GIS insulator partial discharge charge on-line detection method and system

Abstract

The application discloses a GIS insulator partial discharge charge on-line detection method and system, and relates to the technical field of charge detection.The method comprises the following steps: according to a GIS simulation model, simulating the electric field distribution of charge generation in the partial discharge process, determining the performance index of the required electric field sensing probe, and preparing the electric field sensing probe; according to an epoxy resin model, obtaining the charge density, the charge distribution range and the electric field signal collected by the electric field sensing probe under different voltage pulses, and constructing a test data set; according to the GIS simulation model, obtaining the electric field signal under different charge densities and different charge distribution radii, and constructing a training data set; training a neural network model according to the training data set, and correcting according to the test data set; and collecting the partial discharge electric field signal by using the electric field sensing probe array arranged on the outer wall of the GIS insulator, and obtaining the charge density and the charge distribution range on the surface of the insulator by using the neural network. Non-invasive and non-contact on-line measurement is realized.

Description

Technical Field

[0001] This invention relates to the field of charge detection technology, and in particular to an online method and system for detecting partial discharge charge in GIS insulators. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Gas-insulated switchgear (GIS) is a crucial core component of power systems, and its insulation condition directly impacts the stability of the power grid operation. Partial discharge is an early sign and important manifestation of GIS insulation degradation, and a significant cause of insulation failure.

[0004] Studies have shown that the dynamic behavior of surface charges generated during partial discharge and their distortion effect on the electric field may be important causes of insulation breakdown. Therefore, online detection of charges during partial discharge can clarify the dynamic evolution of charge density and distribution, which is beneficial for revealing the correlation mechanism between charge characteristics and discharge development, clarifying the mapping relationship between charge characteristics and insulation state, and is of great significance for online monitoring of GIS insulation state.

[0005] Existing surface charge measurement techniques mainly include capacitive electrostatic probe method and pulse acoustic electrophysiological method, but none of these methods are suitable for online detection of surface charge on a large number of real-life GIS insulators in operation.

[0006] The electrostatic probe method requires the capacitive probe to be in close contact with the insulator surface for scanning. Since the probe is made of metal, inside operating GIS equipment, the high-voltage conductors and insulators are in a completely enclosed, insulated environment, making it neither permissible nor possible to install any moving parts (especially metal parts) for scanning measurements. The pulsed acoustic-electric method inverts the charge distribution by detecting the acoustic signal generated by surface charges under a high-voltage pulse. This requires applying a pulse voltage to the device under test, but operating GIS equipment does not allow for the application of additional pulse voltages. Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes an online detection method and system for partial discharge charge in GIS insulators, achieving non-invasive and non-contact online measurement without damaging the original GIS structure.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides an online detection method for partial discharge charge in GIS insulators, comprising: Based on the constructed GIS simulation model, the electric field distribution generated by the charge during partial discharge is simulated to determine the performance indicators of the required electric field sensing probe, and the electric field sensing probe is prepared accordingly. Based on the pre-built epoxy resin model, the charge density, charge distribution range and electric field signal collected by the electric field sensing probe under different voltage pulses are obtained to construct a test dataset. Based on the GIS simulation model, electric field signals under different charge densities and different charge distribution radii are obtained to construct a training dataset. The neural network model is trained based on the training dataset and then corrected based on the test dataset to obtain the charge inversion neural network model. The partial discharge field signal is collected by an array of electric field sensing probes installed on the outer wall of the GIS insulator. The charge inversion neural network is used to obtain the charge density and charge distribution range generated on the surface of the insulator during the discharge process.

[0009] As an alternative implementation method, a GIS simulation model is established based on the geometric structure and material parameters of a real GIS. By setting the discharge charge quantity and charge distribution pattern on the surface of the insulator in the GIS simulation model, the electric field signal at the location of the electric field sensing probe on the outer wall of the insulator is collected. In this way, the performance indicators of the electric field sensing probe required for discharge electric field measurement, such as sensitivity, electric field measurement range, measurement bandwidth and time resolution, are determined.

[0010] As an optional implementation method, the determination of performance indicators specifically includes: The peak value of the electric field signal is used as the upper limit of the electric field measurement range of the electric field sensing probe; With the smallest measurable electric field difference As the lower limit of sensitivity of the electric field sensing probe; in, It is the peak electric field when the surface charge density is at its minimum and the distribution range is the smallest. This is the power frequency background electric field value; The time interval between the rise of the electric field signal pulse from 10% peak value to 90% peak value is taken as the pulse rise time, and 1 / 100 of the pulse rise time is taken as the lower limit of time resolution. The electric field signal is subjected to a fast Fourier transform to obtain the spectral distribution. The frequency value at which the accumulated energy reaches 90% of the total energy is used as the upper limit of the measurement bandwidth.

