A GIS partial discharge type identification method

Abstract

The application provides a GIS partial discharge type identification method, and relates to the field of electric variable measurement of power transmission and transformation equipment. The method collects PRPD pattern data containing four types of discharge signals through a sensor, constructs an invalid pattern feature library, dynamically adjusts and filters a threshold value by using a random forest algorithm, and removes invalid data by manual fine screening; processes effective patterns to generate gray value and 3D correlation coefficient dual-channel feature data; constructs a lightweight residual network embedded with a channel attention module, a dynamic channel pruning mechanism and a working condition adaptive subnetwork, introduces transfer learning and a dynamic weighted cross-entropy loss function for optimization training; based on a three-class data index system, parameters are dynamically adjusted by a PID algorithm to realize closed-loop optimization. The problems of inaccurate invalid data removal, loss of spatial correlation features, degradation of deep network and poor working condition adaptability are solved, the accuracy, real-time performance and reliability of GIS partial discharge type identification are improved, and the stable operation of the power system is ensured.

Description

Technical Field

[0001] This invention relates to the field of electrical variable measurement technology for power transmission and transformation equipment, and in particular to a method for identifying partial discharge types in GIS (Gas Transmission and Transformation Equipment). Background Technology

[0002] Gas-insulated switchgear (GIS) has become a core component of power systems due to its compact structure, reliable operation, and small footprint. However, insulation defects are easily introduced into GIS during manufacturing, transportation, and installation. These defects can induce partial discharges, damage insulation performance, and lead to equipment breakdown accidents, seriously threatening power grid safety. Therefore, accurate identification of partial discharge types in GIS is crucial for ensuring the stable operation of power systems.

[0003] Currently, GIS partial discharge identification mainly relies on Phase Resolved Partial Discharge (PRPD) analysis and deep learning techniques. However, existing technologies contain invalid data such as field verification signals, interference signals, and incomplete signals in PRPD maps. Existing methods often employ single screening rules, which are inefficient and prone to misjudgment, leading to a decrease in the accuracy of subsequent model training. PRPD maps contain three-dimensional information about the phase, amplitude, and number of discharge pulses. Existing technologies often convert this information into a single grayscale image input model, resulting in the loss of spatial correlation features and insufficient accuracy in identifying complex signals. Deep convolutional neural networks exhibit degradation as the number of network layers increases, and they consume large amounts of computational resources, making it difficult to meet the needs of real-time field monitoring. Furthermore, data processing, model training, and operational condition adaptation are independent of each other, and static parameter settings cannot dynamically adapt to changes in the field environment.

[0004] Existing patents, such as the Chinese invention "A Clustering and Recognition Method for Partial Discharge UHF Signals in GIS Based on Time and Frequency Domains" (Publication No.: CN120316623B), employ dual-channel encoding in both the time and frequency domains. However, they rely on the VGG16 model, limiting their generalization ability and failing to address efficient cleaning of invalid data. Therefore, a technical solution that can collaboratively solve the aforementioned complex problems is urgently needed. Summary of the Invention

[0005] The main objective of this invention is to provide a method for identifying partial discharge types in GIS, which solves the problems of accurate removal of invalid data, complex signal analysis, stability of deep network training, and multi-objective collaborative optimization of defect identification.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for identifying partial discharge types in GIS, comprising the following steps:

[0007] S1. Collect PRPD map data uploaded by GIS substation. The PRPD map data includes discharge signals corresponding to floating discharge, tip discharge, particle discharge, and internal discharge. S2. Construct an invalid map feature library; S3. The random forest algorithm is used to learn the cleaning rules and dynamically adjust the screening threshold to automatically screen the PRPD map, remove invalid maps, and then manually screen the map after automatic screening. S4. Divide the effective spectrum into intervals, normalize the number of pulses and map the gray values. After generating gray value channel data, calculate the spatial correlation coefficients of pulse phase, amplitude and number of pulses in the PRPD spectrum. S5. Construct a lightweight residual network, embed a channel attention module, construct a dynamic channel pruning mechanism, and set up a composite mechanism of identity mapping and working condition attention in the residual module. S6. Input dual-channel feature data and output partial discharge type identification results through the trained lightweight residual network. S7. Construct a data indicator system and dynamically optimize the model.

[0008] In the preferred embodiment, step S1 specifically includes: First, based on the monitoring requirements of GIS equipment, ultra-high frequency sensors are deployed at key monitoring locations on the GIS casing to cover all equipment intervals to be monitored and eliminate monitoring blind spots; these sensors are used to capture ultra-high frequency electromagnetic wave signals radiated during partial discharge of GIS and convert them into processable electrical signals. Before the data acquisition process, the sensor's operating parameters are preset, including the detection frequency band, sampling frequency, and data format. The detection frequency band must cover the typical frequency band of the partial discharge ultra-high frequency signal, the sampling frequency must meet the signal acquisition requirements, and the data format must include phase information, pulse amplitude information, and pulse quantity information. After the sensor collects the raw analog signal, it transmits it to the data processing unit. During the transmission process, anti-interference measures are taken to avoid data distortion caused by electromagnetic interference. After being transmitted to the data processing unit, the analog signal is converted into a digital signal to form PRPD spectrum data containing four types of target partial discharge defect discharge signals: floating discharge, tip discharge, particle discharge and internal discharge. In the preferred embodiment, in step S2: First, historical PRPD (Presentation Point Detection) data were classified and analyzed. Three typical invalid spectrum types were identified: field verification signals, interference signals, and incomplete signals. For each type of invalid spectrum, typical characteristic parameters were extracted to form a feature set, including: The on-site verification signal pulses occur frequently. The frequency threshold is set based on the typical signal frequency of the on-site verification operation to extract the signal. The signal does not carry energy, that is, the energy value is zero. The proportion of high and low frequency components in the signal is fixed. This fixed proportion is determined by the verification equipment itself and is used to distinguish the verification signal from the actual discharge signal. The interference signal does not carry energy and has an energy value of zero; the pulse occurrence frequency is low, and the signal is extracted based on the typical frequency of background interference in the non-discharge state; and the amplitude of the interference signal is usually relatively stable with small fluctuations. The incomplete signal data has a high missing rate, so a threshold is set based on the impact of data integrity on the subsequent recognition effect; the phase distribution cannot completely cover the power frequency cycle, and there will be cases where some phase intervals have no data. The above feature parameters are organized by category and stored in the form of key-value pairs to construct an invalid map feature library.

