A method for testing the sensitivity of high voltage bushings to partial discharges simulating the effects of dust intrusion

Abstract

The application provides a high-voltage bushing discharge sensitivity test method simulating dust invasion influence, belongs to the technical field of high-voltage bushings, and is based on a partition asynchronous propulsion algorithm to solve a charge density gradient field and a dust motion trajectory field, carries out wavelet packet multi-scale decomposition on a partial discharge signal to obtain a time-frequency characteristic atlas, inputs the time-frequency characteristic atlas and sparse dust trajectory mark data into an artificial intelligence discharge type identification model trained by a sparse connection learning framework based on dynamic topology reconstruction and long-interval aggregation to output a discharge starting voltage prediction value, monitors dust fractal dimension by using a dust fractal aggregate parallel computing algorithm to trigger particle size distribution updating, adjusts a dust injection rate according to a deviation between the prediction value and a measured value and records discharge sensitivity reference data, and solves the technical problem of insufficient high-voltage bushing discharge sensitivity evaluation precision under dust invasion conditions.

Description

Technical Field

[0001] This invention belongs to the field of high-voltage bushing technology, and more specifically, relates to a method for testing the discharge sensitivity of high-voltage bushings by simulating the effects of dust intrusion. Background Technology

[0002] Ultra-high voltage direct current (UHVDC) bushings are critical insulation devices in power systems, and inevitably face dust intrusion issues during long-term operation. Traditional discharge sensitivity testing methods mainly rely on the assumption of static dust distribution. They measure the discharge initiation voltage by uniformly injecting dust particles of a fixed size and applying a stepped voltage, using a single sensor to collect partial discharge signals and identifying the discharge type based on a threshold. However, in current high-voltage bushing insulation performance evaluations, dust particles exhibit dynamic agglomeration characteristics under the influence of electric field force and airflow drag force. Their spatial distribution and particle size evolution show significant time-varying and non-uniformity. Traditional methods cannot accurately capture the coupling relationship between dust movement trajectory and electric field distortion, resulting in low accuracy in discharge type identification and large deviations in discharge initiation voltage prediction. In other words, existing technologies suffer from insufficient accuracy in evaluating the discharge sensitivity of high-voltage bushings under dust intrusion conditions. Summary of the Invention

[0003] In view of this, the present invention provides a high-voltage bushing discharge sensitivity test method simulating the effects of dust intrusion, which can solve the technical problem of insufficient accuracy in evaluating the discharge sensitivity of high-voltage bushings under dust intrusion conditions in the prior art.

[0004] This invention is implemented as follows: It provides a method for testing the discharge sensitivity of high-voltage bushings by simulating the effects of dust intrusion. A dust simulation injection system is established inside the ultra-high voltage DC through-wall bushing to inject simulated dust particles into the SF6 gas insulation area. Simultaneously, a space charge sensor array is arranged on the surface of the epoxy core. The dust simulation injection system is activated and a DC voltage is applied. A disturbance is generated by a dust trajectory disturbance wave generator. A high-speed camera system and a synchrotron X-ray tomography device are used to acquire the three-dimensional spatial coordinates of the dust particles and the interface charge density distribution. The three-dimensional spatial coordinates of the dust particles are input into a charge transport-dust dynamics coupled calculation module to solve the charge density problem based on a partitioned asynchronous propagation algorithm. The gradient field and dust trajectory field are used to obtain the evolution sequence of dust spatial distribution and electric field distortion coefficient. Wavelet packet multi-scale decomposition is performed on the partial discharge signal. The time-frequency feature spectrum, sparse dust trajectory marker data and electric field distortion coefficient are input into the discharge type identification model to output the dust bridging discharge probability distribution and the predicted value of discharge initiation voltage. The dust agglomeration morphology data are processed by the parallel computing algorithm of dust fractal aggregates. The time series of dust fractal dimension is obtained by recursive box dimension calculation. When the dust fractal dimension exceeds the fractal dimension threshold, the dust particle size distribution update process is triggered. The dust injection rate is adjusted according to the deviation between the predicted value of discharge initiation voltage and the measured value of discharge initiation voltage, and the discharge sensitivity benchmark data is recorded.

[0005] The dust simulation injection system includes a dust storage bin, a pneumatic conveying pipeline, a flow control valve, and an electric field-flow field coupling device. The dust storage bin is connected to the dust inlet of the electric field-flow field coupling device through the pneumatic conveying pipeline, and the flow control valve is installed in the middle section of the pneumatic conveying pipeline to adjust the dust injection rate.

[0006] The electric field-flow field coupling device includes a dust trajectory disturbance wave generator, an electric field modulation electrode grid, and a flow field distribution adjustment chamber. The dust trajectory disturbance wave generator applies a periodic disturbance electric field to the simulated dust particles through the electric field modulation electrode grid.

[0007] Among them, the dust trajectory disturbance wave generator outputs sparse dust trajectory marker data through the dust marker signal generation function, which randomly selects some particles in the dust group to apply identifiable electric field disturbance markers.

[0008] The dust marking signal generation function is expressed as follows: the number of marked dust particles is equal to the total number of dust particles divided by the square root of the standard dust quantity reference value, multiplied by the sparsity coefficient; the amplitude of the marked dust particle perturbation electric field is equal to the current electric field strength divided by the standard electric field strength, multiplied by the average dust particle size divided by the reference particle size, and then multiplied by the perturbation amplitude coefficient.

[0009] The space charge sensor array consists of capacitive charge detection units, which are embedded in grooves on the surface of the epoxy core. The spacing between adjacent capacitive charge detection units is 8 to 15 mm.

[0010] The synchrotron radiation X-ray tomography device includes an X-ray source, a rotating sample stage, a detector array, and an image reconstruction computing unit. The image reconstruction computing unit uses an improved Tikhonov regularized inversion algorithm to reconstruct multi-angle projection data into a three-dimensional charge density distribution.

[0011] Among them, the charge transport-dust dynamics coupled calculation module decomposes the internal calculation domain of the UHVDC through-wall bushing into a charge transport-dominated region and a dust movement-dominated region, where the charge density gradient is greater than 100%. The region is divided into charge transport-dominated regions.

[0012] Among them, the partitioned asynchronous propagation algorithm uses the fourth-order implicit Runge-Kutta scheme to solve the drift diffusion equation in the charge transport-dominated region, and the explicit frog-jump scheme to track the dust particle trajectory in the dust motion-dominated region. The boundary between the two regions is achieved by cubic spline interpolation to realize the loosely coupled transfer of charge density and electric field strength.

[0013] Among them, the partitioned asynchronous advancement algorithm adopts a Warp-level synchronization mechanism, which organizes the 32 threads of the graphics processor into a thread bundle unit. At the end of each time step, the Warp-level synchronization instruction ensures that all threads in the same thread bundle complete the calculation.

[0014] Among them, the wavelet packet multi-scale decomposition adopts the Daubechies wavelet basis function, and the number of decomposition layers is set to 5 to 7. After each layer of decomposition, coefficients with a frequency band range of 50kHz to 2MHz are retained as time-frequency features.

[0015] The structure of the discharge type identification model is as follows: the input layer receives the time-frequency feature spectrum, sparse dust trajectory marker data and electric field distortion coefficient, extracts spatial features through the convolutional layer, enhances gradient propagation through the residual connecting block, aggregates long-interval spatiotemporal features through the multi-head adaptive attention layer, and outputs the dust bridging discharge probability distribution and the predicted value of the discharge initiation voltage through the fully connected layer.

[0016] Among them, the discharge type identification model training adopts a weighted combination of cross-entropy loss function and mean square error loss function as the total loss function, and applies a sparse connection learning framework based on dynamic topology reconstruction and long-interval clustering during the training process.

[0017] Among them, the sparse connection learning framework based on dynamic topology reconstruction and long-interval clustering establishes a fully connected neural network topology in the initial stage of training the firing type identification model. After each training cycle, the activation frequency and activation amplitude of neurons are counted to calculate the mutual information between neurons. When the mutual information is lower than a dynamic threshold, the connection is pruned, and the weights of the remaining connections are grouped. Regularization constraints.

[0018] Among them, the parallel computing algorithm for dust fractal aggregates uses the quadtree spatial decomposition method to recursively subdivide the bounding boxes. In the shared memory of the graphics processor, a 64-bit integer is allocated to each sub-box for bitwise operations to determine the box occupancy status. The box count results of each level are summarized through the Warp reduction technique.

[0019] The dust particle size distribution update process includes calculating the agglomeration factor based on the fractal dimension of the dust, multiplying the average dust particle size by the agglomeration factor and adding 1 to obtain the updated equivalent particle size, recalculating the distribution histogram of the equivalent particle size and fitting it to a log-normal distribution function, and passing the updated particle size distribution parameters to the charge transport-dust dynamics coupled calculation module.

