feature extraction module, used to extract a second feature from the surface image;

[0102] A feature fusion module, used for fusing the first feature and the second feature to obtain a fused feature;

[0103] A classification module, used for classifying the insulator status based on the fusion features;

[0104] The risk calculation module is used to calculate the insulator flashover risk based on the insulator state classification results.

[0105] Since the features corresponding to the current signal and the acoustic signal and some features of the surface image are numerical values, and the surface image is image data, it is necessary to extract features separately. During implementation, the first feature extraction module can adopt an existing machine learning model or a deep learning model, such as a feedforward neural network structure, and the second feature extraction module can adopt a convolutional network structure.

[0106] The feature fusion module fuses the extracted first feature and the second feature. When implemented, the fusion can be performed in a splicing manner.

[0107] The classification module classifies the insulator state according to the fused features, that is, predicts the probability of the insulator belonging to the state of no obvious discharge, corona and flashover. The risk calculation module calculates the flashover risk according to the probability of each state.

[0108] Specifically, the risk calculation module uses the following formula to calculate the flashover risk probability of the insulator:

[0109] Risk=ω1p1+ω2p2+ω3p3

[0110] Among them, p1 represents the probability that the insulator is in the flashover state predicted by the model, p2 represents the probability that the insulator is in the corona state predicted by the model, p3 represents the probability that the insulator is in the no obvious discharge state predicted by the model, ω1, ω2 and ω3 represent weights, where ω1>ω2>ω3.

[0111] During long-term service, insulators gradually change from a state of no obvious discharge to a corona state, and then develop into a flashover. Therefore, by judging the current state of the insulator, the probability of flashover can be predicted.

[0112] Specifically, the trained flashover risk prediction model is obtained in the following manner:

[0113] Collect multi-dimensional feature data of insulators in different states; use the multi-dimensional feature data as input data of training samples and use the states of insulators as labels to construct a training sample set; the different states include no obvious discharge, corona and flashover states;

[0114] Constructing a multi-classification neural network model; training the multi-classification neural network model based on the training sample set, and using the trained multi-classification neural network model as a trained flashover risk prediction model.

[0115] During implementation, multi-dimensional feature data of solid insulating materials in different states are collected, that is, the electrical signal, acoustic signal and surface image each time leakage occurs in the state of no obvious discharge, corona and flashover, and the corresponding multi-dimensional features are extracted to construct a training sample set. During implementation, feature data of different insulators can be collected to form big data, and big data storage and management can be performed to provide a basis for subsequent training of accurate and highly generalized models.

[0116] Then, a multi-classification neural network is trained based on the training sample set to obtain a trained flashover risk prediction model.

[0117] It should be noted that when training a multi-classification neural network model, the network loss is calculated based on the output results of the classification module.

[0118] Since the number of samples of the flashover state is small compared to other states in the collected data, there is a problem of sample imbalance. If the model is trained directly, overfitting will occur, affecting the accuracy of the model. Therefore, after constructing the training sample set, before training the multi-classification neural network model based on the training sample set, it also includes:

[0119] Extracting samples labeled as flashover from the training sample set, and splicing the multi-dimensional features of the extracted samples and the corresponding surface images to form source data;

[0120] The source data is input into the trained generative adversarial network model to generate new flashover state samples, and the generated new flashover state samples are added to the training sample set to expand the training sample set.

[0121] During implementation, firstly, samples of flashover states are generated based on the trained generative adversarial network model, and then the samples are expanded to improve the accuracy of the subsequent classification model.

[0122] Specifically, the constructed generative adversarial network model includes a first generator, a second generator, a first discriminator, and a second discriminator;

[0123] The first generator is used to generate a first sample according to input data and perform arc segmentation on the surface image in the first sample; the first discriminator is used to perform authenticity discrimination on the result generated by the first generator;

[0124] The second generator is used to generate a second sample according to the input first sample and perform arc segmentation on the surface image in the second sample; the second discriminator is used to perform authenticity discrimination on the result generated by the second generator.

[0125] By combining the arc segmentation task with the sample generation task, the two tasks assist each other, thereby improving the efficiency of sample generation, and the generated samples have a high similarity with the real samples, thus providing a data basis for the subsequent training of the classification model.

