Method for detecting ablation resistance of burning-resistant nozzle and arc contact

Abstract

The invention is suitable for the technical field of performance detection, and provides a method for detecting the ablation resistance of a burning-resistant nozzle and an arc contact, and the method comprises the steps: collecting a temperature field matrix, an arc image sequence, an electrical parameter time sequence and a surface displacement curve of an ablation region, and generating a multi-modal data set; based on the data set, extracting spatial features of the ablation core region through a cavity convolution layer and a channel attention mechanism; inputting a preset bidirectional LSTM network, fusing thermal stress data of the COMSOL simulation model to drive gating weight update, and extracting time sequence evolution characteristics; generating a space-time fusion feature vector containing the temperature field gradient, the arc form change rate and the surface roughness evolution feature; inputting the space-time fusion feature vector, the thermal stress distribution data and the material phase change threshold parameter into a constructed ablation prediction model for training; the ablation prediction model after training is completed is optimized; and a performance prediction result is generated in real time by using the optimized ablation prediction model, so that the detection accuracy is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of performance detection, and in particular to a method for detecting the ablation resistance of a burn-resistant nozzle and an arc contact. Background Art

[0002] In high-voltage switchgear, arcing contacts and arc-extinguishing nozzles are core components responsible for breaking current and controlling arcs. Their erosion resistance directly determines the reliability and service life of the circuit breaker. When the equipment is opening and closing, the high temperature and impact of the arc can cause erosion of the contact material and damage the nozzle structure, leading to risks such as poor contact and interruption failure. With the increase in grid voltage levels and the growing demand for intelligent operation and maintenance, higher requirements are being placed on the performance testing of erosion-resistant components: not only must static erosion resistance be evaluated, but also the dynamic capture of performance evolution under the influence of thermal, electrical, and mechanical coupling during the erosion process.

[0003] At present, the relevant detection methods mainly use drive devices, temperature control devices and detection devices to measure parameters such as the inner diameter and bounce time of the moving arc contact before and after closing offline. However, it can only simulate the static performance under a single temperature condition and cannot capture the dynamic coupling effect of the arc morphology and temperature field during the ablation process in real time; some methods quantify the ablation amount of the arc extinguishing nozzle through weighing and dimension measurement, but rely on manual operation and can only obtain discrete data before and after the test. It is difficult to capture the transient dynamic process of arc erosion, and it is impossible to output remaining life assessment and risk warning. Summary of the Invention

[0004] The present invention provides a method for detecting the ablation resistance of a burn-resistant nozzle and an arc contact, which is used to solve the problem of incomplete performance detection.

[0005] The present invention provides a method for detecting the ablation resistance of a burn-resistant nozzle and an arc contact, comprising:

[0006] Using an infrared thermal imager array, high-speed camera, arc sensor, and laser displacement meter, the temperature field matrix, arc image sequence, electrical parameter time series, and surface displacement curve of the ablation area are synchronously collected to generate a multimodal dataset with time stamp alignment.

[0007] The temperature field matrix and arc image sequence are input into a preset multi-scale residual network, and the spatial features of the ablation core area are extracted through the void convolution layer and channel attention mechanism. The electrical parameter time series and surface displacement curve are input into a preset bidirectional LSTM network, and the thermal stress data of the COMSOL simulation model is integrated to drive the gate weight update and extract the time series evolution characteristics.

[0008] Generate a spatiotemporal fusion feature vector including temperature field gradient, arc morphology change rate and surface roughness evolution characteristics according to the spatial characteristics and temporal evolution characteristics;

[0009] Inputting the spatiotemporal fusion feature vector, thermal stress distribution data generated by the COMSOL simulation model, and material phase change threshold parameters into the constructed ablation prediction model for training;

[0010] Knowledge distillation compression and sliding window calibration are used to optimize the trained ablation prediction model.

[0011] The multimodal data collected in real time is used as input, and the optimized ablation prediction model is used to generate the corresponding ablation risk heat map, remaining life probability curve and critical failure warning signal.

[0012] Furthermore, the temperature field matrix and the arc image sequence are input into a preset multi-scale residual network, and the spatial features of the ablation core area are extracted through a hole convolution layer and a channel attention mechanism, including:

[0013] The preset multi-scale residual network includes a multi-scale dilated convolution module, which is three parallel residual sub-modules, using dilated convolution layers with expansion rates of 2, 4, and 6, respectively, to capture the correlation between the ablation area and the surrounding temperature field at different spatial scales;

[0014] A compression and excitation module is embedded at the output of each residual submodule, and a channel weight vector is generated through global average pooling to enhance the features of ablation-sensitive areas where the temperature gradient exceeds the set threshold.

[0015] The output feature maps of the multi-scale dilated convolution module are concatenated according to the channel dimension and then reduced in dimension through a 1×1 convolution layer to generate a multi-scale fusion feature map.

[0016] Furthermore, the specific steps of outputting the spatial feature vector through the preset multi-scale residual network include:

[0017] Inputting the temperature field matrix and the arc image sequence into the multi-scale dilated convolution module to extract local morphological features;

[0018] Calculate the channel weight vector for the local morphological features output by each residual submodule;

[0019] After the feature dimension is reduced, it is input into the feature pyramid network, and the low-level high-resolution detail features and high-level semantic features are fused through upsampling and skip connections;

[0020] Output the spatial feature vector that represents the temperature distribution pattern and morphology evolution law of the ablation core area.