[0020] This invention establishes a coupled computational model of charge transport and dust dynamics, combining sparse dust trajectory marking technology and multi-scale time-frequency analysis methods to achieve accurate tracking of the dynamic evolution of dust. The dust trajectory perturbation wave generator selectively tracks some dust particles by applying identifiable electric field perturbation markers, reducing the computational complexity of global tracking while preserving key information about dust agglomeration characteristics. The partitioned asynchronous advancement algorithm uses different time steps and numerical formats for the charge transport-dominated and dust motion-dominated regions, improving the efficiency of the coupled solution and accurately describing the interaction between the charge density gradient field and the dust motion trajectory field. The sparse connection learning framework based on dynamic topology reconstruction with long-interval aggregation enhances the sensitivity of the discharge type identification model to intermittent dust agglomeration characteristics through adaptive pruning and connection restoration mechanisms, avoiding over-smoothing of sparse events by fully connected networks. Real-time monitoring of the dust fractal dimension time series triggers the particle size distribution update process, ensuring consistency between the computational model and the actual dust state. In summary, this invention solves the technical problem mentioned in the background art of insufficient accuracy in assessing the discharge sensitivity of high-voltage bushings under dust intrusion conditions.

[0021] It should be noted that the dust trajectory disturbance wave generator outputs sparse dust trajectory marker data through a dust marker signal generation function. This function applies identifiable electric field disturbance waves to random areas inside the UHVDC through-wall bushing, rather than precisely controlling the dust particles. The inputs include the total number of dust particles, the average particle size, and the current electric field strength. The outputs are the coordinate sequence of the applied area and the amplitude sequence of the electric field disturbance wave. The electric field disturbance wave acts on the dust particles within the applied area, causing their trajectories to randomly deviate, forming sparse trajectory disturbance features. This improves the robustness of the discharge type identification model to the uncertainty of dust movement. Since dust particles in the actual operating environment are affected by various random factors such as airflow disturbances, temperature gradients, and mechanical vibrations, resulting in unpredictable trajectory deviations, the random electric field disturbance wave is artificially applied during the training phase to simulate these trajectory deviations. This allows the discharge type identification model to learn the correlation between dust trajectory disturbances and discharge characteristics during training, thus maintaining stable identification accuracy even when faced with random changes in dust trajectories during actual operation.

[0022] Furthermore, during the actual testing process after training, an electric field disturbance wave was added to cause the dust particle trajectory to shift randomly within a controlled range. By comparing the changes in the predicted discharge initiation voltage and the dust bridging discharge probability distribution output by the discharge type identification model before and after the disturbance was applied, the anti-interference ability of the discharge type identification model against the uncertainty of dust trajectory was verified. When the change amplitude was less than 8%, the discharge type identification model was deemed to have sufficient robustness. The verification process provides additional assurance for the reliability of the discharge sensitivity benchmark data. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method of the present invention.

[0024] Figure 2 A schematic diagram of the composition and structure of a dust simulation injection system.

[0025] Figure 3 This is a schematic diagram of the partitioned solution of the charge transport-dust dynamics coupled calculation module.

[0026] Figure 4 The dust bridging discharge probability distribution map output by the discharge type identification model.

[0027] Figure 5 This is a comparative analysis chart of discharge initiation voltages under different operating conditions. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0029] like Figure 1 The diagram shows a flowchart of a high-voltage bushing discharge sensitivity test method for simulating the effects of dust intrusion, provided by the present invention. This method includes the following steps: S01. A dust simulation injection system is established inside the UHVDC through-wall bushing. Simulated dust particles with a particle size distribution of 1 to 50 μm are injected into the SF6 gas insulation area through an electric field-flow field coupling device. At the same time, a space charge sensor array is uniformly arranged on the surface of the epoxy core. S02. Start the dust simulation injection system and apply DC voltage. Generate sparse dust trajectory marking data through the dust trajectory disturbance wave generator inside the electric field-flow field coupling device. Use a high-speed camera system and a synchrotron radiation X-ray tomography device to collect the three-dimensional spatial coordinates of dust particles and the interface charge density distribution in real time. S03. Input the collected three-dimensional spatial coordinates of dust particles into the charge transport-dust dynamics coupled calculation module, and solve the charge density gradient field and dust motion trajectory field based on the partitioned asynchronous propagation algorithm to obtain the dust spatial distribution evolution sequence and electric field distortion coefficient at different times. S04. Perform wavelet packet multi-scale decomposition on the collected partial discharge signal, input the time-frequency feature spectrum, sparse dust trajectory marking data and electric field distortion coefficient into the discharge type identification model, and output the dust bridging discharge probability distribution and the predicted value of the discharge initiation voltage. S05. Use the parallel computing algorithm for dust fractal aggregates to process dust agglomeration morphology data. Obtain the time series of dust fractal dimension through recursive box dimension calculation. When the dust fractal dimension exceeds the fractal dimension threshold, trigger the dust particle size distribution update process. S06. Adjust the dust injection rate according to the deviation between the predicted value of the discharge initiation voltage and the measured value of the discharge initiation voltage. When the absolute value of the deviation is greater than 12%, increase the dust injection rate and repeat steps S02 to S05. When the absolute value of the deviation is less than 5%, record the current operating parameters as the discharge sensitivity reference data.

[0030] The dust simulation injection system includes a dust storage bin, a pneumatic conveying pipeline, a flow control valve, and an electric field-flow field coupling device. The dust storage bin is connected to the dust inlet of the electric field-flow field coupling device through the pneumatic conveying pipeline. The flow control valve is installed in the middle section of the pneumatic conveying pipeline to adjust the dust injection rate. The gas outlet of the electric field-flow field coupling device is connected to the SF6 gas insulation area of ​​the ultra-high voltage DC through-wall bushing.

[0031] The electric field-flow field coupling device is equipped with a dust trajectory disturbance wave generator, an electric field modulation electrode grid, and a flow field distribution adjustment chamber. The dust trajectory disturbance wave generator applies a periodic disturbance electric field to the simulated dust particles through the electric field modulation electrode grid. The flow field distribution adjustment chamber is equipped with multiple adjustable-angle guide plates to control the flow direction of SF6 gas.

[0032] The dust trajectory disturbance wave generator outputs sparse dust trajectory marking data through a dust marking signal generation function. The dust marking signal generation function is used to randomly select some particles in the dust group to apply identifiable electric field disturbance markings. The inputs include the total number of dust particles, the average particle size of the dust, and the current electric field strength. The outputs are the marked dust particle number sequence and the marked dust particle disturbance electric field amplitude sequence.

[0033] The dust marking signal generation function is expressed as follows: the number of marked dust particles is equal to the total number of dust particles divided by the square root of the standard dust quantity reference value, multiplied by a sparsity coefficient; the amplitude of the perturbation electric field of the marked dust particles is equal to the current electric field strength divided by the standard electric field strength, multiplied by the average dust particle size divided by the reference particle size, and then multiplied by the perturbation amplitude coefficient, where the standard dust quantity reference value is taken as... The sparsity coefficient is 0.05 to 0.15, the standard electric field strength is 500 kV / m, the reference particle size is 10 μm, and the perturbation amplitude coefficient is 0.2 to 0.4.

[0034] The space charge sensor array consists of several capacitive charge detection units, which are embedded in grooves with a depth of 0.5 to 1.5 mm on the surface of the epoxy core. The spacing between adjacent capacitive charge detection units is 8 to 15 mm. Each capacitive charge detection unit is connected to a charge signal amplifier via a shielded coaxial cable.

[0035] The high-speed camera system includes three orthogonally arranged high-speed cameras and a synchronous trigger controller. The optical axes of the three orthogonally arranged high-speed cameras are perpendicular to each other and their intersection point is located on the central axis of the UHVDC through-wall bushing. The shooting frame rate is set to 5,000 to 10,000 frames per second, and the exposure time is set to 20 to 50 μs.

[0036] The synchrotron radiation X-ray tomography device includes an X-ray source, a rotating sample stage, a detector array, and an image reconstruction computing unit. The X-ray source emits a monochromatic X-ray beam with an energy of 15 to 25 keV. The rotating sample stage rotates at an angular velocity of 0.5 to 2 degrees per second to acquire multi-angle projection data. The image reconstruction computing unit uses an improved Tikhonov regularized inversion algorithm to reconstruct the multi-angle projection data into a three-dimensional charge density distribution.

[0037] The interface charge density distribution refers to the three-dimensional distribution of space charge density at the interface between the epoxy core and SF6 gas. It is obtained by acquiring multi-angle projection data through a synchrotron X-ray tomography device and processing it through an image reconstruction computing unit.