[0011] As an alternative implementation, the constructed neural network model includes an input layer, a hidden layer, and an output branch layer; The input layer receives the electric field signal. After standardizing and preprocessing the electric field signal, it uses multi-channel one-dimensional convolution of different sizes to extract the partial discharge pulse features along the time axis within a sliding window. The hidden layer extracts features from the input layer, extracts the spatial distribution pattern of electric field distortion through stacked residual convolutional blocks, and then inputs them into the output branch layer after attention weighting by the spatial attention module. The output branch layer includes two parallel fully connected branches. The first fully connected branch outputs the predicted maximum surface charge density and distribution gradient, and the second fully connected branch outputs the diffusion radius of the discharge charge on the insulator surface.

[0012] As an alternative implementation method, the process of making corrections based on the test dataset includes: For each test sample Calculate the relative error of charge density respectively. The relative error of the charge distribution range : ; ; in, and Let be the model prediction value for the i-th test sample. and The actual preset value for the i-th test sample; The deviation distribution is analyzed based on different charge density ranges, the spatial position of the electric field sensing probe, and the deviation attributes. The causes of the deviation are determined, including physical parameter mismatch, sensor and hardware noise, and assembly process errors. To address the discrepancies caused by physical parameter mismatch and assembly process errors, the weights of the first half of the input layer and hidden layer in the neural network model are frozen, while only the second half of the hidden layer and the output branch layer are unfrozen. Iterative updates are then performed on the test dataset using a set minimum learning rate. To address the discrepancies between sensor and hardware noise, white noise following a Gaussian distribution or background noise segments extracted from the actual test environment are superimposed on the original training dataset. At the same time, spatial displacement perturbation is introduced to reconstruct the enhanced training dataset and perform secondary training. If the prediction accuracy of charge density meets the standard, but the prediction accuracy of charge distribution range does not, then adjust the weight coefficients in the total loss function of the model, increasing the penalty for the weight with the larger deviation between the two, until the prediction accuracy of both meets the requirements.

[0013] As an alternative implementation method, in the GIS simulation model, surface charge parameters with different charge densities and different charge distribution radii are set. For each set of charge parameters, electric field signals are collected at the location of the sensing probe array on the outer wall of the insulator, and a training dataset is constructed with the electric field signals as input and the corresponding charge parameters as supervision labels.

[0014] As an alternative implementation, charges with different charge densities and charge distribution ranges are pre-placed in selected areas on the outer wall surface of the insulator of the epoxy resin model to simulate the charge accumulation state after partial discharge. The electric field signal of the partial discharge process is synchronously collected by a sensor probe array to construct a test dataset. The electric field sensor probes are placed on the outer wall of the insulator of the epoxy resin model, and one electric field sensor probe is arranged at a set interval to form a sensor probe array.

[0015] As an alternative implementation, a needle electrode is set in the normal direction on the outer wall surface of the epoxy resin model insulator. Different levels of voltage pulses are applied using the needle electrode, and the surface charge density is measured using a probe. The electric field signal of the discharge process is collected synchronously using a sensor probe array. After the experiment, the surface charge distribution range is determined based on the charge density of all measuring points.

[0016] Secondly, the present invention provides an online detection system for partial discharge charge of GIS insulators, comprising: The simulation module is configured to determine the performance indicators of the required electric field sensing probe by simulating the electric field distribution generated by the charge during partial discharge based on the constructed GIS simulation model, thereby preparing the electric field sensing probe. The test dataset construction module is configured to acquire the charge density, charge distribution range and electric field signal collected by the electric field sensing probe under different voltage pulses based on a pre-built epoxy resin model, thereby constructing the test dataset. The training dataset construction module is configured to acquire electric field signals under different charge densities and different charge distribution radii based on the GIS simulation model, thereby constructing the training dataset. The model training module is configured to train the constructed neural network model based on the training dataset and correct it based on the test dataset to obtain the charge inversion neural network model. The detection module is configured to collect partial discharge field signals using an array of electric field sensing probes installed on the outer wall of the GIS insulator. The partial discharge field signals are then processed using a charge inversion neural network to obtain the charge density and charge distribution range generated on the surface of the insulator during the discharge process.

[0017] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0018] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0019] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes an online method and system for detecting partial discharge charge in GIS insulators. Through simulation analysis, the required performance indicators of the electric field sensing probe for measuring the discharge electric field are clarified, including sensitivity, bandwidth, and dynamic range. Based on this, a high-performance electric field sensing probe is fabricated. Simulation and measured data are used to construct training and testing datasets, respectively, to train and correct the neural network, resulting in a charge inversion neural network. Multiple electric field sensing probes are arranged on the outer wall of a real GIS insulator to form a sensor array, acquiring the electric field signal during the discharge process in real time. The charge inversion neural network is then used to obtain the charge density and charge distribution range generated on the insulator surface during the discharge. By arranging the electric field sensing probes on the outer wall of the GIS insulator, non-invasive and non-contact online measurement is achieved without damaging the original GIS structure. Furthermore, since the sensing probes do not penetrate the cavity, they do not affect the original electric field inside the cavity, ensuring accurate and reliable measurement results and not affecting the development of partial discharge within the cavity.