[0009] In the preferred embodiment, step S3 specifically includes the following steps: S31, Random Forest Selection Model Training; First, the model structure is designed by integrating multiple decision trees. The number of decision trees is determined based on the data scale and recognition accuracy requirements to balance model performance and computational efficiency. The maximum depth of the decision trees is set so that the model can capture the key features of invalid maps without learning noise, thus avoiding overfitting caused by excessively deep decision trees. The minimum number of samples for each leaf node is set, and splitting stops when the number of samples for a node is less than this value to avoid generating an overly complex model. Subsequently, PRPD maps labeled "valid" or "invalid" were selected as training data, and the training set and validation set were divided proportionally. The training set was used for model parameter learning, and the validation set was used to evaluate model performance. Finally, the feature parameters in the invalid graph feature library constructed in step S2 are used as the input features of the model, and the "valid" or "invalid" label of the graph is used as the output target to train the random forest model. During the training process, the recognition accuracy of the model is monitored in real time through the validation set. When the accuracy reaches the preset standard, the training is stopped and the trained model is saved.

[0010] S32. Automatic primary screening; The PRPD map data obtained in step S1 is input into the trained random forest model. Each map is labeled, and maps that are determined to be invalid are automatically removed and not included in subsequent processing. The set of maps that are determined to be valid is included in the subsequent manual screening process. S33, Initial screening threshold correction; A dynamic threshold adjustment mechanism is designed to optimize the accuracy of automatic initial screening based on the false judgment rate of invalid spectra and the retention rate of valid spectra. When the false positive rate of invalid spectra is higher than the preset standard, it indicates that the screening rules are too lenient. Tighten the screening threshold and increase the criteria for judging invalid feature parameters to reduce the number of invalid spectra being mistakenly judged as valid. When the retention rate of valid spectra is lower than the preset standard, it indicates that the screening rules are too strict. Relax the screening threshold and lower the criteria for judging invalid feature parameters to avoid misjudging valid spectra as invalid. The threshold adjustment step size is adjusted proportionally according to the original threshold to smooth the adjustment process and avoid large fluctuations. By statistically analyzing the results of the automatic initial screening, the threshold is dynamically adjusted, and the adjusted threshold is updated into the screening rules of the random forest model for automatic initial screening of subsequent batches of data, forming a cyclical optimization.

[0011] S34, manual fine screening; Automatic initial screening may miss a small number of cases, so a manual fine screening step is also provided as a supplement. The manual screening only re-examines the spectra that are automatically screened as valid, avoiding manual processing of all spectra. Based on the typical characteristics of invalid spectra, and combined with the parameters in the feature library of step S2, the manual staff checks the phase integrity, amplitude rationality, and pulse distribution pattern of the spectra, and removes invalid spectra that have been missed.

[0012] In the preferred embodiment, step S4 specifically includes the following steps: S41, Interval division; Discretizing the continuous phase and amplitude parameters into a grid lays the foundation for subsequent statistical pulse counts and characteristic calculations. Using the power frequency phase as the horizontal axis, the phase axis is divided into multiple intervals at uniform intervals to capture the differences in phase distribution of different discharge types; Using the pulse signal amplitude as the vertical axis, the amplitude axis is divided into multiple intervals at uniform intervals to cover the dynamic range of the pulse amplitude, so that there is a reasonable sample distribution in each interval; By dividing the phase axis and amplitude axis into intervals, a two-dimensional grid structure is formed. Each grid corresponds to a specific phase and amplitude interval. The number of pulses in each grid is counted to obtain three-dimensional statistical data containing phase, amplitude, and number of pulses.

[0013] S42, Pulse count normalization processing; Because the absolute values ​​of pulse counts differ significantly across different spectra, direct use would affect the stability of model training. Therefore, normalization is required, as shown in the formula: (1); in, This represents the normalized number of pulses. The original number of pulses in the u-th row and v-th column of the grid; This represents the maximum number of pulses within all grids in the graph. Map the pulse counts of all spectra to the same numerical range to eliminate the influence of differences in the absolute values ​​of pulse counts between different spectra, and normalize the values ​​to between zero and one.

[0014] S43, Grayscale value mapping; To convert the normalized pulse count into image form for use as input to a deep learning model, grayscale value mapping is performed using the following formula: (2); in, The pixel grayscale value corresponding to the u-th row and v-th column grid; This represents the normalized number of pulses. This represents the upper limit of the image's grayscale levels; The more pulses there are, the closer the normalized value is to one, the closer the corresponding gray value is to zero, and the darker the pixel color. Conversely, the fewer pulses there are, the closer the gray value is to the maximum gray level, and the lighter the pixel color. This reflects the distribution characteristics of the number of pulses, making the active discharge area appear dark in the grayscale image, which makes it easier for the model to capture key features.

[0015] S44. Calculate the spatial correlation coefficient; Grayscale images cannot retain three-dimensional spatial correlation information. For the identification of multiple superimposed signals, it is necessary to calculate the spatial correlation coefficient, the formula of which is: (3); in, Let be the spatial correlation coefficient of the number of pulses between the (i,j)th grid and the adjacent (i+1,j+1)th grid; This represents the normalized number of pulses. This is the average number of normalized pulses across all grids in the spectrum; The spatial correlation coefficient ranges from -1 to 1. A positive value indicates a positive correlation between the number of pulses in two adjacent grids, while a negative value indicates a negative correlation. The larger the absolute value, the stronger the correlation. It is used to capture the variation pattern of the number of pulses in adjacent phase and amplitude intervals, preserve the spatial correlation characteristics of three-dimensional data, and form 3D correlation coefficient channel data.

[0016] S45. Construct dual-channel data; The grayscale channel data generated in step S43 and the 3D correlation coefficient channel data generated in step S44 are concatenated according to the channel dimension to form dual-channel feature data containing the distribution characteristics of the number of pulses and spatial correlation information.