[0038] The charge transport-dust tail dynamics coupled calculation module decomposes the internal calculation domain of the UHVDC through-wall bushing into a charge transport-dominated region and a dust movement-dominated region. The charge density gradient near the epoxy core surface is greater than... The area is divided into a charge transport-dominated region, and the area where the SF6 gas flow velocity is greater than 0.5 m / s is divided into a dust movement-dominated region.

[0039] The partitioned asynchronous propagation algorithm uses a fourth-order implicit Runge-Kutta scheme to solve the drift-diffusion equation in the charge transport-dominant region, with a time step set to [value missing]. to In the dust-dominant region, an explicit frog-jump scheme is used to track the trajectory of dust particles, with the time step set to [value missing]. to The two regions are connected by a loosely coupled transfer of charge density and electric field strength through cubic spline interpolation.

[0040] The partitioned asynchronous advancement algorithm adopts a Warp-level synchronization mechanism, which organizes the 32 threads of the GPU into a thread bundle unit. At the end of each time step, the Warp-level synchronization instruction ensures that all threads in the same thread bundle have completed their calculations before proceeding to the next time step.

[0041] The charge density gradient field is the gradient vector field of charge density in three-dimensional space, which is obtained by solving the drift-diffusion equation through the charge transport-dust dynamics coupled calculation module.

[0042] The dust motion trajectory field is a sequence of position coordinates of dust particles in three-dimensional space that changes over time, and is obtained by tracking them using an explicit frog-jump scheme through a charge transport-dust dynamics coupling calculation module.

[0043] The dust spatial distribution evolution sequence is a sequence formed by arranging the distribution state of dust particles in three-dimensional space at different times in chronological order, and is output by the charge transport-dust dynamics coupling calculation module.

[0044] The electric field distortion coefficient is defined as the absolute value of the local electric field strength divided by the average electric field strength minus 1. When the electric field distortion coefficient is greater than 0.3, the region is considered to have significant electric field distortion and is marked as a discharge sensitive region.

[0045] The partial discharge signal is an electrical pulse signal generated when partial discharge occurs inside the UHVDC through-wall bushing, and is acquired by a partial discharge detection sensor installed outside the UHVDC through-wall bushing.

[0046] The wavelet packet multi-scale decomposition uses the Daubechies wavelet basis function, with the number of decomposition layers set to 5 to 7. After each layer of decomposition, coefficients with a frequency range of 50kHz to 2MHz are retained as time-frequency features, and the coefficients of each layer are arranged according to the time axis to form a two-dimensional time-frequency feature map.

[0047] The time-frequency feature map is a two-dimensional matrix image formed by arranging the coefficients of each layer of the partial discharge signal according to the time axis and frequency axis after wavelet packet multi-scale decomposition. The matrix size is 128×128 pixels.

[0048] Optionally, when recording discharge sensitivity benchmark data in step S06, an electric field disturbance wave is continuously applied and steps S02 to S05 are repeated 3 to 5 times. The measured values ​​of the discharge initiation voltage at different electric field disturbance wave amplitudes and application locations are collected. The standard deviation of the multiple measurement results is calculated. When the standard deviation is less than 6% of the average value, it is determined that the discharge sensitivity benchmark data still maintains stability under the influence of dust trajectory disturbance. The average value of the multiple measurement results is stored as the final discharge sensitivity benchmark data in the discharge sensitivity lookup table. The repeated measurement mechanism enhances the adaptability of the discharge sensitivity benchmark data to uncertainties in the actual operating environment by covering multiple random states of dust movement, ensuring that the insulation performance assessment and operation and maintenance decisions based on the benchmark data have higher credibility and engineering practical value.

[0049] The structure of the discharge type identification model is as follows: the input layer receives a 128×128 pixel time-frequency feature map, sparse dust trajectory marker data, and electric field distortion coefficient. Spatial features are extracted through three convolutional layers. Each convolutional layer is followed by a batch normalization layer and an activation layer. Then, gradient propagation is enhanced through two residual connection blocks. Next, long-interval spatiotemporal features are aggregated through a multi-head adaptive attention layer. Finally, two fully connected layers are connected to output the dust bridging discharge probability distribution and the predicted discharge initiation voltage. The number of heads in the multi-head adaptive attention layer is dynamically determined based on the average dust particle size, electric field distortion coefficient, and dust fractal dimension.

[0050] The steps for establishing the training dataset for the discharge type identification model specifically include: applying a DC voltage ranging from 200 to 800 kV to the UHVDC through-wall bushing under laboratory conditions, injecting simulated dust particles with a particle size distribution of 1 to 50 μm, collecting partial discharge signals and dust distribution images under different dust concentrations and voltage levels, identifying three types of discharge—corona discharge, surface discharge, and dust bridging discharge—through manual annotation, collecting 3,000 to 5,000 samples for each type, dividing the samples into a training set and a validation set at an 8:2 ratio, and using sparse dust trajectory marker data generated by a dust trajectory perturbation wave generator as auxiliary annotation information to enhance the diversity of the training set.

[0051] The specific steps for training the discharge type identification model include: using a weighted combination of cross-entropy loss function and mean squared error loss function as the total loss function, with a weight ratio of 3:7; using the Adam optimizer to update parameters; setting the initial learning rate to 0.001; reducing the learning rate to 0.8 times the original value every 10 training epochs; setting the batch size to 32; and having a total of 150 to 200 training epochs. During the training process, a sparse connection learning framework based on long-interval clustering of dynamic topology reconstruction is applied.

[0052] The principle of the sparse connection learning framework based on dynamic topology reconstruction with long-interval aggregation is as follows: simulating the intermittent aggregation characteristics and non-uniform spatial distribution characteristics of dust particles under the action of electric field force and airflow drag force, that is, simulating the dense aggregation of dust particles at certain times and spatial locations while sparsely distributed at other times and locations. The long-interval aggregation leads to a sparse and strongly correlated pattern in the time-frequency characteristics of partial discharge signals in both time and spatial dimensions. Traditional fully connected networks cannot effectively capture the sparsity characteristics and cause a waste of computational resources.

[0053] The reason for the long-interval aggregation of simulated dust particles is that, in terms of time, the simulated dust particles are affected by the periodically changing eddies, resulting in alternating agglomeration and dispersion processes. The agglomeration time lasts for several milliseconds before entering the dispersion stage for tens of milliseconds. In terms of space, the non-uniformity of the electric field causes the simulated dust particles to preferentially migrate to areas with large electric field intensity gradients and form high-concentration clusters locally, while the dust concentration in other areas remains at a low level. The spatiotemporal intermittent distribution characteristics require the discharge type identification model to have the ability to model sparse long-range dependencies.

[0054] The implementation of the sparse connection learning framework based on dynamic topology reconstruction and long-interval clustering in the discharge type identification model is as follows: A fully connected neural network topology is established at the initial stage of training the discharge type identification model. After each training cycle, the activation frequency and activation amplitude of each neuron are statistically analyzed. The mutual information between any two neurons is calculated. When the mutual information is lower than a dynamic threshold, the connection is pruned probabilistically. Simultaneously, the weights of the remaining connections are grouped. Regularization constraints encourage the weights of connections within the same group to become consistent. Every 5 training cycles, the necessity of pruned connections is reassessed based on the performance of the validation set, and some connections are restored with a small probability, ultimately forming a sparse network topology that retains key long-range dependencies.

[0055] The proposed sparse connection learning framework based on dynamic topology reconstruction and long-interval aggregation brings the following technical benefits to the discharge type identification model and the entire scheme: The framework significantly reduces the number of parameters and computational complexity of the discharge type identification model through adaptive sparse network connections, enabling the model to process high-resolution time-frequency feature maps with limited computational resources. Simultaneously, the sparse connection structure enhances the sensitivity to the intermittent characteristics of simulated dust particle aggregation, avoiding excessive smoothing of sparse events by fully connected networks. The dynamic topology reconstruction mechanism endows the discharge type identification model with the ability to adjust its internal representation according to actual dust distribution patterns, improving the generalization performance of discharge type identification under different operating conditions. Grouping regularization constraints maintain the expressive capacity of the discharge type identification model, preventing underfitting caused by sparsity. The introduction of sparse dust trajectory labeling data provides physical prior guidance for sparse connections, making the pruning process more consistent with the inherent laws of dust dynamics. Overall, the framework achieves synergistic optimization of the efficiency and accuracy of the discharge type identification model, providing reliable technical support for real-time discharge sensitivity assessment.

[0056] The number of heads in the multi-head adaptive attention layer is calculated as follows: the number of heads in the multi-head adaptive attention layer is equal to the base number of heads multiplied by the average dust particle size divided by the reference particle size, then multiplied by the electric field distortion coefficient divided by the standard distortion coefficient, plus the sum of the dust fractal dimension divided by the reference fractal dimension. The calculation result is rounded up and limited to between 4 and 16, where the base number of heads is 4, the reference particle size is 10 μm, the standard distortion coefficient is 0.3, and the reference fractal dimension is 1.8.