[0021] Furthermore, the electrical parameter time series and surface displacement curve are input into a preset bidirectional LSTM network, and the thermal stress data of the COMSOL simulation model is integrated to drive the gating weight update and extract the time series evolution characteristics, including:

[0022] In the calculation of the forget gate and input gate of LSTM, the thermal stress data output by the COMSOL simulation model is used as an additional input to dynamically modify the gate unit weights to make the timing modeling conform to the laws of thermodynamic diffusion.

[0023] The forward LSTM is used to process the data in the forward time sequence to capture the cumulative effect of ablation, and the backward LSTM is used to process the data in the reverse time sequence to capture the deformation hysteresis characteristics. The bidirectional hidden states are concatenated to generate a temporal feature vector.

[0024] Furthermore, the specific steps of extracting the time series evolution features include:

[0025] The electrical parameter time series and surface displacement curve are segmented using a sliding window that matches the arcing period;

[0026] Introducing thermal stress gradient constraints in LSTM cell state updates, combining the current thermal stress value and its gradient to adjust the gating calculation and suppress state mutations under abnormal working conditions;

[0027] The hidden states of the forward and reverse LSTMs are fused element by element, and the time series feature vector is output after maximum pooling in the time dimension.

[0028] Furthermore, generating a spatiotemporal fusion feature vector including temperature field gradient, arc morphology change rate and surface roughness evolution features according to the spatial features and temporal evolution features includes:

[0029] The spatial feature vector and the temporal feature vector are concatenated according to the channel dimension and input into the self-attention module to calculate the correlation weight matrix of the spatial and temporal features, and assign the fusion weights of the temperature field gradient, arc morphology change rate and surface roughness evolution features;

[0030] Performing weighted summation on the concatenated features according to the association weight matrix to generate an initial fused feature vector, and mapping it to a unified feature space through a fully connected layer;

[0031] The fused feature vector is layer-normalized and the spatiotemporal fusion feature vector is output after dimensionality reduction.

[0032] Furthermore, the constructed ablation prediction model includes:

[0033] The constructed ablation prediction model includes a physical constraint input layer, a multi-task learning branch, and a joint optimization layer;

[0034] The physical constraint input layer receives the spatiotemporal fusion feature vector, the thermal stress distribution data of the COMSOL simulation model, and the material phase change threshold parameter, and encodes the thermal stress data and the phase change threshold into a physical feature vector;

[0035] The multi-task learning branch includes an ablation rate prediction branch and a critical failure judgment branch. The ablation rate prediction branch is composed of a three-layer fully connected network, which inputs a spatiotemporal fusion feature vector and a physical feature vector and outputs the ablation rate under the current working condition. The critical failure judgment branch is composed of a bidirectional LSTM and a Sigmoid classifier, which inputs a time series evolution feature and a phase change threshold and outputs a critical failure probability:

[0036] The joint optimization layer performs weighted fusion of the outputs of the ablation rate prediction branch and the critical failure determination branch to generate a final ablation depth prediction value.

[0037] Furthermore, the training process of the ablation prediction model includes:

[0038] The spatiotemporal fusion feature vector, thermal stress distribution data generated by the COMSOL simulation model, material phase change threshold parameters, and ablation depth labels measured by the laser displacement meter are used as training inputs.

[0039] A composite loss function is used to calculate the weighted sum of the mean square error of the ablation rate prediction and the cross entropy loss of the critical failure probability, and an L2 regularization term is added to constrain the network weights.

[0040] The thermal stress gradient data of the simulation model is aligned with the characteristic distribution of the network hidden layer through the feature alignment loss function, and the prediction results are consistent with the thermal-electric coupling law;

[0041] In the steady-state phase, the learning rate is reduced to the preset initial value, and the Adam optimizer is used to adaptively adjust the parameter update step size;

[0042] Update network weights in real time and use the Kalman filter to dynamically correct the residual deviation between the predicted value and the measured value;

[0043] The trained ablation prediction model outputs the predicted value of ablation rate, critical failure probability and dynamically corrected ablation depth.

[0044] Furthermore, the method of optimizing the trained ablation prediction model by using knowledge distillation compression and sliding window calibration includes:

[0045] The original ablation prediction model is used as the teacher network to construct a lightweight student network. The soft target distillation loss function is used to constrain the output prediction value of the student network to align with the teacher network, and the number of residual module channels is reduced and redundant network layers are removed.

[0046] The latest ablation data collected in real time is cached at preset time intervals, input into the student network for incremental fine-tuning, and the Kalman filter is used to dynamically correct the deviation between the predicted value and the measured value;

[0047] The optimized student network is converted into a format compatible with embedded devices and deployed to industrial edge computing nodes to support real-time inference and data feedback.

[0048] Furthermore, the multimodal data collected in real time is used as input, and the optimized ablation prediction model is used to generate the corresponding ablation risk heat map, remaining life probability curve and critical failure warning signal, including:

[0049] Reconstruct the thermal distribution of the nozzle or contact surface based on the spatiotemporal fusion feature vector to generate a high-resolution ablation risk thermal map, marking the risk level and weak area coordinates;

[0050] Combining the ablation rate prediction value with the material phase change threshold, a Monte Carlo simulation is used to output the remaining life probability distribution curve, including the confidence interval.