[0021] This invention constructs a training dataset by combining simulation and experiment. It generates a large number of samples covering different charge densities and distribution ranges through simulation, solving the problem of obtaining a large number of measured samples. At the same time, it uses measured data to test and correct the neural network, enabling the model to adapt to the complex electromagnetic environment and noise interference under real working conditions, thus enhancing the engineering practicality of the method.

[0022] This invention clarifies the performance indicators of the sensor required for measuring the discharge electric field through simulation, such as sensitivity, bandwidth, and dynamic range. The optical electric field sensor prepared has advantages such as high sensitivity, wide bandwidth, low system noise, and large dynamic range. It can realize the electric field measurement of the entire process from the power frequency background field and the start of partial discharge to the final insulation breakdown, providing a high-precision and high-reliability raw dataset for charge inversion.

[0023] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0025] Figure 1 This is a flowchart of the online detection method for partial discharge charge of GIS insulators provided in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the GIS simulation model provided in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the electric field sensing probe arrangement provided in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the connection of the electric field sensing probe provided in Embodiment 1 of the present invention; Figure 5 This is an electric field signal diagram of a GIS insulator during partial discharge, provided in Embodiment 1 of the present invention. Figure 6 This is a schematic diagram of the charge inversion results provided in Embodiment 1 of the present invention; Among them, 1. cover plate; 2. high voltage conductor; 3. insulator; 4. terminal controller; 5. photodetector; 6. laser; 7. electric field sensing probe; 8. outer wall of insulator. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0028] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. Furthermore, it should be understood that the terms “comprising” and “including”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0029] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0030] Example 1 like Figure 1 As shown, this embodiment provides an online detection method for partial discharge charge in GIS insulators, including: S1: Based on the constructed GIS simulation model, the electric field distribution generated by the charge during partial discharge is simulated to determine the performance indicators of the required electric field sensing probe, and the electric field sensing probe is prepared accordingly. S2: Based on the pre-built epoxy resin model, obtain the charge density, charge distribution range and electric field signal collected by the electric field sensing probe under different voltage pulses to construct a test dataset; S3: Based on the GIS simulation model, obtain electric field signals under different charge densities and different charge distribution radii to construct a training dataset; S4: Train the constructed neural network model based on the training dataset, and correct it based on the test dataset to obtain the charge inversion neural network model; S5: The partial discharge field signal is collected by an array of electric field sensing probes set on the outer wall of the GIS insulator. The partial discharge field signal is then processed by a charge inversion neural network to obtain the charge density and charge distribution range generated on the surface of the insulator during the discharge process.

[0031] In this embodiment, in step S1, based on the constructed GIS simulation model, the electric field distribution generated by the charge during partial discharge is simulated to determine the performance indicators of the required electric field sensing probe, thereby preparing the electric field sensing probe.

[0032] Specifically, it includes: S11: Based on the geometric structure and material parameters of a real GIS, establish a high-precision, fine-grid, multi-physics GIS simulation model to simulate the electric field distribution generated by charge during partial discharge. For example... Figure 2 As shown, it includes a cover plate 1, a high-voltage conductor 2, an insulator 3, and other parts.

[0033] S12: By setting the discharge charge quantity and charge distribution pattern on the surface of the insulator in the GIS simulation model, the electric field signal at the location of the electric field sensing probe on the outer wall of the insulator is collected, and the performance indicators such as the sensitivity, electric field measurement range, measurement bandwidth and time resolution of the electric field sensing probe required for discharge electric field measurement are determined accordingly.

[0034] In this embodiment, the specific implementation process of step S12 is as follows: The electric field signal collected at the location of the sensor probe array (multiple measuring points along the circumference of the outer wall of the insulator) Time-domain and frequency-domain analyses were performed, and relevant performance indicators were determined by extracting the following key features.

[0035] (1) Calculate the peak electric field, i.e. the maximum electric field strength, to evaluate the maximum electric field strength that the electric field sensing probe needs to withstand, and use it as the upper limit of the electric field measurement range of the electric field sensing probe.

[0036]

[0037] in, This represents the peak value of the electric field.

[0038] (2) Calculate the minimum measurable electric field difference This is used to evaluate the sensitivity of the probe, and serves as the lower limit of the sensitivity of the electric field sensing probe.

[0039]

[0040] in, It is the peak electric field when the surface charge density is at its minimum and the distribution range is the smallest. It is the power frequency background electric field value.

[0041] It is the smallest effective electric field signal that the electric field sensing probe can resolve, corresponding to the lower limit of sensitivity. The probe resolution must be better than... Only then can the weak electric field of the target be detected from the background.

[0042] (3) Extract the time interval from the 10% peak value to the 90% peak value of the electric field signal pulse, and record it as the pulse rise time. This is used to evaluate the time resolution requirement of the electric field signal. The lower limit of the time resolution of the electric field sensing system is 1 / 100 of the pulse rise time. That is, the time resolution of the electric field sensing system should not be less than 1 / 100 of the pulse rise time in order to meet the requirements of complete acquisition and waveform restoration of the electric field pulse signal.

[0043] (4) For electric field signals Perform a Fast Fourier Transform:

[0044] in, It is the spectral distribution of the electric field signal. For frequency.