[0017] In the preferred embodiment, step S5 specifically includes the following steps: S51. Construct the basic network architecture; Based on the ResNet-50 architecture, this paper addresses the degradation problem of deep networks through residual modules, which include an input layer, convolutional layers, residual module groups, pooling layers, and fully connected layers. The input layer sets the number of input channels to 2, corresponding to dual-channel feature data, and the input size is consistent with the spatial size of the dual-channel data; The number of convolutional kernels in the first convolutional layer is set according to the feature extraction requirements, the stride is set to 1, and SAME padding is used to keep the spatial size of the feature map unchanged after convolution. A batch normalization layer is added after the convolutional layer to accelerate model training convergence, and an activation function layer using the ReLU function is added to enhance the non-linear expressive ability of the model. The residual module group retains the four residual module groups of ResNet-50. Each module group contains multiple residual modules, and its structure is as follows: convolutional layer, batch normalization, activation function, convolutional layer, batch normalization, identity mapping, activation function; depending on the change in the number of input and output channels, the identity mapping adopts two connection methods: The first method uses solid line connections. When the number of input channels and output channels of the residual module are the same, a direct identity mapping is used. The input data is not processed by convolution and is directly added to the convolution output of the module. The formula is: (4); in, This is the output of the residual module; This refers to the output of the convolutional layer within the module; For the module's input; The second method uses dashed lines for connection. When the number of input channels and output channels of the residual module differs, an identity mapping with convolution is used. A one-dimensional convolutional layer is used to adjust the number of input channels to match the number of output channels before summing them. The formula is: (5); in, These are the weight parameters of a one-dimensional convolutional layer; This helps maintain gradient stability during deep training of the network and avoids degradation.

[0018] S52, Embedded attention module; Embed a channel attention module at the output of each residual module group to make the model focus on the more important feature channels; First, global average pooling is performed on the output feature map of the residual module group to transform the feature map of each channel into a one-dimensional feature value, which reflects the global information of the corresponding channel. Subsequently, the feature vector after global average pooling is input into two fully connected layers, where: The first fully connected layer compresses the features, reducing the number of parameters, and uses ReLU as the activation function; The second fully connected layer activates the features and outputs a weight vector with the same number of channels. The activation function is Sigmoid, which maps the weight values ​​to the range of zero to one. Finally, the activated weight vector is multiplied by the feature map output by the residual module group according to the channel dimension, and the features of each channel are weighted. The channel with the larger weight value will have a higher weight in subsequent processing.

[0019] S53. Design a dynamic channel pruning mechanism; The pruning threshold is set based on the weight values ​​output by the channel attention module. Channels with weight values ​​lower than this threshold are identified as redundant channels and are pruned. The first pruning is performed after a set number of training iterations, which must be sufficient for the model to have learned effective features. Subsequent pruning is performed at fixed iteration intervals to remove redundant channels. The initial pruning rate is set at a low level to avoid over-pruning and loss of effective features; the pruning rate is then gradually increased until the network computational cost is reduced to the target level; after each pruning, the weights of the remaining channels are fine-tuned to control the loss of model accuracy. After pruning, the network channel configuration is updated, and the convolutional kernel weights and bias parameters corresponding to redundant channels are deleted to further simplify the network structure.

[0020] S54. Construct sub-networks; Based on the two complex operating conditions of phase shift and multi-signal superposition, a dedicated sub-network is constructed to work in conjunction with the basic network, wherein: The phase-shifting adapter subnetwork is used to extract stable features in phase-shifting scenarios. Its structure consists of convolutional layers, batch normalization, activation functions, pooling layers, and fully connected layers. The convolutional layers use different kernel sizes depending on the situation to capture changes in feature distribution after phase shifting. The pooling layers use max pooling to enhance the model's robustness to different degrees of phase shifting. The fully connected layers integrate the extracted local features into global features and concatenate them with the output features of the base network. The multi-signal superposition adapter subnetwork is used to separate the features of a single signal after multiple signals are superimposed. The structure consists of a dilated convolutional layer, batch normalization, activation function, adaptive pooling layer, and fully connected layer. The dilated convolutional layer uses different dilation rates to expand the receptive field, capture the long-distance dependencies after multiple signals are superimposed, and separate the features of different signals. The adaptive pooling layer dynamically adjusts the pooling window according to the size of the feature map to keep the dimension of the output features consistent. The fully connected layer integrates the features and then concatenates them with the output features of the base network and the phase-shift adapter subnetwork.

[0021] S55. Introduce transfer learning strategies; Transfer learning strategies include a pre-training phase and a fine-tuning phase, wherein: In the pre-training phase, the basic network is pre-trained using the GIS partial discharge dataset. The dataset needs to contain a combination of various discharge types and working conditions so that the model can learn the general characteristics of partial discharge. The pre-trained optimizer uses stochastic gradient descent, the learning rate is set to an appropriate level, and the number of iterations is determined based on the size of the dataset and the model's convergence. The loss function is the cross-entropy loss function, and the optimization objective is to minimize the model's classification error on the public dataset. During the fine-tuning phase, the pre-trained base network weights are used as initial weights. A dataset containing single defects, composite defects, and working condition interference samples is used to fine-tune the overall network, which includes the base network, channel attention module, dynamic pruning unit, and working condition adaptation sub-network. The learning rate for fine-tuning is set to one-tenth of the pre-training learning rate to avoid excessive modification of the initial weights. The optimizer uses an adaptive momentum estimation algorithm to accelerate model convergence. Only the upper-layer weights and sub-network weights are fine-tuned to reduce the number of training parameters and improve training efficiency.