[0057] The dust bridging discharge probability distribution is the probability distribution of dust bridging discharge occurring at different spatial locations inside the UHVDC through-wall bushing, obtained through the output of the discharge type identification model.

[0058] The predicted discharge initiation voltage is the lowest voltage value for partial discharge predicted by the discharge type identification model based on the time-frequency characteristic spectrum, sparse dust trajectory marking data, and electric field distortion coefficient.

[0059] The dust agglomeration morphology data is the spatial geometric morphology information of dust particles during the agglomeration process. It is obtained by acquiring the three-dimensional spatial coordinates of dust particles through a high-speed camera system and then processing the images.

[0060] The parallel computing algorithm for dust fractal aggregates uses a quadtree spatial decomposition method to recursively subdivide the bounding box. The initial bounding box side length is set to the inner diameter of the UHVDC through-wall bushing. Each recursion divides the bounding box into 8 sub-boxes along the three coordinate axes. The recursion depth is set to 8 to 12 layers. In the GPU shared memory, a 64-bit integer is allocated to each sub-box for bitwise operations to determine the box occupancy status.

[0061] The parallel computing algorithm for dust fractal aggregates uses Warp reduction technology to summarize the box counting results of each level. The 32 threads in each thread bundle are responsible for counting 32 sub-boxes respectively. The 32 count values ​​are accumulated into a total value through 5 rounds of reduction operations. The time of a single reduction operation is less than 10ns, and the total reduction time increases logarithmically with the recursion depth.

[0062] The recursive box dimension is calculated by statistically analyzing the number of occupied boxes at different scales. A scatter plot is drawn with the logarithm of the box side length as the x-axis and the logarithm of the number of occupied boxes as the y-axis. The least squares method is used to fit a straight line and the negative value of the slope is calculated as the estimated value of the dust fractal dimension.

[0063] The dust fractal dimension time series is a numerical sequence formed by the change of dust fractal dimension over time. It is obtained by recursively calculating the box dimension of dust agglomeration morphology data at different times using a parallel computing algorithm for dust fractal aggregates.

[0064] The fractal dimension threshold is set to 2.2 to 2.5. When the fractal dimension of the dust exceeds the fractal dimension threshold, it indicates that the agglomeration degree of the simulated dust particles has reached a level that affects the discharge characteristics, and the dust particle size distribution function needs to be updated to reflect the increase in equivalent particle size caused by agglomeration.

[0065] The dust particle size distribution update process includes: calculating the agglomeration factor based on the dust fractal dimension, where the agglomeration factor is equal to the dust fractal dimension divided by the reference fractal dimension minus 1; multiplying the average dust particle size by the agglomeration factor and adding 1 to obtain the updated equivalent particle size; recalculating the distribution histogram of the equivalent particle size and fitting it to a log-normal distribution function; and passing the updated particle size distribution parameters to the charge transport-dust dynamics coupled calculation module.

[0066] The measured value of the discharge initiation voltage is the lowest voltage value at which partial discharge occurs, obtained by actual measurement using a partial discharge detection sensor.

[0067] The absolute value of the deviation is calculated as follows: the absolute value of the deviation is equal to the predicted value of the discharge initiation voltage minus the measured value of the discharge initiation voltage, divided by the measured value of the discharge initiation voltage, and then multiplied by 100%. When the absolute value of the deviation is greater than 12%, it is determined that the prediction accuracy of the discharge type identification model is insufficient and the dust injection rate needs to be increased to expand the coverage of the training data. When the absolute value of the deviation is less than 5%, it is determined that the prediction accuracy of the discharge type identification model meets the requirements and the current operating parameters are representative.

[0068] The current operating parameters include dust injection rate, average dust particle size, electric field distortion coefficient, and dust fractal dimension.

[0069] The discharge sensitivity benchmark data includes dust injection rate, average dust particle size, electric field distortion coefficient, dust fractal dimension, and measured discharge initiation voltage. The discharge sensitivity benchmark data is used to establish a discharge sensitivity lookup table under different dust conditions, providing a reference for the insulation performance evaluation and operation and maintenance of UHVDC through-wall bushings.

[0070] As an optional implementation, the present invention also provides a high-voltage bushing discharge sensitivity testing system implemented by a computer, wherein the computer is provided with a readable storage medium storing program instructions, and the program instructions execute the above-described high-voltage bushing discharge sensitivity testing method simulating the influence of dust intrusion when the computer is run.

[0071] The specific implementation methods of the above steps are described in detail below.

[0072] The specific implementation of step S01 is as follows: First, a dust injection position is selected in the SF6 gas insulation area of ​​the UHVDC through-wall bushing. The position is located 20 to 30 mm radially outward of the epoxy core. Simulated dust particles with a particle size distribution of 1 to 50 μm from the dust storage chamber are transported to the electric field-flow field coupling device at a controllable flow rate through a pneumatic conveying pipeline. The flow control valve adjusts the cross-sectional area of ​​the pipeline according to the preset dust injection rate to achieve precise flow control. The reference value for the dust injection rate is... to The charge density is measured at a rate of kg / s. Simultaneously, capacitive charge detection units are arranged at uniform intervals of 8 to 15 mm on the surface of the epoxy core to form a space charge sensor array. The arrangement density ensures that the spatial resolution of the charge density measurement reaches the sub-millimeter level. Each capacitive charge detection unit is connected to a charge signal amplifier through a shielded coaxial cable to reduce electromagnetic interference. The purpose of this step is to establish the basic conditions for controllable dust injection and real-time monitoring of charge density, providing a stable experimental environment and high-quality measurement data for subsequent discharge characteristic analysis.

[0073] The specific implementation of step S02 is as follows: After starting the dust simulation injection system, a DC voltage of 200 to 800 kV is applied to the UHVDC through-wall bushing. The dust trajectory disturbance wave generator in the electric field-flow field coupling device calculates the number of marked dust particles and the amplitude sequence of the marked dust particle disturbance electric field according to the dust mark signal generation function. A periodic disturbance electric field with an amplitude of 0.2 to 0.4 times the current electric field strength is applied to the selected simulated dust particles through the electric field modulation electrode grid. The frequency of the disturbance electric field is set to 100 to 500 Hz to form identifiable trajectory feature marks. Three orthogonally arranged high-speed cameras of the high-speed camera system synchronously acquire dust particle motion images at a shooting frame rate of 5000 to 10000 frames per second. The synchronous trigger controller ensures... The exposure time deviation of the three cameras is less than 1 μs. The rotating sample stage of the synchrotron X-ray tomography device rotates at an angular velocity of 0.5 to 2 degrees per second to acquire projection data at 180 angles. The image reconstruction calculation unit uses an improved Tikhonov regularized inversion algorithm to iteratively reconstruct the projection data. The algorithm effectively suppresses the influence of projection data noise on the reconstruction results by introducing a regularization parameter to balance the data fitting accuracy and solution smoothness. Finally, a three-dimensional charge density distribution with a spatial resolution better than 0.5 mm is obtained. The purpose of this step is to simultaneously acquire the three-dimensional spatial coordinates of dust particles, sparse dust trajectory marker data, and interface charge density distribution, so as to provide complete initial and boundary conditions for the coupled calculation of charge transport and dust dynamics.

[0074] The specific implementation of step S03 is as follows: The three-dimensional spatial coordinates of the dust particles and the interface charge density distribution collected in step S02 are input into the charge transport-dust dynamics coupled calculation module. This module first divides the calculation domain into a charge transport-dominated region and a dust motion-dominated region based on the charge density gradient and SF6 gas flow velocity. The charge density gradient threshold is set to... C / The SF6 gas flow velocity threshold was set to 0.5 m / s. In the charge transport-dominated region, the time derivative terms of the drift-diffusion equation were discretized using a fourth-order implicit Runge-Kutta scheme. This scheme possesses A-stability and can handle rigid problems. The time step was set to... to To ensure the stability of the numerical solution, an explicit leapfrog scheme is used to update the position and velocity of dust particles in the dust-dominated region. This scheme reduces phase error through staggered time steps of velocity and position, with the time step size set to [value missing]. to The s satisfies the Courant condition. The boundary between the two regions is continuously transferred through a cubic spline interpolation function to achieve the transfer of charge density and electric field strength. The Warp-level synchronization mechanism synchronizes the GPU thread bundle at the end of each time step to avoid performance loss caused by thread differentiation. The calculation module outputs the charge density gradient field and dust motion trajectory field at different times. Based on the trajectory field, the distribution density of dust particles in the spatial grid is statistically analyzed to form a dust spatial distribution evolution sequence. At the same time, the ratio of local electric field strength to average electric field strength is calculated and 1 is subtracted and the absolute value is taken to obtain the electric field distortion coefficient. The purpose of the above steps is to efficiently solve the strongly coupled problem of charge transport and dust dynamics through the partitioned asynchronous advancement algorithm, obtain the dynamic evolution law of dust particle motion and electric field distribution, and provide key physical field parameters for subsequent discharge type identification.