[0045] The frequency at which the accumulated energy reaches 90% of the total energy is calculated and used as the upper limit of the measurement bandwidth of the electric field sensing probe.

[0046] S13: Prepare electric field sensing probes of different sizes and shapes. Based on the performance indicators of the required electric field sensing probes, calibrate the performance indicators of probes with different parameters in the laboratory. Select electric field sensing probes that meet the performance indicators in step S12 for integration of the electric field sensing system. Equip a high sampling rate oscilloscope based on the measurement bandwidth requirements proposed in step S12.

[0047] In this embodiment, in step S2, the charge density, charge distribution range and electric field signal collected by the electric field sensing probe under different voltage pulses are obtained according to the pre-constructed epoxy resin model, so as to construct a test dataset.

[0048] Specifically, it includes: S21: Prepare an epoxy resin model of the same size as the real GIS. Attach the electric field sensing probes 7 prepared in step S1 to the outer wall 8 of the insulator on the epoxy resin model using silicone. Attach one electric field sensing probe 7 every 50 mm to form a sensing probe array. Figure 3 As shown.

[0049] Laser 6 and electric field sensing probe 7 are connected via polarization-maintaining fiber; electric field sensing probe 7 and photodetector 5 are connected via single-mode fiber; photodetector 5 is electrically connected to terminal controller 4. Figure 4 As shown.

[0050] The electric field sensing probe 7 is composed of a collimator, a prism, a quarter glass slide, an electro-optic crystal, a prism, and a collimator bonded together in sequence.

[0051] S22: By pre-setting charges with different charge densities and charge distribution ranges in selected areas on the outer wall surface of the insulator of the epoxy resin model, the charge accumulation state after partial discharge is simulated, and the electric field signal of the partial discharge process is collected synchronously by a sensor probe array to construct a test dataset.

[0052] Specifically: (1) A needle electrode was placed in the normal direction on the outer wall surface of the epoxy resin model insulator, and a controllable voltage pulse was applied using the needle electrode. A movable electrostatic probe was used to perform grid-like scanning measurements along the insulator surface: the probe was placed perpendicular to the surface, with a fixed lift-off distance of 1 mm, and the surface charge density at each location was measured point by point, and the charge density value at each measurement point was recorded. At the same time, an array of sensor probes was used to synchronously acquire the electric field signal of the discharge process. After the experiment, a surface charge density distribution map was generated based on the charge density data of all measurement points, thereby determining the charge distribution range.

[0053] (2) Apply voltage pulses of different levels, starting from 10kV and increasing to 100kV in 1kV increments. Repeat step (1) to obtain multiple sets of data and construct a test dataset containing charge density, charge distribution range and discharge electric field signal.

[0054] In this embodiment, in step S3, electric field signals under different charge densities and different charge distribution radii are obtained according to the GIS simulation model, thereby constructing a training dataset.

[0055] Specifically, it includes: S31: In the GIS simulation model, surface charge parameters with different charge densities and different charge distribution radii are set, and a large number of samples are generated using a parametric scanning method.

[0056] S32: For each set of charge parameters, calculate the electric field signal at the location of the sensing probe array on the outer wall of the insulator, and construct a training dataset with the electric field signal as input and the corresponding charge parameters as supervision labels.

[0057] In this embodiment, in step S4, a neural network model is trained and constructed based on the training dataset, and then corrected based on the test dataset to obtain a charge inversion neural network model.

[0058] Specifically, it includes: S41: Using the training dataset, a deep neural network is trained to establish a charge inversion neural network model that transforms the electric field signal on the outer wall into the charge density and charge distribution range on the insulator surface.

[0059] The charge inversion neural network model based on electric field characteristics adopts a multi-task learning architecture, consisting of an input layer, a hidden layer, and an output branch layer in a progressive manner. This architecture aims to decouple and accurately regress the charge density and distribution range on the insulator surface from complex high-dimensional spatiotemporal electric field signals.

[0060] Specifically, it includes: (1) Input layer (data preprocessing and multi-scale shallow feature mapping).

[0061] The core task of the input layer is to receive the raw physical signals and perform preliminary cleaning, alignment, and multi-scale feature extraction to provide high-quality feature representations for subsequent deep networks.

[0062] (1-1) Signal reception and matrix construction.

[0063] The input layer receives multi-channel time-domain electric field signals synchronously acquired from the GIS insulator outer wall sensing probe array (50mm interval). The time-series data of each physical node are synchronously truncated and aligned to form a two-dimensional spatiotemporal feature matrix (number of spatial nodes). (Number of time sampling points), this matrix directly reflects the transient evolution of partial discharge in space and time.

[0064] (1-2) Data processing and multi-channel one-dimensional convolution calculation.

[0065] Standardization: The input electric field signal is Z-score standardized. This step effectively eliminates the huge differences in signal amplitude magnitude under different partial discharge intensities, making the model more stable during gradient descent and forcing the model to focus on the relative distortion rate and high-frequency phase characteristics of the electric field waveform, rather than a single absolute amplitude.