[0022] In the preferred embodiment, step S6 specifically includes the following steps: S61. Design a dynamic weighted cross-entropy loss function; We design a dynamic weighted cross-entropy loss function to dynamically adjust sample weights during training, thereby solving the problems of sample imbalance and low recognition accuracy of difficult-to-distinguish samples. The basic loss function formula is: (6); in, Based on cross-entropy loss; This represents the number of partial discharge types. This is the true label for the i-th type of discharge; To predict the probability of the i-th type of discharge for the model; The weights of samples in each category are calculated using the following formula: (7); in, Let be the class weight of the c-th sample; The total number of samples in the training set; Number of categories; The number of samples in class c; The class with the fewest samples is given a larger weight, so that the model can balance the weights of minority class samples during training. Calculate the difficulty weight for each sample to improve the accuracy of identifying difficult samples. The formula is as follows: (8); in, Let be the difficulty weight of the i-th sample; This represents the cumulative prediction accuracy for this sample. To find the minimum value, avoid the denominator being zero; The lower the cumulative prediction accuracy of a sample, the greater the difficulty weight, thus reinforcing the characteristics of samples that are difficult to distinguish using reinforcement learning. Multiply the category weight by the difficulty weight to obtain the total weight of each sample, using the following formula: (9); in, For dynamic weighted cross-entropy loss; The total weight of the i-th sample; Let be the basic cross-entropy loss for the i-th sample.

[0023] S62, Feature enhancement processing; Feature enhancement is performed before the feature map is input into the fully connected layer; First, we determine whether it is an extreme working condition by analyzing whether the phase offset angle of the sample and the number of signal superposition layers exceed the threshold. Next, for extreme working condition samples, a dynamic mask with the same dimension as the feature map is generated, where the mask value of extreme working condition features is greater than 1, and the mask value of non-extreme working condition features is 1; the mask value is dynamically adjusted according to the training effect to prevent overfitting of extreme working condition features. Finally, the feature map is multiplied element-wise with the dynamic mask to amplify the feature values ​​under extreme conditions, making it easier for the model to capture them.

[0024] S63, Model Training; First, the dual-channel feature data generated in step S4 is input into the model, enters the network through the input layer, and passes through the convolutional layer, residual module group, channel attention module, dynamic pruning unit, working condition adaptation sub-network, feature enhancement layer and fully connected layer to output the predicted probability distribution of various discharge types. Subsequently, the dynamic weighted cross-entropy loss is calculated according to formula (9) as the optimization objective for model training; the gradient of the loss function with respect to the parameters of each layer of the network is calculated using the backpropagation algorithm; the network parameters are updated according to the gradient using the optimizer, and the weights and biases are adjusted to reduce the value of the weighted loss function; the learning rate is reduced as the number of iterations increases during training to make the model converge. Finally, a termination condition is introduced: when the recognition accuracy of the validation set is less than a threshold after continuous iterations, training is stopped to prevent overfitting, and the current network parameters are saved as the final training result.

[0025] S64. Identify the discharge type; First, for the PRPD map to be identified, invalid maps are removed in step S3, and dual-channel feature data is generated in step S4. The preprocessing method is the same as that of the training data to ensure that the data distribution is consistent. Subsequently, the trained lightweight residual network parameters are loaded, the preprocessed dual-channel feature data is input into the model, and the predicted probability distribution of various discharge types is output through forward propagation calculation. Finally, the category with the highest predicted probability is selected as the partial discharge type identification result of the spectrum, and the probability value of each category is output to provide a reference for operation and maintenance personnel; if the maximum probability is lower than the confidence threshold, it is determined to be an uncertain type and is manually reviewed.

[0026] In the preferred embodiment, step S7 specifically includes the following steps: S71. Construct a data indicator system; Three types of data metrics are constructed, including data cleaning quality metrics, model performance metrics, and operating condition adaptability metrics, among which: Data cleaning quality indicators include the false positive rate of invalid maps and the retention rate of valid maps, reflecting the cleaning effect of step S3; the lower the false positive rate of invalid maps and the higher the retention rate of valid maps, the better the data cleaning quality and the higher the quality of the model input data. Invalid spectrum misclassification rate This refers to the proportion of invalid maps that were mistakenly judged as "valid" out of the total number of invalid maps. The calculation formula is: (10); in, The number of invalid maps that were mistakenly judged as "valid"; This represents the total number of invalid maps submitted for screening. Effective spectral retention rate The percentage of valid maps correctly identified as "valid" out of the total number of valid maps is calculated using the following formula: (11); in, The number of valid maps that are correctly identified as "valid"; The total number of valid maps participating in the screening; The model performance metrics include recognition accuracy and convergence speed, reflecting the training effect of step S5 and the recognition effect of step S6; the higher the recognition accuracy, the more accurately the model can distinguish the discharge type; the faster the convergence speed, the higher the model training efficiency. Recognition accuracy The proportion of correctly identified samples out of the total number of test samples reflects the overall recognition ability of the model. The calculation formula is as follows: (12); in, The number of samples correctly identified in the test set; This represents the total number of samples in the test set. Convergence speed The number of iterations required to reduce the loss function value to a preset error reflects the efficiency of model training, and is calculated using the following formula: (13); in, This represents the number of training iterations. For the first The loss function value of the next iteration; This represents the preset expected error. The operating condition adaptability index includes extreme condition recall rate and operating condition adaptability index, which reflect the method's ability to adapt to complex field conditions. The higher the extreme condition recall rate, the better the identification effect of extreme conditions. The higher the operating condition adaptability index, the stronger the performance stability of the method under different operating conditions. Extreme operating condition recall rate The proportion of correctly identified extreme condition samples out of the total number of extreme condition samples reflects the model's ability to identify extreme scenarios. The calculation formula is as follows: (14); in, The number of extreme operating condition samples that were correctly identified; This represents the total number of extreme operating condition samples. Operating condition adaptability index To assess the stability of the model's recognition accuracy under different operating conditions, reflecting the model's adaptability to changes in operating conditions, the calculation formula is as follows: (15); in, It is a function of standard deviation; For the model in The recognition accuracy under different working conditions; for Average recognition accuracy under various working conditions.

[0027] S72, Optimize algorithm design; The optimization algorithm uses the PID algorithm for dynamic parameter adjustment, and calculates the parameter adjustment amount based on the proportion, integral and derivative information of the index deviation. First, calculate the deviation between the actual value and the target value of each core indicator, using the following formula: (16); in, The index deviation at time t; The target value of the indicator; The actual value of the index at time t; When the deviation is positive, it indicates that the actual value is lower than the target value, and the parameters should be adjusted to improve the indicator; when the deviation is negative, it indicates that the actual value is higher than the target value, and the parameters should be adjusted to optimize other indicators.