[0126] During implementation, the first generator and the second generator have the same structure. The first generator and the second generator can adopt a symmetrical structure model, that is, a model including an encoder and a decoder, such as using Unet. After the decoder extracts the decoding features, a sample generation module and an arc segmentation module are added to realize sample generation and arc segmentation.

[0127] Specifically, the first generator and the second generator have the same structure; the first generator includes:

[0128] An encoder, for extracting shallow features of input source data using a multi-layer downsampling structure;

[0129] Decoder; used to decode the input features using a multi-layer upsampling structure and output decoded features; a jump connection is connected between the encoder and the decoder, which is used to transmit the shallow features of each layer downsampled by the encoder to the upsampling layer corresponding to the encoder as the input of the upsampling layer corresponding to the encoder;

[0130] A generation module, used to generate a first sample according to the decoding features output by the last upsampling layer of the decoder;

[0131] The segmentation module is used to generate a contour segmentation image of the surface image according to the decoded features.

[0132] During implementation, a preprocessing layer may be added before the encoder to preprocess the input source data, extract initial features, and send the initial features to the encoder.

[0133] The encoder uses a multi-layer downsampling structure to extract shallow features of the input source data. In implementation, the encoder includes multiple convolution modules and downsampling layers. A convolution module is connected to a downsampling layer to gradually extract shallow features and reduce the size of the feature map. Each convolution module includes two instance normalization layers, two convolution layers with a stride of 1, and two Leaky ReLU activation functions. Each downsampling layer contains a 2D convolution layer.

[0134] The decoder includes multiple convolution modules and upsampling layers. One convolution module is connected to an upsampling layer, which is used to decode the input features and output decoded features using a multi-layer upsampling structure. The convolution module of the decoder has the same structure as the convolution module of the encoder, and uses deconvolution as the upsampling layer. The decoder decodes the features step by step and gradually increases the size of the feature map.

[0135] In order to improve the quality of the generated image, a skip connection is connected between the encoder and the decoder to transmit the shallow features of each layer downsampled by the encoder to the upsampling layer corresponding to the encoder as the input of the upsampling layer corresponding to the encoder. The skip connection fuses the features of the corresponding positions of the encoder and the decoder, thereby improving the quality of generation and segmentation.

[0136] After the decoder extracts the decoding features, a sample generation module and an arc segmentation module are added to realize sample generation and arc segmentation.

[0137] In implementation, the sample generation module includes a 2D convolution layer and a LeakyReLU activation layer, and the arc segmentation module structure is the same as the sample generation module.

[0138] During implementation, both the first discriminator and the second discriminator may use PatchGAN. During implementation, the surface image of the sample in the flashover state is annotated with arc contours, and the contour of the arc is annotated, that is, the entire arc part is annotated as 1, and the other parts are annotated as 0.

[0139] During implementation, the input of the first generator is the sample data of the flashover state. Since the sample input data includes the surface image and voltage amplitude, leakage current peak value, leakage current time series correlation coefficient, leakage current peak standard deviation normalized value, the center frequency and amplitude of the main spectral components in the acoustic signal spectrum, the average power of the acoustic signal, the discharge brightness and the proportion of the electrical damage area. Therefore, other features except the surface image are converted into a one-dimensional vector, and the converted one-dimensional vector is converted into the form of LxW and spliced ​​with the surface image as the input of the first generator, where W is the width of the surface image.

[0140] During implementation, other features and surface images are spliced ​​together as the entire sample to generate a new sample, and the relationship between other features and surface images is mined to make the generated sample more realistic and more similar to the real sample.

[0141] When implemented, the loss of the generative adversarial network model is calculated using the following formula:

[0142] Loss=θ1(‖Xp-Yp‖1+β‖Xp·S-Yp·S‖1)+θ2L dice

[0143] +θ3L feat (Xp,X p ′ )+θ4L D

[0144] Wherein, Xp represents the surface image in the input data of the first generator, Yp represents the surface image in the first sample generated by the first generator, S represents the arc segmentation map corresponding to the surface image in the input data of the first generator, and L dice Indicates arc segmentation loss, L feat (·,·) represents the feature reconstruction loss, X p ′ represents the surface image in the second sample generated by the second generator, L D represents the discriminator loss, θ1, θ2, θ3 and θ4 represent weight coefficients, β represents parameters, and ‖·‖1 represents the norm of the matrix.