[0066] Initial feature extraction (multi-scale 1D-CNN configuration): Multi-scale multi-channel 1D-CNN is employed, utilizing a sliding window calculation along the time axis using 1D convolution kernels. To accommodate the bandwidth variations and transient changes in the electric field signal at different discharge stages, the input layer incorporates three parallel convolution branches: Branch 1: The kernel size is set to 3, the number of kernels (channels) is 64, and the stride is set to 1. This is used to accurately capture the steep pulse leading edge at the moment of discharge.

[0067] Branch 2: The kernel size is set to 5, the number of kernels is 64, and the stride is set to 1. This is used to extract the main high-frequency oscillation features of the electric field waveform.

[0068] Branch 3: The kernel size is set to 9, the number of kernels is 64, and the stride is set to 2. This is used to sense the envelope and attenuation trend of the signal within a slightly longer time window.

[0069] The feature maps output from the three branches are concatenated along the channel dimension to produce a shallow feature map with 192 channels.

[0070] (2) Hidden layer (deep spatiotemporal feature extraction and anti-overfitting).

[0071] The hidden layer employs a hybrid topology combining a deep residual network (ResNet) and a spatial attention mechanism. This layer is the core of the model's feature abstraction, enabling the extraction of highly nonlinear spatial distribution patterns of electric field distortion without losing information.

[0072] (2-1) Core topology: Residual Blocks.

[0073] To avoid gradient vanishing and feature degradation while increasing the number of network layers, four residual convolutional blocks are stacked in the hidden layers. Each residual block contains two one-dimensional convolutional layers (each with a kernel size of 3). As the network depth increases, the number of convolutional kernels (feature channels) in the four residual blocks increases in a stepwise manner, configured as 128, 256, 512, and 1024 respectively. Skip connections are introduced within each residual block to connect when the number of input and output channels is inconsistent. Dimension matching is performed on the convolutional layers.

[0074] (2-2) Spatial Attention Mechanism.

[0075] After feature extraction, a spatial attention module is connected to assign greater adaptive weights to sensor probe data that are closer to the center of the power source.

[0076] Specifically as follows: The input feature map (size is) ,in For the number of channels, Global max pooling and global average pooling are performed along the channel dimension (for spatial length) to generate two... One-dimensional feature vectors. These two vectors are concatenated along the channel dimension. The feature matrix is ​​then passed through a standard one-dimensional convolutional layer with a kernel size of 7 and an output channel of 1 to compress the number of channels to 1. Finally, it is mapped to a Sigmoid activation function. The spatial weight vector of the interval is then multiplied element-wise with the original input feature map to enhance the features of the key spatial node data.

[0077] (2-3) Overfitting prevention and generalization improvement mechanism.

[0078] Batch Normalization and Activation Function: Strict batch normalization (BN) is introduced after each convolutional layer and before the activation function to reconstruct the data distribution and smooth the loss pattern, accelerating network convergence. The activation function is uniformly Leaky-ReLU, effectively avoiding neuron "death" and feature truncation issues.

[0079] Dropout and L2 Regularization: After flattening the features using Global Average Pooling (GAP) at the end of the hidden layer, a Dropout layer with a dropout rate of 0.4 is connected. Simultaneously, an L2 regularization penalty term (Weight Decay) is introduced into the weight updates of each convolutional and fully connected layer. The L2 regularization coefficient (… ) Specifically set as These mechanisms force the model to learn general, robust features across nodes, significantly improving its generalization and perception capabilities in the complex electromagnetic environments of real GIS.

[0080] (3) Output branch layer (multi-objective decoupling and regression calculation).

[0081] The shared high-dimensional feature vector (1024 dimensions) extracted and flattened by the hidden layer is then split into two parallel fully connected network branches to achieve decoupled mapping of the target physical quantity.

[0082] (3-1) The number of neurons and layer configuration of network branches.

[0083] The first branch (charge density regression) is responsible for predicting the maximum surface charge density and its distribution gradient. This branch consists of three fully connected (FC) layers with decreasing numbers of neurons: 512, 128, and 1. The hidden FC layers use Leaky-ReLU activation, and the last layer strictly uses the ReLU activation function to ensure that the predicted charge density value is non-negative.

[0084] The second branch (charge distribution range regression) is responsible for outputting the diffusion radius or equivalent coverage area of ​​the discharge charge on the insulator surface. Since this task's relative density prediction is more sensitive to spatial geometry, it also consists of three independent fully connected layers with a neuron configuration of 256, 64, and 1. The final layer also uses the ReLU activation function.

[0085] (3-2) Design of joint loss function.

[0086] To balance the regression tasks involving two physical quantities with different dimensions and to avoid a decrease in the accuracy of distribution range prediction due to the dominance of gradient updates by charge density (which may be of a large magnitude), the model's total loss function is designed with adaptive weight parameters. and Multi-task loss function:

[0087] in, and For the predicted charge density and charge distribution range, and This represents the actual charge density and charge distribution range.

[0088] (3-3) Model training parameters and Early Stopping mechanism.