[0051] When the critical failure probability or the remaining life prediction value exceeds the preset safety threshold, a graded warning signal is triggered and pushed to the monitoring terminal.

[0052] It can be seen from the above technical solutions that the present invention has the following advantages:

[0053] The present invention inputs the temperature field matrix and arc image sequence generated in the multimodal dataset into a preset multi-scale residual network, and extracts the spatial features of the ablation core area through the void convolution layer and channel attention mechanism; inputs the electrical parameter time series and the surface displacement curve into a preset bidirectional LSTM network, integrates the thermal stress data of the COMSOL simulation model to drive the gate weight update, and extracts the time series evolution characteristics; constructs a spatiotemporal coupling feature vector by integrating the temperature field gradient, arc morphology change rate and surface roughness evolution characteristics, and accurately characterizes the nonlinear evolution law under the interaction of electrothermal and mechanical multi-fields during the ablation process; inputs the spatiotemporal fusion characteristics, COMSOL thermal stress distribution and material phase change threshold into the ablation prediction model to achieve a deep integration of data drive and physical mechanism; compresses the complex deep network into a micro network that can be run by edge devices, while maintaining detection accuracy, reducing computing power consumption, and adapting to the distributed online monitoring needs of high-voltage switchgear; based on the optimized model, generates ablation risk heat map, remaining life probability curve and critical failure warning signal in real time. The present invention realizes online monitoring of high-voltage switchgear throughout its life cycle through multimodal dynamic perception, spatiotemporal feature fusion, physical constraint modeling, and lightweight intelligent deployment, significantly improving the operational reliability and maintenance efficiency of key components of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 The figure is a flow chart of an embodiment of a method for detecting the ablation resistance of a burn-resistant nozzle and an arc contact in the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0056] The method for detecting the ablation resistance of the burn-resistant nozzle and arc contact in this embodiment is used to improve the timeliness and accuracy of performance detection. The implementation method in this embodiment can be implemented in the system, on the server, or on the terminal, without specific limitation.

[0057] Example 1

[0058] See also Figure 1 In the present invention, a method for detecting the ablation resistance of a burn-resistant nozzle and an arc contact includes the following steps:

[0059] S11. Using an infrared thermal imager array, high-speed camera, arc sensor, and laser displacement meter, synchronously collect the temperature field matrix, arc image sequence, electrical parameter time series, and surface displacement curve of the ablation area to generate a timestamp-aligned multimodal dataset.

[0060] Install the infrared thermal imager array around the nozzle / contact, calibrate the spatial coordinate system and establish a three-dimensional temperature field mapping relationship; adjust the viewing angle and focal length of the high-speed camera to ensure that the arc morphology completely covers the field of view; calibrate the range and phase delay of the arc sensor, and align the laser displacement meter's spot at the center of the ablation area.

[0061] A trigger signal is sent through the synchronization module to simultaneously start all sensor acquisitions; the arcing process is monitored in real time, and the following data is collected: 1. Infrared thermal imager: temperature field matrix (dimensions: X×Y×T, X / Y are spatial coordinates, T is time series); 2. High-speed camera: arc image sequence (RGB or grayscale format); 3. Arc sensor: current / voltage waveform (time-amplitude series); 4. Laser displacement meter: surface displacement curve (time-displacement series).

[0062] Gaussian filtering is applied to the temperature field matrix to eliminate interference from ambient thermal radiation. Inter-frame registration and motion compensation are performed on high-speed camera images to eliminate image jitter caused by mechanical vibration. A sliding average filter is applied to current and voltage data to suppress high-frequency noise. Baseline correction is performed on laser displacement data to eliminate zero-point drift errors. All sensor data streams are aligned according to a unified timestamp to generate a structured multimodal dataset, which is stored in HDF5 or NetCDF format.

[0063] This step uses multi-sensor collaborative acquisition and time synchronization technology to obtain thermal, electrical, and mechanical multimodal data during the ablation process, constructing a high-precision dataset aligned in time and space, and providing a basis for subsequent feature extraction and model training.

[0064] S12. Input the temperature field matrix and arc image sequence into a preset multi-scale residual network, extracting the spatial features of the ablation core area through a dilated convolutional layer and a channel attention mechanism. Input the electrical parameter time series and surface displacement curve into a preset bidirectional LSTM network, integrating the thermal stress data of the COMSOL simulation model to drive the gating weight update and extract the temporal evolution characteristics.

[0065] It should be noted that the COMSOL simulation model is COMSOL Multiphysics, a multi-physics coupled simulation software based on the finite element method. It supports modeling and numerical solutions for various physical phenomena in complex engineering systems, including heat, electricity, force, and fluid flow. In the scenario of testing the ablation resistance of burn-resistant nozzles and arc contacts, the COMSOL model is used to simulate the multi-physics coupled behavior of the ablation process, including the temperature field distribution, thermal stress evolution, and material phase change effects under the action of the arc heat source, providing physical law constraints for deep learning models.