[0145] When implementing, the arc splitting loss L is calculated using the following formula: dice :

[0146]

[0147] Among them, S Yp represents the arc segmentation image obtained by the first generator performing arc segmentation on the surface image in the first sample, It represents the arc segmentation image obtained by the second generator performing arc segmentation on the surface image in the second sample.

[0148] During implementation, the feature reconstruction loss L is calculated using the following formula: feat (Xp,X p ′ ):

[0149]

[0150] Among them, N L represents the number of feature extraction layers of the pre-trained neural network, represents the feature map of the j-th layer output of the pre-trained neural network, C j H represents the number of channels of the feature map output by the jth layer of the pre-trained neural network. j Represents the height of the feature map output by the jth layer of the pre-trained neural network, W j represents the width of the feature map output by the jth layer of the pre-trained neural network, and ‖·‖2 represents the binary norm of the matrix.

[0151] When implemented, the feature reconstruction loss is calculated by dividing the surface image Xp in the input data of the first generator by the surface image Xp in the second sample generated by the second generator. p ′ A 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 a high-dimensional space. The feature reconstruction loss is used to constrain the high-dimensional space to improve the generalization of the model.

[0152] When implemented, the discriminator loss L is calculated using the following formula: D :

[0153] L D =E[log(1-D1(Y))]+E[log(1-D2(X ′ ))]

[0154] Where E[·] represents the expectation, D1(Y) represents the discrimination result of the first discriminator on the first sample generated by the first generator, and D2(X ′ ) represents the discrimination result of the second discriminator on the second sample generated by the second generator.

[0155] By adding arc segmentation constraints and feature reconstruction loss constraints, the model can more accurately capture the arc characteristics of the flashover surface image, making the generated samples more realistic and reliable.

[0156] When the training loss reaches the preset accuracy requirement or the number of training times is reached, the training of the generative adversarial network model is stopped to obtain a trained generative adversarial network model. The first sample obtained by inputting the real flashover sample into the trained generative adversarial network model is added as a generated sample to the training sample set to expand the training sample set.

[0157] Based on the expanded sample set, a multi-classification neural network model is trained to obtain a trained flashover risk prediction model.

[0158] Based on the trained flashover risk prediction model, the multi-dimensional features of the insulator when generating leakage current are obtained and input into the flashover risk prediction model to obtain the flashover risk prediction result of the insulator. The insulator is monitored in a timely manner to improve the accuracy and efficiency of monitoring.

[0159] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0160] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. An insulator flashover risk warning system, characterized in that: include: A current collection unit is used to collect current signals when leakage current is generated on the surface of the insulator; When the current signal exceeds a preset threshold, a trigger signal is sent to the sound signal collection unit and the shooting unit, and the collected current signal is sent to the processing unit; an acoustic signal collecting unit, configured to collect the acoustic signal after receiving the trigger signal, and send the collected acoustic signal to the processing unit; A shooting unit, used for collecting an image of the insulator surface after receiving the trigger signal, and sending the captured surface image to a processing unit; a processing unit, configured to perform analog-to-digital conversion on the current signal and the acoustic signal and to send the converted current signal, acoustic signal and surface image to a monitoring center; A monitoring center is used to extract multi-dimensional feature data according to the converted current signal, acoustic signal and surface image; predict the current flashover risk of the insulator based on the trained flashover risk prediction model; and send an alarm message to the alarm unit when the flashover risk is greater than a preset risk threshold; An alarm unit, used for generating an alarm according to the alarm information; The multi-dimensional feature data is electrical-acoustic-optical multi-dimensional features; The flashover risk prediction model is a multi-classification neural network model; The multi-classification neural network model includes: A first feature extraction module, used for extracting a first feature according to the converted current signal and acoustic signal; A second feature extraction module, used to extract a second feature from the surface image; A feature fusion module, used for fusing the first feature and the second feature to obtain a fused feature; A classification module, used for classifying the insulator status based on the fusion features; The risk calculation module is used to calculate the insulator flashover risk based on the insulator state classification results.