[0089] Adam optimizer parameter settings: The Adam optimizer is used for end-to-end supervised learning. The initial learning rate is set to... A cosine annealing learning rate decay strategy was used to avoid getting trapped in local optima. The batch size for model training was set to 64, and the maximum number of epochs was set to 100.

[0090] Early Stopping specific stopping conditions: After each training epoch, the joint loss function (Validation Loss) is calculated on an independent validation set. The tolerance is set to 20 epochs, and the minimum change threshold (Min Delta) is [value missing]. That is: if the decrease in validation set loss is less than 1% over 20 consecutive epochs. If the training process is interrupted or begins to show an upward trend, the training process will be forcibly terminated, and the model weights of the Epoch with the best performance on the validation set will be automatically rolled back and saved.

[0091] S42: Input the electric field signal from the test dataset obtained in step S2 into the charge inversion neural network model trained in step S41 to obtain the predicted charge density and charge distribution range, and compare it with the actual preset charge parameters to test the accuracy of the neural network.

[0092] If the prediction error exceeds the allowable range, the charge inversion neural network model is corrected based on the measured data, and the corrected charge inversion neural network is finally obtained.

[0093] The specific correction process includes the following four progressive steps: (1) Calculate the absolute and relative errors of the model on the test dataset and establish the error matrix.

[0094] For each test sample Calculate the relative error of charge density respectively. The relative error of the charge distribution range : ; ; in, and Let be the model prediction value for the i-th test sample. and This is the actual preset value for the i-th test sample.

[0095] At the same time, the root mean square error (RMSE) of the entire test dataset is calculated to macroscopically assess the overall deviation of the model.

[0096] (2) Analyze the deviation distribution (feature localization): Based on the global error calculation, a multi-dimensional distribution characteristic analysis is performed on the deviation data.

[0097] (2-1) Analysis by magnitude range: The charge density is divided into three magnitude ranges: the slight discharge range is when the charge density is less than The moderate discharge range is The severe discharge range is not less than The mean, standard deviation, and root mean square error of the prediction error within each interval are statistically analyzed to determine whether the model has prediction bias at a specific discharge intensity.

[0098] The overall positive or negative deviation is determined by the mean error, the prediction stability within the interval is determined by the standard deviation of the error, and the overall prediction accuracy is measured by the root mean square error. By comparing the indicators of each interval, it is determined whether the positioning model has significant prediction bias in a specific discharge intensity range. If the absolute value of the mean error in a certain interval is higher than that in other intervals, there is a fixed offset at that discharge intensity; if the standard deviation of the error in a certain interval is higher than that in other intervals, the model generalizes poorly and fluctuates greatly at that discharge intensity; if the root mean square error in a certain interval is higher than that in other intervals, that discharge intensity is a weak range for model prediction.

[0099] (2-2) Spatial Location Analysis: Establishing the Topological Coordinates of the Sensor Array: Using the central axis of the insulator as a reference, establish a two-dimensional polar coordinate system and record the spatial coordinates of each sensor probe. Then, using the charge density distribution map obtained by probe gridding measurement, the coordinates of the location where the maximum charge density is located are regarded as the discharge location. Calculate the Euclidean distance between this discharge location and each sensor. Set distance thresholds based on the actual size of the insulator and the array layout; for example, the near zone is defined as no more than 5 cm, the middle zone as 5 to 15 cm, and the far zone as greater than 15 cm. The principle for determining the thresholds is to ensure that the sample size in each interval is basically balanced. Statistically analyze the prediction errors corresponding to all sensors falling within the same distance interval and compare the error magnitudes of the near, middle, and far zones.

[0100] (2-3) Determining the bias attribute: Evaluate whether the error is random (zero mean distribution) or systematic (with obvious positive or negative bias), so as to provide a basis for subsequent cause identification.

[0101] (3) Identify the causes of the deviation (mechanism attribution).

[0102] Based on the characteristics of the deviation distribution, the sources of error can be attributed to the following typical differences between the simulation and the real domain: (3-1) Mismatch of physical parameters: There are slight differences between the dielectric constant of the ideal insulating material and the ambient temperature and humidity set in the simulation model and the actual epoxy resin model, which leads to the deviation of the electric field attenuation and distortion law.

[0103] (3-2) Sensor and hardware noise: Actual electric field sensing probes and photoelectric detection systems have thermal noise, dark current and signal transmission attenuation, while simulation data are often ideal "clean" signals.

[0104] (3-3) Assembly and process errors: In actual operation, there may be millimeter-level installation deviations in the bonding position of the sensor probe array, resulting in the spatial sampling points and simulation model nodes not being perfectly aligned.

[0105] (4) Implement targeted correction operations and specific correction strategies.

[0106] Based on the identified core causes of bias, perform one or more of the following combined correction operations on the charge inversion neural network: Strategy 1: Fine-tuning using transfer learning based on small sample test data.

[0107] To address systematic deviations caused by physical parameter mismatch and assembly errors, the model is fine-tuned using a test set (actual measurement data).