[0066] S121. Input the temperature field matrix and arc image sequence into the preset multi-scale residual network, and extract the spatial features of the ablation core area through the void convolution layer and channel attention mechanism, including:

[0067] This step uses a multi-scale dilated convolution module and a channel attention mechanism to adaptively enhance the spatial characteristics of the ablation core area, suppress background noise interference, and improve sensitivity to temperature gradients and morphology changes. The construction of the multi-scale dilated convolution module includes:

[0068] The preset multi-scale residual network includes a multi-scale dilated convolution module, which consists of three parallel residual submodules. The dilated convolution layers use dilated convolution layers with expansion rates of 2, 4, and 6, respectively, to capture the correlation between the ablation area and the surrounding temperature field at different spatial scales.

[0069] Residual submodule 1: dilated convolution layer with dilation rate = 2 (convolution kernel 3×3, number of channels 64) + BatchNorm + ReLU; residual submodule 2: dilated convolution layer with dilation rate = 4 (convolution kernel 3×3, number of channels 64) + BatchNorm + ReLU; residual submodule 3: dilated convolution layer with dilation rate = 6 (convolution kernel 3×3, number of channels 64) + BatchNorm + ReLU; the three submodules process input data in parallel to expand the receptive field to cover different spatial scales.

[0070] A compression and excitation module is embedded at the output of each residual submodule, and a channel weight vector is generated through global average pooling to enhance the features of ablation-sensitive areas where the temperature gradient exceeds the set threshold.

[0071] Global average pooling (GAP) is performed on the output feature map of each residual submodule to generate a channel description vector. The channel weight vector is calculated through compression and excitation of the fully connected layer (FC). The weight vector is multiplied with the original feature map channel by channel to enhance the response of sensitive areas with temperature gradients ≥100℃ / mm.

[0072] The output feature maps of the multi-scale dilated convolution module are concatenated according to the channel dimension and then reduced in dimension through a 1×1 convolution layer to generate a multi-scale fusion feature map.

[0073] The weighted feature maps output by the three residual submodules are concatenated according to the channel dimension (number of channels = 64 × 3 = 192); the dimension is reduced to 64 channels through a 1 × 1 convolutional layer to generate a multi-scale fusion feature map; the output feature map resolution is consistent with the input.

[0074] S122. The specific steps of outputting the spatial feature vector through the preset multi-scale residual network include:

[0075] 1. Input the temperature field matrix and arc image sequence into the multi-scale dilated convolution module to extract local morphological features;

[0076] The temperature field matrix and the arc image are spliced ​​channel by channel into a 4-channel input and input into the multi-scale dilated convolution module to extract the local morphological features corresponding to the expansion rates of 2 / 4 / 6, and output three sets of feature maps.

[0077] 2. Calculate the channel weight vector for the local morphological features output by each residual submodule;

[0078] Apply the SE module to each set of feature maps to generate a 64-dimensional channel weight vector; perform channel weighting on the feature maps according to the weight vector.

[0079] 3. After reducing the dimensionality of the features, they are input into the feature pyramid network, which fuses low-level high-resolution detail features and high-level semantic features through upsampling and skip connections;

[0080] Extract high-resolution details from the multi-scale fusion feature map (256×256×64); generate 32×32×256 high-level semantic features through 3 downsampling (stride=2 convolution); upsample the high-level features to 256×256 and add them element-wise with the low-level features to fuse global semantics with local details.

[0081] 4. Output the spatial feature vector that represents the temperature distribution pattern and morphology evolution law of the ablation core area.

[0082] The fused feature map (256×256×64) is subjected to global maximum pooling (GMP) to generate a 64-dimensional vector; it is then mapped to a 256-dimensional spatial feature vector through a fully connected layer.

[0083] The above steps ensure the feasibility and innovation of the technical solution through modular design and quantitative parameter limitation, and provide high-precision and high-robust spatial feature extraction capabilities for ablation resistance detection.

[0084] S13. Generate a spatiotemporal fusion feature vector including temperature field gradient, arc morphology change rate, and surface roughness evolution characteristics based on spatial characteristics and temporal evolution characteristics;

[0085] 1. The spatial feature vector and the temporal feature vector are concatenated according to the channel dimension and input into the self-attention module to calculate the correlation weight matrix of the spatial and temporal features and assign the fusion weights of the temperature field gradient, arc morphology change rate, and surface roughness evolution features;

[0086] 2. Perform weighted summation on the concatenated features according to the associated weight matrix to generate the initial fused feature vector, which is then mapped to a unified feature space through a fully connected layer.

[0087] 3. Perform layer normalization on the fused feature vector and output the spatiotemporal fusion feature vector after dimensionality reduction.

[0088] First, the 256-dimensional spatial feature vector extracted by the multi-scale residual network and the 128-dimensional temporal features output by the bidirectional LSTM are concatenated into a 384-dimensional vector. This vector is then fed into a self-attention module to calculate the associated weight matrix, dynamically enhancing the weights of highly sensitive features such as temperature field gradients and arc morphology mutations. The weighted features are then mapped to a 128-dimensional unified semantic space using a fully connected layer to eliminate scale differences across modal data. Finally, layer normalization is used to stabilize the feature distribution, and principal component analysis (PCA) is used to reduce the dimensionality to 64 dimensions, generating a lightweight spatiotemporal fusion feature vector. This process combines attention-driven dynamic fusion with data compression optimization to significantly improve feature discrimination and model convergence speed. It also meets the real-time processing requirements of industrial edge devices (inference latency ≤ 20ms), providing highly robust input for ablation prediction.