[0075] The specific implementation of step S04 is as follows: The partial discharge signal acquired in step S02 is decomposed into 5 to 7 layers of wavelet packets using the Daubechies wavelet basis function. Each layer of decomposition divides the signal frequency band into high-frequency and low-frequency parts, retaining wavelet coefficients in the frequency band range of 50kHz to 2MHz. This frequency band covers the main spectral components of dust bridging discharge. The wavelet coefficients of each layer are arranged according to the time axis and frequency axis to form a 128×128 pixel time-frequency feature map. This map reflects the energy distribution characteristics of the discharge signal in the time-frequency domain. The time-frequency feature map, the sparse dust trajectory marker data generated in step S02, and the electric field distortion coefficient calculated in step S03 are input into the discharge type identification model. The three convolutional layers of the discharge type identification model use 3×3 convolutional kernels to extract the local spatial distribution of the time-frequency feature map. In this process, a batch normalization layer normalizes the output of the convolutional layer to accelerate training convergence. Two residual connection blocks alleviate the gradient vanishing problem in deep networks through skip connections. A multi-head adaptive attention layer calculates the number of heads based on the average particle size of the dust, the electric field distortion coefficient, and the fractal dimension of the dust, with a reference value of 4 to 16. The attention layer captures long-interval spatiotemporal dependencies by calculating the similarity weights between different positions of time-frequency features. Two fully connected layers map the feature vector output by the attention layer to the dust bridging discharge probability distribution and the predicted discharge initiation voltage. The purpose of these steps is to extract the feature patterns of dust bridging discharge from the time-frequency feature map using a deep learning model, and to improve the accuracy of discharge type identification by combining dust trajectory marking data and electric field distortion coefficient, thereby providing quantitative prediction results for discharge sensitivity assessment.

[0076] The specific implementation of step S05 is as follows: Dust agglomeration morphology data are extracted from the three-dimensional spatial coordinates of dust particles collected in step S02. An initial bounding box is established using a quadtree spatial decomposition method. The side length of the bounding box is equal to the inner diameter of the UHVDC through-wall bushing. The bounding box is recursively subdivided into 8 sub-boxes along the three coordinate axes, with a recursion depth set to 8 to 12 levels. A 64-bit integer is allocated to each sub-box in the GPU shared memory. Bitwise operations are used to determine whether dust particles exist in the sub-box and record the occupancy status. Warp reduction technology is used to summarize the sub-box counts handled by 32 threads into a total value through 5 rounds of binary tree reduction. A single reduction operation takes less than 10 ns. The number of occupied boxes at different recursion levels is counted. A scatter plot is drawn with the logarithm of the box side length as the x-axis and the logarithm of the number of occupied boxes as the y-axis. The least squares method is used to fit a straight line and the negative value of the slope is calculated to obtain the dust fractal dimension. The fractal dimension reflects the spatial filling characteristics and complexity of dust agglomerates. The above calculation process is repeated for dust agglomeration morphology data at different times to form a time series of dust fractal dimension. When the dust fractal dimension exceeds the threshold of 2.2 to 2.5, the dust particle size distribution update process is triggered. The agglomeration factor is calculated as the dust fractal dimension divided by the reference fractal dimension of 1.8 and then subtracted by 1. The average dust particle size is multiplied by the agglomeration factor and then added by 1 to obtain the updated equivalent particle size. The distribution histogram of the equivalent particle size is recalculated and fitted to a log-normal distribution function. The updated particle size distribution parameters are passed to the charge transport-dust dynamics coupled calculation module for the next round of iteration calculation. The purpose of this step is to monitor the degree of dust agglomeration in real time through the parallel calculation algorithm of dust fractal agglomerates. When agglomeration causes a significant increase in the equivalent particle size, the dust particle size distribution parameters are updated to ensure the consistency between the dust dynamics model and the actual physical process in subsequent calculations.

[0077] The specific implementation of step S06 is as follows: Calculate the difference between the predicted discharge initiation voltage output in step S04 and the measured discharge initiation voltage actually measured by the partial discharge detection sensor, divide by the measured discharge initiation voltage, and multiply by 100% to obtain the absolute value of the deviation. When the absolute value of the deviation is greater than 12%, it is determined that the prediction accuracy of the discharge type identification model is insufficient. Increase the dust injection rate to increase the dust concentration by 20% to 50%, and re-execute steps S02 to S05 to collect new training data to expand the coverage of the training dataset of the discharge type identification model. When the absolute value of the deviation is less than 5%, it is determined that the prediction accuracy of the discharge type identification model meets the engineering requirements. The application requires recording the current dust injection rate, average dust particle size, electric field distortion coefficient, dust fractal dimension, and measured discharge initiation voltage as discharge sensitivity benchmark data. This benchmark data is used to establish a discharge sensitivity lookup table under different dust conditions, providing a quantitative basis for evaluating the insulation performance of UHVDC through-wall bushings and making preventive maintenance decisions in actual operation. The purpose of this step is to dynamically adjust the experimental conditions based on the prediction error through a closed-loop feedback mechanism to ensure that the obtained discharge sensitivity benchmark data has sufficient representativeness and accuracy, while verifying the predictive robustness of the discharge type identification model under different dust concentrations.

[0078] It should be noted that one of the key technical ideas of this invention is a sparse connection learning framework based on dynamic topology reconstruction and long-interval aggregation. This framework targets the spatiotemporal intermittent aggregation characteristics of dust particles under the influence of electric fields and airflow. It constructs a sparse network topology structure that retains key long-range dependencies through adaptive pruning and recovery mechanisms of neural network connections. Compared with traditional fully connected networks, this framework significantly reduces the number of model parameters and computational complexity, enabling the model to process high-resolution time-frequency feature maps with limited computing resources. At the same time, the sparse connection structure enhances the sensitivity to the intermittent features of dust aggregation, avoiding the over-smoothing of sparse events by fully connected networks. The dynamic topology reconstruction mechanism gives the model the ability to adjust its internal representation according to the actual dust distribution pattern, improving the generalization recognition performance of discharge types under different operating conditions. Grouping regularization constraints maintain the network's expressive capacity and prevent underfitting caused by sparsity.

[0079] The second key technical idea of ​​this invention is a partitioned asynchronous propagation algorithm for charge transport-dust dynamics coupling. The algorithm decomposes the computational domain into a charge transport-dominated region and a dust motion-dominated region based on the charge density gradient and gas flow velocity. In different regions, numerical formats and time steps adapted to the physical process characteristics are adopted. An implicit format is used in the charge transport-dominated region to ensure the numerical stability of rigid problems, while an explicit format is used in the dust motion-dominated region to reduce the computational load. Loose coupling between the two regions is achieved through interpolation boundary conditions. Each region can independently select a time step for asynchronous propagation. Compared with traditional strongly coupled iterative methods, the algorithm avoids the Jacobian matrix dimension explosion problem and optimizes GPU thread execution efficiency through a Warp-level synchronization mechanism, significantly shortening the computation time of multiphysics coupling simulation while ensuring computational accuracy.

[0080] The third key technical idea of ​​this invention is the mechanism of generating sparse dust trajectory label data by a dust trajectory perturbation wave generator. This mechanism forms trajectory feature labels by applying identifiable periodic perturbation electric fields to some dust particles. The sparse label data serves as auxiliary annotation information to enhance the diversity of the training dataset, providing physical prior guidance for sparse connection learning in neural networks. This makes the pruning process more in line with the inherent laws of dust dynamics. Compared with the traditional label-free training method, this mechanism improves the model's learning efficiency of the correlation between dust trajectory and discharge characteristics without increasing the cost of manual annotation. At the same time, the sparse labeling strategy avoids changes in the electric field distribution caused by perturbation to all dust particles, ensuring the authenticity of the experimental conditions.

[0081] The synergistic effect of the three key technical approaches is manifested in the following ways: the sparse labeled data generated by the dust trajectory disturbance wave generator provides physical prior guidance for the sparse connection learning framework based on dynamic topology reconstruction, making the sparsification process of the neural network consistent with the dynamic laws of dust, and improving the model's ability to capture long-interval aggregation features; the partitioned asynchronous propagation algorithm coupled with charge transport and dust dynamics efficiently solves the charge density gradient field and dust trajectory field, providing high-quality input features for the neural network; at the same time, the real-time calculation results of the dust fractal dimension are used to dynamically adjust the number of heads in the multi-head attention layer of the neural network, realizing the adaptive response of the model structure to the dust aggregation state. The three form a closed-loop optimization system of data acquisition, physical field calculation and intelligent recognition. Compared with the traditional single technical path, the synergistic effect significantly improves the accuracy of discharge type identification and the reliability of discharge sensitivity assessment while ensuring computational efficiency, providing a systematic technical solution for the accurate assessment of the insulation performance of UHVDC through-wall bushings.