[0028] Then, set the proportional coefficient, integral coefficient, and derivative coefficient, and obtain reasonable parameter values ​​through on-site data debugging to balance the adjustment speed and stability. The parameter adjustment amount is calculated based on the PID algorithm, using the following formula: (17); in, The parameter adjustment amount at time t; This is the proportionality coefficient; The integral coefficient; These are the differential coefficients; This is the cumulative deviation; This represents the rate of change of deviation.

[0029] S73, Parameter Adjustment and Continuous Optimization; Adjust key parameters based on the deviations of different indicators; When the false positive rate of invalid maps is higher than the target value, adjust the screening threshold in step S3, tighten the screening rules, and reduce the false positive rate. When the effective spectrum retention rate is lower than the target value, relax the screening threshold to increase the retention rate; When the recognition accuracy is lower than the target value, adjust the number of interval divisions, normalization method or spatial correlation coefficient calculation method in step S4 to optimize feature quality, improve feature resolution, and adjust the normalization formula to enhance feature discrimination. When the convergence speed is lower than the target value, adjust the learning rate, batch size, or network structure parameters in step S5 to accelerate model convergence. When the recall rate under extreme conditions is lower than the target value, adjust the feature enhancement layer mask value in step S6 to enhance the enhancement degree of the features under extreme conditions. When the network computational load exceeds the target value, adjust the dynamic pruning rate in step S5 to further eliminate redundant channels. Parameter adjustments are performed at fixed intervals, with the interval length determined based on the data processing batches. Parameter boundary ranges are also set to prevent excessive parameter adjustments from causing system instability. After parameter adjustment, the system enters collaborative closed-loop optimization. After each task cycle iteration, data of various core indicators are collected, the indicator deviation is calculated, the parameter adjustment amount is obtained through the PID algorithm, and the relevant parameters are adjusted. In subsequent iteration cycles, steps S1-S6 are executed according to the adjusted parameters, indicator data is collected again, the deviation is calculated, the parameters are adjusted, and this process is repeated until the indicator deviation stabilizes within an acceptable range. The optimization results of each iteration are stored in the database to form an optimization log for easy subsequent analysis and tracing.

[0030] This invention provides a GIS partial discharge type identification method. It constructs an invalid spectrum feature library by extracting parameters of on-site verification, interference, and incomplete signals, and learns cleaning rules using a random forest algorithm. It innovatively designs a dynamic threshold adjustment mechanism based on the false judgment rate of invalid spectrum and the retention rate of valid spectrum, which overcomes the problem that static thresholds cannot adapt to complex on-site environments. It realizes a dynamic closed loop of invalid spectrum removal and automatic initial screening rules, thereby improving the quality and purity of the input data for subsequent model training from the source. Meanwhile, while dividing the spectrum into intervals, normalizing the number of pulses, and mapping gray values ​​to generate gray value channel data, this invention also calculates the spatial correlation coefficient of the number of pulses between adjacent grids in terms of phase, amplitude, and number of pulses. Then, it splices the data according to the channel dimension to construct dual-channel feature data, which completely preserves the three-dimensional spatial correlation information of the original data, avoids the information loss caused by feature dimensionality reduction, and improves the model's ability to analyze and recognize complex multi-signal superposition features. Furthermore, this invention introduces identity mapping based on ResNet-50 architecture to alleviate the degradation of deep networks, and embeds channel attention modules at the output of residual module groups. Based on the weight values ​​output by the attention modules, a dynamic channel pruning mechanism is innovatively designed to remove redundant channels at fixed intervals, enabling the network to autonomously and dynamically strip away invalid computation branches during training, thereby achieving lightweight deployment of deep learning models on low-computing-power edge devices in the field. Furthermore, this invention constructs a phase-shifting adapter subnetwork and a multi-signal superposition adapter subnetwork, and adds a feature enhancement operation before the fully connected layer. By determining extreme working conditions, a dynamic mask is generated and interactively calculated with the feature map. At the same time, it is combined with a dynamic weighted cross-entropy loss function composed of category weights and difficulty weights. This not only enhances the weak features under extreme working conditions, but also separates overlapping signals, achieving highly robust composite defect identification.

[0031] In summary, this invention proposes a complete multi-objective collaborative technical solution, encompassing dynamic data cleaning, multi-dimensional feature construction, lightweight network design, extreme condition adaptation, and closed-loop parameter optimization throughout the entire process. It systematically addresses the technical shortcomings of existing technologies, such as inaccurate invalid data removal, loss of spatial correlation features, deep network degradation, and poor adaptability to operating conditions. This improves the accuracy, real-time performance, and reliability of GIS partial discharge type identification, providing technical support for the efficient deployment of low-computing-power equipment in substations and the long-term stable operation of power systems. Attached Figure Description

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of a GIS partial discharge type identification method according to the present invention; Figure 2 This is a flowchart of step S3 of the GIS partial discharge type identification method of the present invention; Figure 3 This is a flowchart of step S4 of the GIS partial discharge type identification method of the present invention; Figure 4 This is a flowchart of step S5 of the GIS partial discharge type identification method of the present invention; Figure 5 This is a flowchart of step S6 of the GIS partial discharge type identification method of the present invention; Figure 6 This is a flowchart of step S7 of the GIS partial discharge type identification method of the present invention. Detailed Implementation

[0033] Example 1 This embodiment is based on field data from GIS substations, involving 220kV and 500kV GIS equipment. The monitoring area covers 13 substations with a total of 183,570 original PRPD maps. After final screening, 174,216 valid maps were obtained, including four types of target discharge signals: floating discharge, tip discharge, particle discharge, and internal discharge, as well as three types of invalid maps: field verification signals, interference signals, and incomplete signals. Figure 1-6 As shown, a method for identifying partial discharge types in GIS includes the following steps: Step S1: Collect PRPD map data uploaded by GIS substation. The PRPD map data includes discharge signals corresponding to floating discharge, tip discharge, particle discharge, and internal discharge. Based on the GIS equipment structure, six external UHF sensors are deployed at key monitoring locations of the basin-type insulators on the GIS shells of each substation to cover all monitoring equipment intervals, eliminate monitoring blind spots, and control the straight-line distance between adjacent sensors within 20m to avoid missing signals. The sensor detection frequency band is set to 300MHz-1500MHz, covering the typical frequency band of partial discharge ultra-high frequency signals. The sampling rate is set to 200MS / s to meet the signal acquisition accuracy requirements. The data format includes phase information, pulse amplitude and pulse count. The sensor captures the ultra-high frequency electromagnetic wave signal radiated by partial discharge, transmits it to the data processing unit through a shielded cable, and converts the analog signal into a digital signal through analog-to-digital conversion to form PRPD spectrum data containing four types of target discharge defect signals.