[0108] Operation method: Freeze the weights of the input layer and the first half of the hidden layer (shallow feature extraction layer) in the neural network model to maintain its ability to recognize the basic waveform of the electric field; only unfreeze the second half of the hidden layer and the output branch layer, using a very small learning rate ( Mini-batch iterative updates are performed on the measured data to enable the model to quickly adapt to the real physical mapping relationship.

[0109] Strategy 2: Data Augmentation and Noise Injection.

[0110] This addresses random biases or overfitting caused by sensor and hardware noise.

[0111] Operation method: Go back to step S3, artificially superimpose white noise that follows a Gaussian distribution or background noise segments extracted from the actual test environment into the original simulation training dataset; at the same time, introduce small spatial displacement perturbations (simulate probe installation errors), reconstruct the enhanced training set and perform secondary training on the model to improve the network's robustness to complex environmental interference.

[0112] Strategy 3: Adaptive reweighting of multi-task loss functions.

[0113] If the deviation analysis finds that the predicted accuracy of charge density is up to standard, but the deviation of the charge distribution range is too large (or vice versa).

[0114] Operation method: Dynamically adjust the total loss function in step S41 Weighting coefficients and Increase the weight penalty for branches with larger deviations, so that in the next round of fine-tuning, the optimization focus will be shifted towards the weaker tasks, until the prediction accuracy of both meets the engineering requirements.

[0115] In this embodiment, in step S5, an array of electric field sensing probes installed on the outer wall of the GIS insulator is used to collect partial discharge field signals. A charge inversion neural network is then used to obtain the charge density and charge distribution range generated on the insulator surface during the discharge process. Figures 5-6 As shown.

[0116] Specifically, it includes: S51: Conduct partial discharge experiments on a real GIS. Use silicone grease to stick needle-shaped metal particles onto the surface of a real GIS insulator, and apply voltage to induce partial discharge of the particles. S52: The electric field sensing probes prepared in step S1 are bonded to the outer wall of the insulator with silicone. One electric field sensing probe is bonded every 50 mm to form a sensing probe array to collect the partial discharge field signal during the pressurization process.

[0117] S53: Input the partial discharge field signal measured in step S52 into the corrected charge inversion neural network in step S4 to invert the charge density and charge distribution range generated on the surface of the insulator during the discharge process.

[0118] Example 2 This embodiment provides an online detection system for partial discharge charge of GIS insulators, including: The simulation module is configured to determine the performance indicators of the required electric field sensing probe by simulating the electric field distribution generated by the charge during partial discharge based on the constructed GIS simulation model, thereby preparing the electric field sensing probe. The test dataset construction module is configured to acquire the charge density, charge distribution range and electric field signal collected by the electric field sensing probe under different voltage pulses based on a pre-built epoxy resin model, thereby constructing the test dataset. The training dataset construction module is configured to acquire electric field signals under different charge densities and different charge distribution radii based on the GIS simulation model, thereby constructing the training dataset. The model training module is configured to train the constructed neural network model based on the training dataset and correct it based on the test dataset to obtain the charge inversion neural network model. The detection module is configured to collect partial discharge field signals using an array of electric field sensing probes installed on the outer wall of the GIS insulator. The partial discharge field signals are then processed using a charge inversion neural network to obtain the charge density and charge distribution range generated on the surface of the insulator during the discharge process.

[0119] It should be noted that the above modules correspond to the steps described in Embodiment 1, and the examples and application scenarios implemented by the above modules and the corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should also be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0120] In further embodiments, the following is also provided: An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0121] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0122] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0123] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0124] The method in Example 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0125] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0126] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0127] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0128] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0129] Those skilled in the art will recognize that the units and algorithm steps described in connection with the various examples of this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention.

[0130] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for online detection of partial discharge charge in GIS insulators, characterized in that, include: Based on the constructed GIS simulation model, the electric field distribution generated by the charge during partial discharge is simulated to determine the performance indicators of the required electric field sensing probe, and the electric field sensing probe is prepared accordingly. Based on the pre-built epoxy resin model, the charge density, charge distribution range and electric field signal collected by the electric field sensing probe under different voltage pulses are obtained to construct a test dataset. Based on the GIS simulation model, electric field signals under different charge densities and different charge distribution radii are obtained to construct a training dataset. The neural network model is trained based on the training dataset and then corrected based on the test dataset to obtain the charge inversion neural network model. The partial discharge field signal is collected by an array of electric field sensing probes installed on the outer wall of the GIS insulator. The charge inversion neural network is used to obtain the charge density and charge distribution range generated on the surface of the insulator during the discharge process.

2. The method for online detection of partial discharge charge in GIS insulators as described in claim 1, characterized in that, Based on the geometric structure and material parameters of the real GIS, a GIS simulation model is established. By setting the discharge charge quantity and charge distribution pattern on the surface of the insulator in the GIS simulation model, the electric field signal at the location of the electric field sensing probe on the outer wall of the insulator is collected. In this way, the performance indicators of the electric field sensing probe required for discharge electric field measurement, such as sensitivity, electric field measurement range, measurement bandwidth and time resolution, are determined.