[0082] It should be noted that this invention also solves the following technical problem: the low computational efficiency of real-time monitoring of dust agglomeration morphology in existing technologies. Traditional fractal dimension calculation methods use serial box counting algorithms, which can take several seconds to tens of seconds for a single calculation of a three-dimensional spatial distribution containing tens of thousands to hundreds of thousands of dust particles, failing to meet the real-time requirements of discharge sensitivity testing. This invention uses a parallel computation algorithm for dust fractal aggregates, employing a quadtree spatial decomposition method to recursively subdivide bounding boxes. Bitwise operation tags are allocated to each sub-box in the shared memory of the graphics processor, and thread bundle reduction technology is used to summarize the box counting results at each level in parallel, reducing the time for a single fractal dimension calculation to the millisecond level. Simultaneously, the recursive box dimension calculation obtains the dust fractal dimension estimate by statistically analyzing the number of occupied boxes at different scales and fitting a logarithmic scatter plot. The real-time update of the dust fractal dimension time series triggers the dust particle size distribution update process, ensuring that the charge transport-dust dynamics coupled calculation module always solves based on accurate dust state parameters, thus improving the reliability of discharge initiation voltage prediction.

[0083] Specifically, the principle of this invention is as follows: The technical solution of this invention can solve the technical problem of insufficient accuracy in evaluating the discharge sensitivity of high-voltage bushings under dust intrusion conditions. This is because it establishes a physical-data hybrid driving model for the dynamic evolution of dust and discharge characteristics. The sparse dust trajectory marking technology reduces the tracking computation to a few percent of the global tracking while ensuring statistical representativeness by randomly selecting a portion of particles to apply electric field perturbations, thus providing key boundary conditions for coupled calculations of charge transport and dust dynamics. The partitioned asynchronous propagation algorithm decomposes the computational domain into different physically dominant regions based on the charge density gradient and airflow velocity. In the charge transport dominant region, a small-time-step implicit scheme is used to accurately solve the drift diffusion equation, while in the dust motion dominant region, a large-time-step explicit scheme is used to efficiently track particle trajectories. Cubic spline interpolation is used to achieve loosely coupled transfer of physical quantities between regions, ensuring numerical accuracy while avoiding the computational burden caused by small time steps across the entire domain. The sparse connection learning framework based on dynamic topology reconstruction and long-interval aggregation utilizes the spatiotemporal intermittent characteristics of dust agglomeration. It evaluates the necessity of neuron connections through mutual information and dynamically adjusts the network topology, enabling the discharge type identification model to retain the ability to identify key events such as dust bridging discharge under sparse connection conditions. Real-time monitoring of dust fractal dimension provides a physical basis for particle size distribution updates, forming a closed-loop feedback mechanism to ensure the matching of model parameters with actual working conditions.

[0084] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0085] The specific implementation of step S01 is as follows: A dust simulation injection system is established inside the UHVDC through-wall bushing. The dust storage chamber is connected to the dust inlet of the electric field-flow field coupling device through a pneumatic conveying pipeline. A flow control valve is installed in the middle section of the pneumatic conveying pipeline to adjust the dust injection rate. The gas outlet of the electric field-flow field coupling device is connected to the SF6 gas insulation area of ​​the UHVDC through-wall bushing. Simulated dust particles with a particle size distribution of 1 to 50 μm are injected into the SF6 gas insulation area through the electric field-flow field coupling device. At the same time, a space charge sensor array is uniformly arranged on the surface of the epoxy core. The space charge sensor array consists of several capacitive charge detection units. The capacitive charge detection units are embedded in grooves with a depth of 0.5 to 1.5 mm on the surface of the epoxy core. The spacing between adjacent capacitive charge detection units is 8 to 15 mm. Each capacitive charge detection unit is connected to a charge signal amplifier through a shielded coaxial cable.

[0086] The specific implementation of step S02 is as follows: Start the dust simulation injection system and apply a DC voltage. The dust trajectory disturbance wave generator inside the electric field-flow field coupling device outputs sparse dust trajectory marker data through the dust marker signal generation function. The description of the dust marker signal generation function is as follows.

[0087] Mark the number of dust particles The calculation formula is: .

[0088] In the formula, This represents the total number of dust particles, expressed in individual particles. The standard dust quantity benchmark value is set to [value]. indivual; The sparsity coefficient, with a value ranging from 0.05 to 0.15, is a dimensionless parameter. To indicate the number of dust particles, the unit is individual, and the calculation result is rounded up.

[0089] Mark the amplitude of the electric field disturbance of dust particles The calculation formula is: .

[0090] In the formula, The current electric field strength is expressed in kV / m. The standard electric field strength is 500 kV / m. The average particle size of the dust is expressed in μm. The reference particle size is set to 10 μm; This is the disturbance amplitude coefficient, with a value ranging from 0.2 to 0.4, and is a dimensionless parameter. The amplitude of the electric field disturbance caused by dust particles is expressed in kV / m. A high-speed camera system and a synchrotron X-ray tomography device were used to acquire the three-dimensional spatial coordinates of dust particles and the interface charge density distribution in real time. The high-speed camera system consisted of three orthogonally arranged high-speed cameras and a synchronous trigger controller. The optical axes of the three orthogonally arranged high-speed cameras were perpendicular to each other, and their intersection point was located on the central axis of the UHVDC through-wall bushing. The frame rate was set to 5000 to 10000 frames per second, and the exposure time was set to 20 to 50 μs. The synchrotron X-ray tomography device included an X-ray source, a rotating sample stage, a detector array, and an image reconstruction computing unit. The X-ray source emitted a monochromatic X-ray beam with an energy of 15 to 25 keV. The rotating sample stage rotated at an angular velocity of 0.5 to 2 degrees per second to acquire multi-angle projection data. The image reconstruction computing unit used an improved Tikhonov regularized inversion algorithm to reconstruct the three-dimensional charge density distribution from the multi-angle projection data.

[0091] The specific implementation of step S03 is as follows: The collected three-dimensional spatial coordinates of dust particles are input into the charge transport-dust dynamics coupled calculation module. The charge transport-dust dynamics coupled calculation module decomposes the internal calculation domain of the UHVDC through-wall bushing into a charge transport-dominant region and a dust movement-dominant region. The charge density gradient near the surface of the epoxy core is greater than... The region was divided into a charge transport-dominated region and a dust motion-dominated region in which the SF6 gas flow velocity was greater than 0.5 m / s. The charge density gradient field and dust trajectory field were solved using a partitioned asynchronous propagation algorithm. In the charge transport-dominated region, the algorithm employed a fourth-order implicit Runge-Kutta scheme to solve the drift-diffusion equations, with a time step set to [value missing]. to In the dust-dominant region, an explicit frog-jump scheme is used to track the trajectory of dust particles, with the time step set to [value missing]. to The two regions are connected by cubic spline interpolation to achieve loosely coupled transfer of charge density and electric field strength. The partitioned asynchronous advancement algorithm adopts a Warp-level synchronization mechanism, which organizes the 32 threads of the GPU into a thread bundle unit. At the end of each time step, the Warp-level synchronization instruction ensures that all threads in the same thread bundle complete the calculation before entering the next time step, thus obtaining the dust spatial distribution evolution sequence and electric field distortion coefficient at different times.

[0092] electric field distortion coefficient The defining formula is: .

[0093] In the formula, This represents the local electric field strength, expressed in kV / m. The average electric field strength is expressed in kV / m. Let be the electric field distortion coefficient, a dimensionless parameter, when When the value is greater than 0.3, the region is considered to have significant electric field distortion and is marked as a discharge-sensitive region.

[0094] The specific implementation of step S04 is as follows: the acquired partial discharge signal is decomposed into wavelet packets at multiple scales using the Daubechies wavelet basis function. The number of decomposition layers is set to 5 to 7. After each layer of decomposition, coefficients with a frequency range of 50kHz to 2MHz are retained as time-frequency features. The coefficients of each layer are arranged according to the time axis to form a two-dimensional time-frequency feature map. The time-frequency feature map is a two-dimensional matrix image of 128×128 pixels. The time-frequency feature map, sparse dust trajectory marker data, and electric field distortion coefficient are input into the discharge type identification model. The structure of the discharge type identification model is that the input layer receives the 128×128 pixel time-frequency feature map, sparse dust trajectory marker data, and electric field distortion coefficient. Spatial features are extracted through three convolutional layers. Each convolutional layer is followed by a batch normalization layer and an activation layer. Then, the gradient propagation is enhanced through two residual connection blocks. Next, long-interval spatiotemporal features are aggregated through a multi-head adaptive attention layer. Finally, two fully connected layers are connected to output the dust bridging discharge probability distribution and the predicted value of the discharge initiation voltage.