[0089] S14. Input the spatiotemporal fusion feature vector, thermal stress distribution data generated by the COMSOL simulation model, and material phase change threshold parameters into the constructed ablation prediction model for training;

[0090] Specifically, the constructed ablation prediction model includes the following:

[0091] The constructed ablation prediction model includes a physical constraint input layer, a multi-task learning branch, and a joint optimization layer;

[0092] 1. The physical constraint input layer receives the spatiotemporal fusion feature vector, the thermal stress distribution data of the COMSOL simulation model, and the material phase change threshold parameter, and encodes the thermal stress data and phase change threshold into a physical feature vector;

[0093] The COMSOL simulation model is a numerical simulation tool based on multi-physics coupling (thermal, electrical, and mechanical). It simulates the thermal stress distribution, arc diffusion, and material phase change behavior during the ablation process by solving partial differential equations, outputting physical quantities such as the thermal stress field and temperature gradient. The material phase change threshold parameter represents the critical condition for a material to undergo a phase transition (such as melting or vaporization), such as the melting point of copper-tungsten alloy and the thermal conductivity threshold.

[0094] Input is a spatiotemporal fusion feature vector: thermal stress distribution data generated by COMSOL simulation, such as matrix data with dimensions matching the temperature field, and material phase transition threshold parameters, such as scalars or vectors, such as melting point and thermal expansion coefficient. The encoding process includes: global average pooling of the thermal stress matrix to generate a 32-dimensional vector; normalization of the threshold parameters and concatenation to form a 16-dimensional vector; and concatenation of the thermal stress vector (32-dimensional) with the phase transition threshold vector (16-dimensional) to output a 48-dimensional physical feature vector.

[0095] 2. The multi-task learning branch includes an ablation rate prediction branch and a critical failure determination branch. The ablation rate prediction branch consists of a three-layer fully connected network, which inputs a spatiotemporal fusion feature vector and a physical feature vector and outputs the ablation rate under the current working condition. The critical failure determination branch consists of a bidirectional LSTM and a Sigmoid classifier, which inputs time series evolution features and a phase transition threshold and outputs the critical failure probability:

[0096] Here, the ablation rate prediction branch is a regression task, which is used to predict the ablation depth of the material per unit time. The network architecture is a three-layer fully connected network (64→48→32→1), and the activation function is ReLU. The spatiotemporal fusion feature vector (64) and the physical feature vector (48) are concatenated into a 112-dimensional input. The output is the ablation rate scalar value.

[0097] The critical failure judgment branch is a classification task, used to determine whether the current ablation state is close to critical failure. The network architecture is a bidirectional LSTM (64 hidden units) + Sigmoid classifier. The time series evolution features (128 dimensions) and the phase change threshold (16 dimensions) are spliced ​​into a 144-dimensional input. The output is the critical failure probability.

[0098] 3. The joint optimization layer performs weighted fusion of the outputs of the ablation rate prediction branch and the critical failure determination branch to generate the final ablation depth prediction value.

[0099] The weight is set according to the real-time current level (10-100kA), with emphasis on failure probability monitoring under high current and rate prediction under low current. The final output prediction value is the sum of the rate and the corresponding weight and the failure and the corresponding weight, and the output result dynamically adapts to different working conditions.

[0100] Specifically, the training process of the ablation prediction model includes:

[0101] 1. Use the spatiotemporal fusion feature vector, thermal stress distribution data generated by the COMSOL simulation model, material phase transition threshold parameters, and ablation depth labels measured by the laser displacement meter as training input;

[0102] The training set includes spatiotemporal fusion feature vectors, COMSOL thermal stress distribution data, material phase change threshold and ablation depth measured by laser displacement meter.

[0103] 2. Using a composite loss function, calculate the weighted sum of the mean square error of the ablation rate prediction and the cross entropy loss of the critical failure probability, and add an L2 regularization term to constrain the network weights;

[0104] Calculate the error between the predicted and measured ablation rate:

[0105]

[0106] Calculate the classification error for the critical failure probability:

[0107]

[0108] Constrain network weights to prevent overfitting:

[0109]

[0110] Total loss L 总 :

[0111] L 总 =0.7·L 速率 +0.3·L 失效 +0.01·L 正则

[0112] 3. Using the feature alignment loss function, the thermal stress gradient data of the simulation model is aligned with the feature distribution of the network hidden layer, and the prediction results are consistent with the thermal-electric coupling law;

[0113] The distribution of network hidden layer features is consistent with the distribution of COMSOL thermal stress gradient data; KL divergence constraint:

[0114] L align =D KL (f 网络 ||f COMSOL )

[0115] Where: f 网络 is the hidden layer feature, f COMSOL is the encoding vector for the simulated thermal stress gradient. align Assume the total loss function with a weight of 0.1.