3. The method for online detection of partial discharge charge in GIS insulators as described in claim 2, characterized in that, The determination of performance indicators specifically includes: The peak value of the electric field signal is used as the upper limit of the electric field measurement range of the electric field sensing probe; With the smallest measurable electric field difference As the lower limit of sensitivity of the electric field sensing probe; in, It is the peak electric field when the surface charge density is at its minimum and the distribution range is the smallest. This is the power frequency background electric field value; The time interval between the rise of the electric field signal pulse from 10% peak value to 90% peak value is taken as the pulse rise time, and 1 / 100 of the pulse rise time is taken as the lower limit of time resolution. The electric field signal is subjected to a fast Fourier transform to obtain the spectral distribution. The frequency value at which the accumulated energy reaches 90% of the total energy is used as the upper limit of the measurement bandwidth.

4. The method for online detection of partial discharge charge in GIS insulators as described in claim 1, characterized in that, The constructed neural network model includes an input layer, hidden layers, and an output branch layer; The input layer receives the electric field signal. After standardizing and preprocessing the electric field signal, it uses multi-channel one-dimensional convolution of different sizes to extract the partial discharge pulse features along the time axis within a sliding window. The hidden layer extracts features from the input layer, extracts the spatial distribution pattern of electric field distortion through stacked residual convolutional blocks, and then inputs them into the output branch layer after attention weighting by the spatial attention module. The output branch layer includes two parallel fully connected branches. The first fully connected branch outputs the predicted maximum surface charge density and distribution gradient, and the second fully connected branch outputs the diffusion radius of the discharge charge on the insulator surface.

5. The method for online detection of partial discharge charge in GIS insulators as described in claim 4, characterized in that, The process of making corrections based on the test dataset includes: For each test sample Calculate the relative error of charge density respectively. The relative error of the charge distribution range : ； ； in, and Let be the model prediction value for the i-th test sample. and The actual preset value for the i-th test sample; The deviation distribution is analyzed based on different charge density ranges, the spatial position of the electric field sensing probe, and the deviation attributes. The causes of the deviation are determined, including physical parameter mismatch, sensor and hardware noise, and assembly process errors. To address the discrepancies caused by physical parameter mismatch and assembly process errors, the weights of the first half of the input layer and hidden layer in the neural network model are frozen, while only the second half of the hidden layer and the output branch layer are unfrozen. Iterative updates are then performed on the test dataset using a set minimum learning rate. To address the discrepancies between sensor and hardware noise, white noise following a Gaussian distribution or background noise segments extracted from the actual test environment are superimposed on the original training dataset. At the same time, spatial displacement perturbation is introduced to reconstruct the enhanced training dataset and perform secondary training. If the prediction accuracy of charge density meets the standard, but the prediction accuracy of charge distribution range does not, then adjust the weight coefficients in the total loss function of the model, increasing the penalty for the weight with the larger deviation between the two, until the prediction accuracy of both meets the requirements.

6. The method for online detection of partial discharge charge in GIS insulators as described in claim 1, characterized in that, In the GIS simulation model, surface charge parameters with different charge densities and different charge distribution radii are set. For each set of charge parameters, electric field signals are collected at the location of the sensing probe array on the outer wall of the insulator. A training dataset is constructed with electric field signals as input and corresponding charge parameters as supervision labels. By pre-setting charges of different charge densities and charge distribution ranges in selected areas on the outer wall surface of the insulator of the epoxy resin model, the charge accumulation state after partial discharge is simulated, and the electric field signal of the partial discharge process is collected synchronously by a sensor probe array to construct a test dataset. Specifically, the electric field sensor probes are placed on the outer wall of the insulator of the epoxy resin model, and one electric field sensor probe is arranged at a set interval to form a sensor probe array. Needle electrodes were placed in the normal direction on the outer wall surface of the epoxy resin model insulator. Different levels of voltage pulses were applied using the needle electrodes. The surface charge density was measured using a probe, and the electric field signal of the discharge process was collected synchronously using a sensor probe array. After the experiment, the surface charge distribution range was determined based on the charge density of all measurement points.

7. An online detection system for partial discharge charge of GIS insulators, characterized in that, include: The simulation module is configured to determine the performance indicators of the required electric field sensing probe by simulating the electric field distribution generated by the charge during partial discharge based on the constructed GIS simulation model, thereby preparing the electric field sensing probe. The test dataset construction module is configured to acquire the charge density, charge distribution range and electric field signal collected by the electric field sensing probe under different voltage pulses based on a pre-built epoxy resin model, thereby constructing the test dataset. The training dataset construction module is configured to acquire electric field signals under different charge densities and different charge distribution radii based on the GIS simulation model, thereby constructing the training dataset. The model training module is configured to train the constructed neural network model based on the training dataset and correct it based on the test dataset to obtain the charge inversion neural network model. The detection module is configured to collect partial discharge field signals using an array of electric field sensing probes installed on the outer wall of the GIS insulator. The partial discharge field signals are then processed using a charge inversion neural network to obtain the charge density and charge distribution range generated on the surface of the insulator during the discharge process.

8. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the method described in any one of claims 1-6.

Images