[0095] Number of heads in multi-head adaptive attention layer The calculation formula is: .

[0096] In the formula, The base number of heads is 4, and the unit is the number of heads. The average particle size of the dust is expressed in μm. The reference particle size is set to 10 μm; is the electric field distortion coefficient, a dimensionless parameter; The standard distortion coefficient has a value of 0.3 and is a dimensionless parameter. Let fractal dimension be the dust particle, a dimensionless parameter; The reference fractal dimension is set to 1.8, a dimensionless parameter. This represents the floor function; The number of heads in the multi-head adaptive attention layer is expressed in units, and the calculation result is limited to between 4 and 16.

[0097] The specific implementation of step S05 is as follows: The dust agglomeration morphology data is processed using a parallel computing algorithm for dust fractal aggregates. This algorithm recursively subdivides the bounding box using a quadtree spatial decomposition method. The initial bounding box side length is set to the inner diameter of the UHVDC through-wall bushing. Each recursion divides the bounding box into 8 sub-boxes along the three coordinate axes. The recursion depth is set to 8 to 12 layers. A 64-bit integer is allocated to each sub-box in the GPU shared memory for bitwise operations to determine the box's occupancy status. The parallel computing algorithm for dust fractal aggregates summarizes the data from each layer using Warp reduction technology. The box counting results show that 32 threads within each thread bundle are responsible for counting 32 sub-boxes. Five rounds of reduction operations are used to accumulate the 32 counts into a single total value. Each reduction operation takes less than 10 ns, and the total reduction time increases logarithmically with the recursion depth. The dust fractal dimension time series is obtained by calculating the recursive box dimension. The recursive box dimension is calculated by statistically analyzing the number of occupied boxes at different scales. A scatter plot is drawn with the logarithm of the box side length as the x-axis and the logarithm of the number of occupied boxes as the y-axis. The least squares method is used to fit a straight line, and the negative value of the slope is calculated as the dust fractal dimension. Estimated value, when Exceeding the fractal dimension threshold The dust particle size distribution update process is triggered at certain times, with a fractal dimension threshold. The value ranges from 2.2 to 2.5, and it is a dimensionless parameter. The dust particle size distribution update process includes calculating the agglomeration factor based on the dust fractal dimension. .

[0098] Aggregation factor The calculation formula is: .

[0099] In the formula, Let fractal dimension be the dust particle, a dimensionless parameter; The reference fractal dimension is set to 1.8, a dimensionless parameter. The aggregation factor is a dimensionless parameter.

[0100] Equivalent particle size The calculation formula is: .

[0101] In the formula, The average particle size of the dust is expressed in μm. The aggregation factor is a dimensionless parameter. The updated equivalent particle size is shown in μm. The histogram of the equivalent particle size distribution is recalculated and fitted with a log-normal distribution function. The updated particle size distribution parameters are then passed to the charge transport-dust dynamics coupled calculation module.

[0102] The specific implementation of step S06 is as follows: based on the predicted value of the discharge initiation voltage. Measured value of discharge initiation voltage The deviation is adjusted to adjust the dust injection rate.

[0103] absolute value of deviation The calculation formula is: .

[0104] In the formula, This is the predicted discharge initiation voltage, in kV. The measured value of the discharge initiation voltage is expressed in kV. This is a dimensionless parameter in percentage form, representing the absolute value of the deviation. When When the dust injection rate is greater than 12%, increase the dust injection rate and repeat steps S02 to S05. When the dust concentration is less than 5%, the current operating parameters are recorded as discharge sensitivity benchmark data. The current operating parameters include dust injection rate, average dust particle size, electric field distortion coefficient, and dust fractal dimension. The discharge sensitivity benchmark data includes the measured values ​​of dust injection rate, average dust particle size, electric field distortion coefficient, dust fractal dimension, and discharge initiation voltage. The discharge sensitivity benchmark data is used to establish a discharge sensitivity lookup table under different dust conditions, providing a reference for the insulation performance evaluation and operation and maintenance of UHVDC through-wall bushings.

[0105] It should be noted that the variables involved in this embodiment are explained in detail in Table 1.

[0106] Table 1. Variable Explanation Table

[0107] To better understand and implement this invention, a specific application scenario of the invention is provided below as Example 2: To verify the effectiveness of the invention, technicians built a test environment and conducted a dust intrusion simulation test on a certain type of ultra-high voltage DC through-wall bushing to evaluate its discharge sensitivity characteristics under different dust conditions. The test environment included an ultra-high voltage DC through-wall bushing with a rated voltage of 800kV, a dust simulation injection system, a high-speed camera system, and a partial discharge detection system. Figure 2As shown, the dust storage chamber has a volume of 50L and stores alumina simulated dust particles with a particle size distribution of 1 to 50μm. The pneumatic conveying pipeline uses a stainless steel pipe with an inner diameter of 25mm, and the flow control valve is an electric regulating valve. The electric field-flow field coupling device is installed at the SF6 gas inlet of the UHVDC through-wall bushing. Technicians first uniformly arranged 36 capacitive charge detection units on the surface of the epoxy core, with an adjacent unit spacing of 12mm and an embedding depth of 1.0mm. Each detection unit is connected to the charge signal amplifier via a shielded coaxial cable. The high-speed camera system uses three high-speed cameras with a frame rate of 8000 frames per second, an exposure time set to 35μs, and the camera optical axes are perpendicular to each other with their intersection point located on the central axis of the bushing. The synchrotron radiation X-ray tomography device is equipped with an X-ray source with an energy of 20keV, a sample stage rotation speed of 1 degree per second, and acquires projection data from 180 angles.

[0108] Technicians activated the dust simulation injection system and applied a 600kV DC voltage, setting the initial dust injection rate to 0.8g / min. The dust trajectory disturbance wave generator inside the electric field-flow field coupling device began operation, calculating based on the dust marker signal generation function, when the total number of dust particles reached... With an average dust particle size of 15 μm and a current electric field strength of 520 kV / m, 79 dust particles were marked, and the amplitude of the electric field disturbance caused by the marked dust particles was 156 kV / m. The high-speed camera system and synchrotron X-ray tomography device began real-time data acquisition. Within the first 30 seconds of the test, the amount of three-dimensional spatial coordinate data of the dust particles acquired reached [amount missing]. The interface charge density distribution was reconstructed from several data points using an improved Tikhonov regularized inversion algorithm. Technicians input the collected three-dimensional spatial coordinates of dust particles into the charge transport-dust dynamics coupled calculation module. This module decomposes the computational domain into a charge transport-dominated region and a dust motion-dominated region. The charge density gradient near the epoxy core surface is... The area is divided into a charge transport-dominated region, and the area where the SF6 gas flow velocity is 0.75 m / s is divided into a dust movement-dominated region. For example... Figure 3 As shown, the partitioned asynchronous propagation algorithm uses a fourth-order implicit Runge-Kutta scheme to solve the drift-diffusion equation in the charge transport-dominated region, with the time step set to [value missing]. In the dust-dominant region, an explicit leapfrog pattern was used to track the trajectory of dust particles, with a time step set to [value missing]. The calculated charge density gradient field shows that the gradient value reaches its maximum at the edge of the epoxy core. The dust motion trajectory field shows that dust particles gather towards areas with large electric field intensity gradients under the influence of electric field force and airflow drag force.

[0109] Technicians analyzed the spatial distribution evolution sequence of dust particles over a time period of 0 to 60 seconds and found that dust particles exhibited significant agglomeration in three time periods: 10 to 15 seconds, 30 to 38 seconds, and 45 to 52 seconds. The calculated electric field distortion coefficient reached 0.42 in the dust agglomeration region, exceeding the threshold of 0.3, and these regions were marked as discharge-sensitive areas. The partial discharge signal acquired by the partial discharge detection sensor underwent wavelet packet multi-scale decomposition using the Daubechies wavelet basis function, with a decomposition level of 6 layers, retaining coefficients in the frequency range of 50kHz to 2MHz, forming a 128×128 pixel time-frequency feature map. Technicians input the time-frequency feature map, sparse dust trajectory marker data, and electric field distortion coefficient into a discharge type identification model. After 180 training cycles, the model's total loss function converged to 0.0025. Based on an average dust particle size of 15 μm, an electric field distortion coefficient of 0.42, and a dust fractal dimension of 1.95, the number of heads in the multi-head adaptive attention layer was calculated to be 8. Figure 4 As shown, the dust bridging discharge probability distribution output by the discharge type identification model shows probabilities of 0.73, 0.68 and 0.81 in the three regions at the edge of the epoxy core, respectively, with a predicted discharge initiation voltage of 565kV.