[0034] Step S2: Construct an invalid map feature library; A total of 183,570 historical PRPD maps were classified, and 9,354 invalid maps were identified. The pulse occurrence frequency of the on-site verification signal is ≥77751 times / second, the energy value is 0, 50Hz accounts for 28% and 100Hz accounts for 72%. Based on this characteristic, a frequency threshold of 70000 times / second is set to extract this type of signal. The energy value of the interference signal is 0, the pulse occurrence frequency is ≤290911 times / second, the amplitude fluctuation range is ≤5dBm, and the threshold of 300000 times / second is set based on the typical frequency of background interference to extract this type of signal. Incomplete signals with a data loss rate >10% and whose phase distribution cannot cover the complete power frequency cycle are filtered out by setting a loss rate threshold of 10%. Subsequently, the above feature parameters are stored in key-value pairs according to category to construct an invalid map feature library.

[0035] Step S3: Use the random forest algorithm to learn the cleaning rules and dynamically adjust the screening threshold to automatically screen the PRPD map, remove invalid maps, and then manually screen the automatically screened map. First, 100 decision trees are ensembled, with a maximum depth of 15 and a minimum number of samples per leaf node of 5. 50,000 PRPD maps labeled "valid" and "invalid" are selected and divided into training and validation sets in an 8:2 ratio. The feature parameters of the invalid map feature library are used as input and the "valid" or "invalid" label is used as output. The accuracy of the validation set is monitored in real time. When the accuracy reaches 95%, training is stopped and the model is saved. Subsequently, the PRPD map data collected in step S1 is input into the trained random forest model, each map is labeled, invalid maps are removed, and the set of valid maps is retained for manual screening. Next, a dynamic threshold adjustment mechanism was designed. When the false positive rate of invalid spectrum is greater than 5%, the screening threshold is tightened and the frequency threshold of the on-site verification signal is increased to 75,000 times / second. When the effective spectrum retention rate is <95%, the screening threshold is relaxed, and the threshold for incomplete signal loss rate is increased to 15%. The threshold adjustment step size is 10% of the original threshold, ensuring a smooth adjustment process; Finally, the spectra that are automatically screened as valid are manually reviewed. Based on the parameters of the invalid spectra feature library, the phase integrity, amplitude rationality, and pulse distribution pattern of the spectra are checked to see if they conform to the characteristics of the target discharge signal, and invalid spectra that were missed are removed.

[0036] Step S4: Divide the effective spectrum into intervals, normalize the number of pulses, and map the gray values. After generating gray value channel data, calculate the spatial correlation coefficients of pulse phase, amplitude, and number of pulses in the PRPD spectrum. First, the power frequency phase is used as the horizontal axis and the pulse signal amplitude is used as the vertical axis. The signal is divided into 256 intervals to form a 256×256 two-dimensional grid structure. The number of pulses in each grid is counted to obtain three-dimensional statistical data including phase, amplitude and number of pulses. Subsequently, normalization was performed using formula (1) to map the pulse count of all spectra to the 0-1 interval, eliminating the influence of the difference in the absolute value of pulse count between different spectra; grayscale value mapping was performed using formula (2) to set the upper limit of the image grayscale level. The more pulses there are, the darker the pixel color becomes, meaning that the active discharge area appears dark in the grayscale image. Formula (3) is used to calculate the spatial correlation coefficient between adjacent grids, retaining the spatial correlation characteristics of the three-dimensional data, and forming 3D correlation coefficient channel data. If the value is positive, it indicates that the number of pulses between the two grids is strongly positively correlated; if it is negative, it indicates that they are negatively correlated. Finally, the grayscale channel data and the 3D correlation coefficient channel data are concatenated along the channel dimension to form dual-channel feature data containing pulse number distribution characteristics and spatial correlation information.

[0037] Step S5: Construct a lightweight residual network, embed a channel attention module, construct a dynamic channel pruning mechanism, and set up a composite mechanism of identity mapping and working condition attention in the residual module. First, based on the ResNet-50 architecture, the number of input layer channels corresponding to the dual-channel feature data is set to 2, and the input size is 256×256; the number of convolutional kernels in the first convolutional layer is 64, the stride is 1, and SAME padding is used; there are 4 residual module groups, which are connected by solid lines when their input and output channels are the same, and calculated according to formula (4); when the number of channels is different, they are connected by dashed lines, and calculated according to formula (5). The number of channels is adjusted by one-dimensional convolutional layers. Next, a channel attention module is embedded at the output of each residual module group to convert the feature map into one-dimensional feature values ​​through global average pooling. These features are then input into two fully connected layers. The first layer compresses the features using the ReLU activation function, while the second layer outputs a weight vector using the Sigmoid activation function. The weight vector is multiplied by the feature map along the channel dimension, allowing the model to focus on the channel features with higher weights. Then, the pruning threshold is set according to the weight values ​​output by the channel attention module. The first pruning is set to be performed after 100 training iterations, with a pruning rate of 10%. Pruning is performed once every 50 iterations thereafter, with the pruning rate increasing by 5% each time, until the network computation is reduced to the target level. After each pruning, the weights of the remaining channels are fine-tuned to control the loss of accuracy. Subsequently, a subnetwork is constructed, in which: The phase-shifting adapter subnetwork uses 3×3 convolutional kernels for its convolutional layers and max pooling for its pooling layers. The fully connected layers integrate local features and then concatenate them with the output of the base network to enhance robustness to phase shifts. The hole rate of the dilated convolutional layer of the multi-signal superposition adapter subnetwork is set to 2 to expand the receptive field. The adaptive pooling layer dynamically adjusts the pooling window. The fully connected layer integrates the features and splices them with the outputs of the base network and the phase-shifted adapter subnetwork to achieve multi-signal separation. Finally, a transfer learning strategy is introduced, in which: In the first pre-training phase, a basic network was trained using a dataset of 100,000 GIS partial discharge images containing various discharge types and working conditions. The optimizer was SGD, the learning rate was 0.01, the number of iterations was 500, and the loss function was the cross-entropy loss function. The training aimed to minimize the classification error. The second fine-tuning stage uses the pre-trained weights as the initial weights and a dataset of 20,000 images containing single defects, compound defects, and working condition interference to fine-tune the overall network. The learning rate is set to 1 / 10 of the pre-training learning rate, and the optimizer is Adam. The upper layer weights and sub-network weights are fine-tuned.