[0116] 4. In the steady-state phase, the learning rate is reduced to the preset initial value, and the Adam optimizer is used to adaptively adjust the parameter update step size;

[0117] First, the algorithm is divided into stages: the initial ablation stage (0-5ms): learning rate = 1e-3; the steady-state ablation stage (5-15ms): learning rate = 1e-4; and the critical failure stage (>15ms): learning rate = 5e-5. The parameter update step size is adaptively adjusted, with momentum parameters β1 = 0.9 and β2 = 0.999.

[0118] 5. Update network weights in real time and use the Kalman filter to dynamically correct the residual deviation between the predicted value and the measured value;

[0119] Cache the latest 200 sets of real-time data; trigger incremental fine-tuning every 5 minutes to update network weights. Kalman filter correction: y 校准 =K·y pred +(1―K)·y 实测 , the Kalman gain K is dynamically calculated based on the prediction error covariance.

[0120] 6. The trained ablation prediction model outputs the predicted ablation rate, critical failure probability, and dynamically corrected ablation depth.

[0121] Finally, the predicted value of ablation rate, critical failure probability and dynamically corrected ablation depth are output.

[0122] The above steps address the extreme ablation characteristics of burn-resistant nozzles and arcing contacts under high-voltage, high-current conditions by constructing a multi-task ablation prediction model driven by physics and data, enabling precise monitoring. The model utilizes COMSOL multiphysics simulation data and material phase transition thresholds as physical constraints. The physical constraint input layer encodes thermodynamic laws into feature vectors, guiding the collaborative optimization of multiple task branches. The ablation rate prediction branch combines temperature field gradients and arc morphology features to output real-time ablation rates. The critical failure branch integrates temporal electrical parameters and surface deformation data to predict failure probability. The joint optimization layer dynamically weights the output based on current levels, adapting it to the monitoring requirements of 10-100kA operating conditions. During training, a feature alignment loss is used to force the network's hidden layer distribution to match the simulated thermal stress gradient. Dynamic learning rates and online Kalman calibration are combined to address model drift under high-temperature arc interference. The final model, compressed through knowledge distillation, is deployed in an embedded PLC. This achieves industrial-grade, real-time monitoring of nozzle throat ablation depth with an error of ≤5% and a failure warning response of ≤30ms, meeting the life prediction and risk management requirements of high-voltage switchgear under extreme operating conditions.

[0123] S15. Optimize the trained ablation prediction model using knowledge distillation compression and sliding window calibration.

[0124] 1. Using the original ablation prediction model as the teacher network, a lightweight student network is constructed. A soft target distillation loss function is used to constrain the output prediction values ​​of the student network to align with the teacher network. The number of channels in the residual module is reduced and redundant network layers are removed.

[0125] 2. Cache the latest ablation data collected in real time at preset time intervals, input it into the student network for incremental fine-tuning, and use the Kalman filter to dynamically correct the deviation between the predicted value and the measured value;

[0126] 3. Convert the optimized student network into a format compatible with embedded devices and deploy it to industrial edge computing nodes to support real-time inference and data feedback.

[0127] Specifically, the output of the original ablation prediction model (teacher network) is used as a soft target to construct a lightweight student network (the number of residual module channels is reduced by 50%, and redundant bidirectional LSTM layers are removed). The soft target distillation loss is used to constrain the student network prediction value to be aligned with the teacher network. While ensuring accuracy, the model parameters are compressed by 70% (reduced to ≤1MB) to adapt to the computing power limitations of edge devices. Secondly, the nozzle throat temperature field, contact electrical parameters and surface displacement data collected in real time are cached at 5-minute intervals (window length N = 200 groups), and input into the student network for incremental fine-tuning, and The Kalman filter is used to dynamically correct the residual deviation between the predicted ablation depth and the actual measured value of the laser displacement meter (error fluctuation is reduced by ≥40%) to solve the model drift problem caused by material aging and environmental disturbances; finally, the optimized student network is converted into ONNX format and deployed to the embedded PLC device, supporting real-time inference of nozzle / contact ablation risk within 30ms under a short-circuit current of 100kA, and interacting with the substation monitoring system in real time through the OPCUA protocol to trigger protection mechanisms (such as forced tripping) to ensure the safe operation of high-voltage switchgear under extreme arc shocks.

[0128] S16. Use the multimodal data collected in real time as input and use the optimized ablation prediction model to generate the corresponding ablation risk heat map, remaining life probability curve and critical failure warning signal.

[0129] 1. Reconstruct the thermal distribution of the nozzle or contact surface based on the spatiotemporal fusion feature vector, generate a high-resolution ablation risk thermal map, and mark the risk level and weak area coordinates;

[0130] 2. Combining the ablation rate prediction value with the material phase transition threshold, a Monte Carlo simulation is used to output the remaining life probability distribution curve, including the confidence interval;

[0131] 3. When the critical failure probability or remaining life prediction value exceeds the preset safety threshold, a graded warning signal is triggered and pushed to the monitoring terminal.