[0110] Technicians used a parallel computing algorithm for dust fractal aggregates to process dust agglomeration morphology data. A quadtree spatial decomposition method was employed to recursively subdivide bounding boxes, with the initial bounding box side length set to a sleeve inner diameter of 580 mm and a recursion depth of 10 layers. Warp reduction techniques were used to summarize the box counts at each level. Each thread bundle contained 32 threads responsible for counting 32 sub-boxes, with a total reduction time of 0.15 ms. The recursive box dimension was calculated by statistically analyzing the number of occupied boxes at different scales. A least-squares method was used to fit a straight line, and the calculated dust fractal dimension time series is shown in Table 2.

[0111] Table 2 Time Series Data of Dust Fractal Dimension

[0112] Technicians discovered that the dust fractal dimensions at 35 seconds, 50 seconds, and 55 seconds were 2.31, 2.38, and 2.15, respectively, all exceeding the fractal dimension threshold of 2.2, triggering the dust particle size distribution update process. Taking the 50-second mark as an example, dividing the dust fractal dimension of 2.38 by the reference fractal dimension of 1.8 and subtracting 1 yields an aggregation factor of 0.32. Multiplying the average dust particle size of 15 μm by 1.32 gives an updated equivalent particle size of 19.8 μm. Technicians recalculated the histogram of the equivalent particle size distribution and fitted it to a log-normal distribution function, then passed the updated particle size distribution parameters to the charge transport-dust dynamics coupled calculation module. The measured discharge initiation voltage from the partial discharge detection sensor was 548 kV, and the calculated absolute deviation was 3.1%, less than the 5% threshold, indicating that the discharge type identification model's prediction accuracy met the requirements. Figure 5 As shown, technicians recorded the current operating parameters as baseline data for discharge sensitivity, including dust injection rate of 0.8 g / min, average dust particle size of 15 μm, electric field distortion coefficient of 0.42, dust fractal dimension of 1.95, and measured discharge initiation voltage of 548 kV.

[0113] Technicians continued to adjust the dust injection rate to 1.2 g / min and repeated the above test procedure. Under the new operating conditions, the average dust particle size was 22 μm, the electric field distortion coefficient reached 0.56, the predicted discharge initiation voltage was 492 kV, the measured value was 512 kV, and the absolute deviation was 3.9%, which also met the accuracy requirements. By comparing the test results under different dust injection rates, technicians established a discharge sensitivity lookup table containing 8 sets of operating parameters, providing a reliable reference for evaluating the insulation performance of this type of UHVDC through-wall bushing under different dust pollution levels. The advancement of this invention compared to traditional methods lies in the fact that traditional methods can only obtain a single discharge initiation voltage value through static voltage rise tests, and cannot reveal the influence mechanism of the dynamic movement process of dust on discharge characteristics. This invention generates sparse marker data through a dust trajectory disturbance wave generator, and combines high-speed imaging and synchrotron radiation imaging technology to track the three-dimensional spatial position of dust particles and the interface charge density distribution in real time, which greatly improves the observability of the dust movement process. The charge transport-dust dynamics coupled calculation module employs a partitioned asynchronous algorithm, setting separate time steps for different timescales of charge transport and dust movement, thus avoiding the computational inefficiency caused by the traditional uniform step size. The discharge type identification model introduces a sparse connection learning framework based on dynamic topology reconstruction and long-interval aggregation. Addressing the intermittent and spatially non-uniform characteristics of dust agglomeration, it adaptively prunes low-mutual-information neural connections, significantly reducing computational complexity while maintaining model expressiveness, making real-time discharge type identification possible. The parallel computation algorithm for dust fractal aggregates utilizes the GPU's Warp-level synchronization mechanism to accelerate recursive box dimension calculation, reducing the computation time of several minutes required by traditional serial algorithms to milliseconds, providing technical support for the dynamic updating of dust particle size distribution. These synergistic technological innovations enable this invention to accurately assess the discharge sensitivity of high-voltage bushings under real-world conditions of dynamic dust intrusion, providing a scientific basis for equipment operation and maintenance decisions.

[0114] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for testing the discharge sensitivity of high-voltage bushings by simulating the effects of dust intrusion, characterized in that, A dust simulation injection system was established inside the UHVDC through-wall bushing to inject simulated dust particles into the SF6 gas-insulated area. Simultaneously, a space charge sensor array was deployed on the epoxy core surface. The dust simulation injection system was activated and a DC voltage was applied. A dust trajectory disturbance wave generator produced disturbances. A high-speed camera system and a synchrotron X-ray tomography device were used to acquire the three-dimensional spatial coordinates of the dust particles and the interface charge density distribution. The three-dimensional spatial coordinates of the dust particles were input into a charge transport-dust dynamics coupled calculation module. Based on a partitioned asynchronous propagation algorithm, the charge density gradient field and the dust trajectory field were solved to obtain the dust spatial distribution. The partial discharge signal is decomposed into wavelet packets at multiple scales using the evolution sequence and electric field distortion coefficient. The time-frequency feature spectrum, sparse dust trajectory marker data, and electric field distortion coefficient are input into the discharge type identification model to output the dust bridging discharge probability distribution and the predicted discharge initiation voltage. The dust agglomeration morphology data is processed using a parallel computing algorithm for dust fractal aggregates. The dust fractal dimension time series is obtained by recursive box dimension calculation. When the dust fractal dimension exceeds the fractal dimension threshold, the dust particle size distribution update process is triggered. The dust injection rate is adjusted according to the deviation between the predicted discharge initiation voltage and the measured discharge initiation voltage, and the discharge sensitivity benchmark data is recorded.

2. The method according to claim 1, characterized in that, The dust simulation injection system includes a dust storage bin, a pneumatic conveying pipeline, a flow control valve, and an electric field-flow field coupling device. The dust storage bin is connected to the dust inlet of the electric field-flow field coupling device through the pneumatic conveying pipeline. The flow control valve is installed in the middle section of the pneumatic conveying pipeline to adjust the dust injection rate.

3. The method according to claim 2, characterized in that, The electric field-flow field coupling device is equipped with a dust trajectory disturbance wave generator, an electric field modulation electrode grid, and a flow field distribution adjustment chamber. The dust trajectory disturbance wave generator applies a periodic disturbance electric field to the simulated dust particles through the electric field modulation electrode grid.

4. The method according to claim 3, characterized in that, The dust trajectory disturbance wave generator outputs sparse dust trajectory marker data through the dust marker signal generation function. The dust marker signal generation function randomly selects some particles in the dust group and applies identifiable electric field disturbance markers.

5. The method according to claim 4, characterized in that, The dust marking signal generation function is expressed as follows: the number of marked dust particles is equal to the total number of dust particles divided by the square root of the standard dust quantity reference value, multiplied by the sparsity coefficient; the amplitude of the perturbation electric field of the marked dust particles is equal to the current electric field strength divided by the standard electric field strength, multiplied by the average dust particle size divided by the reference particle size, and then multiplied by the perturbation amplitude coefficient.

6. The method according to claim 5, characterized in that, The space charge sensor array consists of capacitive charge detection units, which are embedded in grooves on the surface of the epoxy core. The spacing between adjacent capacitive charge detection units is 8 to 15 mm.

7. The method according to claim 6, characterized in that, The synchrotron radiation X-ray tomography device includes an X-ray source, a rotating sample stage, a detector array, and an image reconstruction computing unit. The image reconstruction computing unit uses an improved Tikhonov regularized inversion algorithm to reconstruct the multi-angle projection data into a three-dimensional charge density distribution.

8. The method according to claim 7, characterized in that, The charge transport-dust dynamics coupled calculation module decomposes the internal computational domain of the UHVDC through-wall bushing into a charge transport-dominated region and a dust movement-dominated region, where the charge density gradient is greater than... The region is divided into charge transport-dominated regions.

9. The method according to claim 8, characterized in that, The partitioned asynchronous propagation algorithm uses a fourth-order implicit Runge-Kutta scheme to solve the drift-diffusion equation in the charge transport-dominated region and an explicit frog-leap scheme to track the dust particle trajectory in the dust motion-dominated region. The boundary between the two regions is achieved by cubic spline interpolation to realize the loosely coupled transfer of charge density and electric field intensity.

10. The method according to claim 9, characterized in that, The partitioned asynchronous advancement algorithm uses a Warp-level synchronization mechanism to organize the 32 threads of the graphics processor into a thread bundle unit. At the end of each time step, Warp-level synchronization instructions ensure that all threads within the same thread bundle complete the calculation.

Images