[0038] Step S6: Input dual-channel feature data and output partial discharge type identification results through the trained lightweight residual network; First, a dynamic weighted cross-entropy loss function is designed. The basic loss is calculated according to formula (6), which reflects the difference between the predicted probability and the true label. The category weight is calculated according to formula (7), which strengthens the weight of a minority of samples. The difficulty weight is calculated according to formula (8), which increases the weight as the cumulative accuracy decreases. The total loss is calculated according to formula (9), which multiplies the category weight and the difficulty weight to dynamically adjust the sample weight.

[0039] Next, before inputting the feature map into the fully connected layer, it is determined whether it is an extreme working condition by whether the phase offset angle is greater than 30° or whether the number of signal stacking layers is greater than or equal to 2, and a dynamic mask with the same dimension as the feature map is generated. The feature mask value for extreme working conditions is set to 1.2, and the mask value for non-extreme working conditions is 1.0. The feature map and the dynamic mask are multiplied element by element to amplify the feature value of extreme working conditions. Then, the dual-channel feature data generated by S4 is input into the model, and the predicted probability distribution of four types of discharge is output. The gradient is calculated by using the dynamic weighted cross-entropy loss as the optimization objective, and the network parameters are updated by using the optimizer. The learning rate is halved every 100 iterations during training. When the recognition accuracy on the validation set improves by less than 0.1% for 10 consecutive iterations, training is stopped and the network parameters are saved.

[0040] Finally, for the PRPD map to be identified, dual-channel feature data is generated through filtering in step S3 and processing in step S4; the trained model is loaded, the input data is forward propagated to output the predicted probability distribution, and the category with the highest predicted probability is selected as the recognition result. At the same time, the probability values ​​of each category are output. If the maximum probability is lower than 0.8, it is determined to be an uncertain type and is manually reviewed.

[0041] Step S7: Construct a data indicator system and dynamically optimize the model; In the data indicator system, For data cleaning quality indicators, the false positive rate of invalid maps is calculated according to formula (10), with a target value of ≤3%, and the retention rate of valid maps is calculated according to formula (11), with a target value of ≥98%. For the model performance indicators, the recognition accuracy is calculated according to formula (12), with a target value of ≥95%, and the convergence speed is calculated according to formula (13), with a target value of ≤300 iterations; For the working condition adaptability index, the extreme working condition recall rate is calculated according to formula (14), with a target value of ≥90%, and the working condition adaptability index is calculated according to formula (15), with a target value of ≥0.9.

[0042] Subsequently, the parameters were dynamically adjusted using the PID algorithm, and the proportional coefficient, integral coefficient and derivative coefficient were adjusted according to the field conditions; the index deviation was calculated according to formula (16), and the parameter adjustment amount was calculated according to formula (17) to balance the adjustment speed and stability. If the false positive rate of invalid spectrum is >3%, then tighten the screening threshold and adjust the interference signal amplitude fluctuation threshold to ≤3dBm; If the recognition accuracy is less than 95%, increase the number of adjustment intervals and optimize the feature quality. If the recall rate is less than 90% under extreme conditions, adjust the feature enhancement layer mask value. If the network computational load is insufficient, increase the dynamic pruning rate; The parameter adjustment cycle is set to 1000 maps. After each iteration cycle, index data is collected, deviation is calculated and parameters are adjusted. Closed-loop optimization is performed until the index deviation stabilizes within an acceptable range.

[0043] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for identifying partial discharge types in GIS, characterized in that, Includes the following steps: Step S1: Collect phase analysis partial discharge spectrum data of discharge signals corresponding to floating discharge, tip discharge, particle discharge and internal discharge uploaded by GIS substation; Step S2: Construct an invalid map feature library; Step S3: Use the random forest algorithm to learn the cleaning rules and dynamically adjust the screening threshold to perform initial screening of phase analysis partial discharge spectra. After removing invalid spectra, perform manual fine screening of the automatically screened spectra. Step S4: Divide the effective spectrum into intervals, normalize the number of pulses and map the gray values. After generating gray value channel data, calculate the spatial correlation coefficients of pulse phase, pulse amplitude and pulse number in the phase analysis partial discharge spectrum to generate three-dimensional correlation coefficient channel data. Then, stitch the gray value channel data and the three-dimensional correlation coefficient channel data to construct dual-channel feature data. Step S5: Construct a lightweight residual network, embed a channel attention module, construct a dynamic channel pruning mechanism, and set identity mapping and operating condition attention mechanism in the residual module. Step S6: Input dual-channel feature data and train the lightweight residual network constructed in step S5. After training, use the lightweight residual network for real-time identification of partial discharge types in GIS.

2. The GIS partial discharge type identification method according to claim 1, characterized in that, In step S1, sensors are deployed at key monitoring locations on the GIS shell to capture electromagnetic wave signals radiated during partial discharge, and the detection band, sampling frequency and data format of the sensors are preset. Raw analog signals are collected by sensors and transmitted to the data processing unit, with anti-interference measures taken simultaneously during the transmission process; The original analog signal transmitted to the data processing unit is converted into phase-analyzed partial discharge spectrum data containing the discharge signal of the target partial discharge defect.