2. The insulator flashover risk warning system according to claim 1, characterized in that: The current acquisition unit includes a current sensor, an amplifier and a trigger circuit; The current sensor adopts a non-contact through-core leakage current sensor, which is used to collect the leakage current signal of the insulator; The amplifier is used to amplify the collected leakage current signal; The trigger circuit is used to compare the amplified leakage current signal with a preset threshold value, and output a high level signal as a trigger signal when the leakage current signal is greater than the preset threshold value.

3. The insulator flashover risk warning system according to claim 1, characterized in that: The trained flashover risk prediction model is obtained in the following way: Collect multi-dimensional characteristic data of insulators under different states; The multi-dimensional feature data is used as the input data of the training sample, and the state of the insulator is used as the label to construct a training sample set; the different states include no obvious discharge, corona and flashover states; Build a multi-classification neural network model; The multi-classification neural network model is trained based on the training sample set, and the trained multi-classification neural network model is used as a trained flashover risk prediction model.

4. The insulator flashover risk warning system according to claim 3, characterized in that: After constructing the training sample set and before training the multi-classification neural network model based on the training sample set, the method further includes: Extracting samples labeled as flashover from the training sample set, and splicing the multi-dimensional features of the extracted samples and the corresponding surface images to form source data; The source data is input into the trained generative adversarial network model to generate new flashover state samples, and the generated new flashover state samples are added to the training sample set to expand the training sample set.

5. The insulator flashover risk warning system according to claim 4, 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 according to input data and perform arc segmentation on the surface image in the first sample; the first discriminator is used to perform authenticity discrimination on the result generated by the first generator; The second generator is used to generate a second sample according to the input first sample and perform arc segmentation on the surface image in the second sample; the second discriminator is used to perform authenticity discrimination on the result generated by the second generator.

6. The insulator flashover risk warning system according to claim 5, characterized in that: The first generator and the second generator have the same structure; the first generator comprises: An encoder, for extracting shallow features of input source data using a multi-layer downsampling structure; A decoder is used to decode the input features and output decoded features using a multi-layer upsampling structure; a jump connection is connected between the encoder and the decoder, which is used to transmit the shallow features of each layer downsampled by the encoder to the upsampling layer corresponding to the encoder as the input of the upsampling layer corresponding to the encoder; A generation module, used to generate a first sample according to the decoding features output by the last upsampling layer of the decoder; The segmentation module is used to generate a contour segmentation image of the surface image according to the decoded features.

7. The insulator flashover risk warning system according to claim 6, characterized in that: The loss of the generative adversarial network model is calculated using the following formula: Loss=θ1(‖Xp-Yp‖1+β‖Xp·S-Yp·S‖1)+θ2L dice +θ3L feat (Xp,X′ p )+θ4L D Wherein, Xp represents the surface image in the input data of the first generator, Yp represents the surface image in the first sample generated by the first generator, S represents the arc segmentation map corresponding to the surface image in the input data of the first generator, and L dice Indicates arc segmentation loss, L feat (·,·) represents the feature reconstruction loss, X′ p represents the surface image in the second sample generated by the second generator, L D represents the discriminator loss, θ1, θ2, θ3 and θ4 represent weight coefficients, β represents parameters, and ‖·‖1 represents the norm of the matrix.

8. The insulator flashover risk warning system according to claim 7, characterized in that: The arc splitting loss L is calculated using the following formula: dice : Among them, S Yp represents the arc segmentation image obtained by the first generator performing arc segmentation on the surface image in the first sample, It represents the arc segmentation image obtained by the second generator performing arc segmentation on the surface image in the second sample.

9. The insulator flashover risk warning system according to claim 1, characterized in that: The risk calculation module calculates the flashover risk probability of the insulator based on the insulator status classification results using the following formula: Risk=ω1p1+ω2p2+ω3p3 Among them, p1 represents the probability that the insulator is in the flashover state predicted by the model, p2 represents the probability that the insulator is in the corona state predicted by the model, p3 represents the probability that the insulator is in the no obvious discharge state predicted by the model, ω1, ω2 and ω3 represent weights, where ω1>ω2>ω3.

Images