[0132] Specifically, in response to the ablation detection needs of burn-resistant nozzles and arc contacts under high voltage and high current conditions, first, the real-time spatiotemporal fusion feature vector is input into the deconvolution network to reconstruct the thermal distribution of the nozzle throat and the contact finger surface, and generate a 512×512 pixel high-resolution ablation risk thermal map. The risk areas such as melting depressions and arc burn points are marked by temperature gradient thresholds (≥100℃ / mm) and deformation rates, and their three-dimensional coordinates are output; secondly, combined with the material phase change threshold, such as the melting point of copper-tungsten alloy ≥3400℃, and the real-time ablation rate prediction value, Monte Carlo simulation is used to perform 1000 random samplings to calculate the different current levels (1 The remaining life probability distribution under the combination of 0-100kA) and arcing duration (1-20ms) is output with a 90% confidence interval curve to quantify the life uncertainty under extreme working conditions; finally, a dynamic safety threshold is set (critical failure probability ≥ 85% or remaining life ≤ 10 operating cycles). When the local temperature rise on the contact surface exceeds the limit or the deformation rate of the nozzle throat suddenly increases, a three-level early warning signal (early warning / alarm / emergency shutdown) is triggered and pushed to the monitoring terminal in real time via the Modbus protocol. The equipment protection mechanism (such as forced tripping) is simultaneously activated to ensure the safe operation of high-voltage switchgear in high-risk scenarios such as arc reignition and material welding.

[0133] The above embodiment effectively improves the efficiency and accuracy of the ablation resistance detection of the burn-resistant nozzle and the arc contact by realizing online monitoring of the entire life cycle of the high-voltage switchgear.

[0134] Those skilled in the art will appreciate that the units of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition of each example has been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0135] In the embodiments provided herein, it should be understood that the division of units is merely a logical functional division. In actual implementation, other division methods may be employed, such as combining multiple units into one unit, splitting a unit into multiple units, or ignoring certain features. Furthermore, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically as a separate unit, or two or more units may be integrated into a single unit. These integrated units may be implemented in either hardware or software functional units.

[0136] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-0nly Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk, etc., various media that can store program code.

[0137] It can be understood that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and description of the present invention.

Claims

1. A method for detecting the ablation resistance of a burn-resistant nozzle and an arc contact, characterized in that: include: Using an infrared thermal imager array, high-speed camera, arc sensor, and laser displacement meter, the temperature field matrix, arc image sequence, electrical parameter time series, and surface displacement curve of the ablation area are synchronously collected to generate a multimodal dataset with time stamp alignment. The temperature field matrix and arc image sequence are input into the preset multi-scale residual network, and the spatial features of the ablation core area are extracted through the void convolution layer and channel attention mechanism; The electrical parameter time series and surface displacement curve are input into the preset bidirectional LSTM network, and the thermal stress data of the COMSOL simulation model is integrated to drive the gating weight update and extract the time series evolution characteristics; Generate a spatiotemporal fusion feature vector including temperature field gradient, arc morphology change rate and surface roughness evolution characteristics according to the spatial characteristics and temporal evolution characteristics; Inputting the spatiotemporal fusion feature vector, thermal stress distribution data generated by the COMSOL simulation model, and material phase change threshold parameters into the constructed ablation prediction model for training; Knowledge distillation compression and sliding window calibration are used to optimize the trained ablation prediction model. The multimodal data collected in real time is used as input, and the optimized ablation prediction model is used to generate the corresponding ablation risk heat map, remaining life probability curve and critical failure warning signal.

2. The method for detecting the ablation resistance of the burn-resistant nozzle and arc contact according to claim 1, characterized in that: The temperature field matrix and the arc image sequence are input into a preset multi-scale residual network, and the spatial features of the ablation core area are extracted through the hole convolution layer and the channel attention mechanism, including: The preset multi-scale residual network includes a multi-scale dilated convolution module, which is three parallel residual sub-modules, using dilated convolution layers with expansion rates of 2, 4, and 6, respectively, to capture the correlation between the ablation area and the surrounding temperature field at different spatial scales; A compression and excitation module is embedded at the output of each residual submodule, and a channel weight vector is generated through global average pooling to enhance the features of ablation-sensitive areas where the temperature gradient exceeds the set threshold. The output feature maps of the multi-scale dilated convolution module are concatenated according to the channel dimension and then reduced in dimension through a 1×1 convolution layer to generate a multi-scale fusion feature map.

3. The method for detecting the ablation resistance of the burn-resistant nozzle and arc contact according to claim 2, characterized in that: The specific steps of outputting the spatial feature vector through the preset multi-scale residual network include: Inputting the temperature field matrix and the arc image sequence into the multi-scale dilated convolution module to extract local morphological features; Calculate the channel weight vector for the local morphological features output by each residual submodule; After the feature dimension is reduced, it is input into the feature pyramid network, and the low-level high-resolution detail features and high-level semantic features are fused through upsampling and skip connections; Output the spatial feature vector that represents the temperature distribution pattern and morphology evolution law of the ablation core area.

4. The method for detecting the ablation resistance of the burn-resistant nozzle and arc contact according to claim 1, characterized in that: The electrical parameter time series and surface displacement curve are input into a preset bidirectional LSTM network, the thermal stress data of the COMSOL simulation model is integrated to drive the gating weight update, and the time series evolution characteristics are extracted, including: In the calculation of the forget gate and input gate of LSTM, the thermal stress data output by the COMSOL simulation model is used as an additional input to dynamically modify the gate unit weights to make the timing modeling conform to the laws of thermodynamic diffusion. The forward LSTM is used to process the data in the forward time sequence to capture the cumulative effect of ablation, and the backward LSTM is used to process the data in the reverse time sequence to capture the deformation hysteresis characteristics. The bidirectional hidden states are concatenated to generate a temporal feature vector.

5. The method for detecting the ablation resistance of the burn-resistant nozzle and arc contact according to claim 4, characterized in that: The specific steps of extracting the time series evolution features include: The electrical parameter time series and surface displacement curve are segmented using a sliding window that matches the arcing period; Introducing thermal stress gradient constraints in LSTM cell state updates, combining the current thermal stress value and its gradient to adjust the gating calculation and suppress state mutations under abnormal working conditions; The hidden states of the forward and reverse LSTMs are fused element by element, and the time series feature vector is output after maximum pooling in the time dimension.

6. The method for detecting the ablation resistance of the burn-resistant nozzle and arc contact according to any one of claims 1 to 5, characterized in that: The generating of a spatiotemporal fusion feature vector including temperature field gradient, arc morphology change rate and surface roughness evolution features according to the spatial features and temporal evolution features includes: The spatial feature vector and the temporal feature vector are concatenated according to the channel dimension and input into the self-attention module to calculate the correlation weight matrix of the spatial and temporal features, and assign the fusion weights of the temperature field gradient, arc morphology change rate and surface roughness evolution features; Performing weighted summation on the concatenated features according to the association weight matrix to generate an initial fused feature vector, and mapping it to a unified feature space through a fully connected layer; The fused feature vector is layer-normalized and the spatiotemporal fusion feature vector is output after dimensionality reduction.

7. The method for detecting the ablation resistance of the burn-resistant nozzle and arc contact according to claim 1, characterized in that: The constructed ablation prediction model includes: The constructed ablation prediction model includes a physical constraint input layer, a multi-task learning branch, and a joint optimization layer; The physical constraint input layer receives the spatiotemporal fusion feature vector, the thermal stress distribution data of the COMSOL simulation model, and the material phase change threshold parameter, and encodes the thermal stress data and the phase change threshold into a physical feature vector; The multi-task learning branch includes an ablation rate prediction branch and a critical failure judgment branch. The ablation rate prediction branch is composed of a three-layer fully connected network, which inputs a spatiotemporal fusion feature vector and a physical feature vector and outputs the ablation rate under the current working condition. The critical failure judgment branch is composed of a bidirectional LSTM and a Sigmoid classifier, which inputs a time series evolution feature and a phase change threshold and outputs a critical failure probability: The joint optimization layer performs weighted fusion of the outputs of the ablation rate prediction branch and the critical failure determination branch to generate a final ablation depth prediction value.

8. The method for detecting the ablation resistance of the burn-resistant nozzle and arc contact according to claim 7, characterized in that: The training process of the ablation prediction model includes: The spatiotemporal fusion feature vector, thermal stress distribution data generated by the COMSOL simulation model, material phase change threshold parameters, and ablation depth labels measured by the laser displacement meter are used as training inputs. A composite loss function is used to calculate the weighted sum of the mean square error of the ablation rate prediction and the cross entropy loss of the critical failure probability, and an L2 regularization term is added to constrain the network weights. The thermal stress gradient data of the simulation model is aligned with the characteristic distribution of the network hidden layer through the feature alignment loss function, and the prediction results are consistent with the thermal-electric coupling law; In the steady-state phase, the learning rate is reduced to the preset initial value, and the Adam optimizer is used to adaptively adjust the parameter update step size; Update network weights in real time and use the Kalman filter to dynamically correct the residual deviation between the predicted value and the measured value; The trained ablation prediction model outputs the predicted value of ablation rate, critical failure probability and dynamically corrected ablation depth.

9. The method for detecting the ablation resistance of the burn-resistant nozzle and arc contact according to claim 1, characterized in that: The method of optimizing the trained ablation prediction model by using knowledge distillation compression and sliding window calibration includes: The original ablation prediction model is used as the teacher network to construct a lightweight student network. The soft target distillation loss function is used to constrain the output prediction value of the student network to align with the teacher network, and the number of residual module channels is reduced and redundant network layers are removed. The latest ablation data collected in real time is cached at preset time intervals, input into the student network for incremental fine-tuning, and the Kalman filter is used to dynamically correct the deviation between the predicted value and the measured value; The optimized student network is converted into a format compatible with embedded devices and deployed to industrial edge computing nodes to support real-time inference and data feedback.

10. The method for detecting the ablation resistance of the burn-resistant nozzle and arc contact according to claim 1, characterized in that: The method uses the real-time collected multimodal data as input and uses the optimized ablation prediction model to generate the corresponding ablation risk heat map, remaining life probability curve and critical failure warning signal, including: Reconstruct the thermal distribution of the nozzle or contact surface based on the spatiotemporal fusion feature vector to generate a high-resolution ablation risk thermal map, marking the risk level and weak area coordinates; Combining the ablation rate prediction value with the material phase change threshold, a Monte Carlo simulation is performed to output the remaining life probability distribution curve, including the confidence interval. When the critical failure probability or the remaining life prediction value exceeds the preset safety threshold, a graded warning signal is triggered and pushed to the monitoring terminal.

Images