3. The GIS partial discharge type identification method according to claim 1, characterized in that, Step S2 specifically includes: Historical phase analysis partial discharge spectrum data is classified and analyzed to extract typical invalid spectra, including on-site verification signals, interference signals, and incomplete signals; Feature parameters of the on-site verification signal are extracted based on pulse occurrence frequency, energy value, and the proportion of high and low frequency components; feature parameters of the interference signal are extracted based on pulse occurrence frequency, energy value, and amplitude fluctuation range; feature parameters of the incomplete signal are extracted based on data missing rate and phase distribution covering the power frequency cycle. The extracted feature parameters are organized by category and stored in key-value pairs to construct an invalid map feature library.

4. The GIS partial discharge type identification method according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31. Design the random forest algorithm model structure, select phase analysis partial discharge spectra labeled "valid" or "invalid" as training data, divide the training set and validation set according to the ratio, take the feature parameters of the invalid spectrum feature library as input features, take the "valid" or "invalid" label of the spectrum as output target, train the random forest model and save it. S32. Input the phase analysis partial discharge spectrum data obtained in step S1 into the trained random forest model, automatically remove the spectrums that are determined to be invalid, and retain the set of valid spectrums. S33. A dynamic threshold adjustment mechanism is established based on the false positive rate of invalid maps and the retention rate of valid maps generated in the initial screening process. The screening threshold is adjusted according to the false positive rate of invalid maps and the retention rate of valid maps, and the adjusted threshold is updated and fed back into the screening rules of the random forest algorithm model. S34. Manually re-examine the spectra that were determined to be valid during the initial screening process. By using the typical characteristics of invalid spectra and the parameters of the feature library, check the phase integrity, amplitude rationality and pulse distribution pattern of the spectra, and remove invalid spectra that were missed.

5. The GIS partial discharge type identification method according to claim 1, characterized in that, Step S4 specifically includes the following steps: S41. Using the power frequency phase as the horizontal axis and the pulse signal amplitude as the vertical axis, the phase axis and amplitude axis are divided into a two-dimensional grid structure at uniform intervals, and the number of pulses in each grid is counted to form three-dimensional statistical data. S42. Calculate the quotient of the original pulse count within a single grid and the maximum pulse count in the entire spectrum, and use it as the pulse count after normalization. S43. Calculate the pixel difference by multiplying the upper limit of the image gray level by the normalized number of pulses, complete the gray value mapping, and generate gray value channel data. S44. Calculate the difference between the normalized pulse count and the global average pulse count for each of two adjacent grids, and obtain the spatial correlation coefficient between grids based on the difference to generate three-dimensional correlation coefficient channel data. S45. Concatenate the grayscale channel data and the three-dimensional correlation coefficient channel data according to the channel dimension to construct dual-channel feature data.

6. The GIS partial discharge type identification method according to claim 1, characterized in that, Step S5 specifically includes the following steps: S51. Using ResNet-50 as the base network, introduce identity mapping with two connection methods, solid line connection and dashed line connection, into the residual module of the base network. When the number of input channels and the number of output channels of the residual module are the same, the identity mapping without convolution is used; when the number of input channels and the number of output channels of the residual module are different, the identity mapping with convolution is used. S52. Embed a channel attention module at the output of each residual module group, perform global average pooling on the feature map, input it into two fully connected layers, and multiply the activated weight vector with the feature map by the channel dimension. S53. Set the pruning threshold based on the weight values ​​output by the channel attention module, perform dynamic channel pruning according to the set number of iterations and intervals, and fine-tune the remaining channel weights after pruning. S54. Construct a phase-shifting adaptation subnetwork and a multi-signal superposition adaptation subnetwork, extract the stable features under the phase-shifting scenario and the signal features after separating the superposition of multiple signals respectively, and concatenate them with the output features of the basic network. S55. A transfer learning strategy including a pre-training phase and a fine-tuning phase is adopted, and the base network is pre-trained and then the overall network is fine-tuned.

7. The GIS partial discharge type identification method according to claim 6, characterized in that, In step S54, the phase shift adapter network uses dynamically sized convolutional kernels to capture the displacement changes of the phase feature distribution, and after downsampling by the max pooling layer, it is integrated by the fully connected layer into global features characterizing the phase shift. The multi-signal superposition adapter network expands the receptive field of features by using dilated convolutional layers of different levels to remove the long-range dependence of superimposed signals. It uses adaptive pooling layers to unify the output spatial dimension under different superposition scenarios and integrates them into independent signal features by fully connected layers.

8. The GIS partial discharge type identification method according to claim 6, characterized in that, Step S6 specifically includes the following steps: S61. Construct a dynamic weighted cross-entropy loss function composed of the basic cross-entropy loss, category weight, and difficulty weight; S62. Before inputting the feature map into the fully connected layer, the extreme working conditions are determined by the phase offset angle and the number of signal superposition layers. A dynamic mask is generated and multiplied with the feature map to amplify the feature values ​​of the extreme working conditions for feature enhancement processing. S63. With minimizing the dynamic weighted cross-entropy loss as the training optimization objective, the backpropagation algorithm and adaptive optimizer are used to update the model parameters of the lightweight residual network until the termination condition is met. S64. After preprocessing the unknown phase analysis partial discharge spectrum to be identified, extract dual-channel feature data and use the dual-channel feature data as the input of the trained lightweight residual network. The output of the lightweight residual network is the predicted probability distribution of various types of discharge. The category corresponding to the highest predicted probability is taken as the identification result, and manual review is prompted when the highest probability is lower than the preprocessing reliability.

9. The GIS partial discharge type identification method according to claim 8, characterized in that, In step S61, the class weight is obtained by calculating the quotient of the total number of training set samples divided by the product of the number of discharge types and the number of samples of the specified class. The difficulty weight is obtained by calculating the reciprocal of the sum of the cumulative accuracy of a single sample and the minimum zero constant. The class weight, difficulty weight and basic cross-entropy loss corresponding to a single sample are multiplied and summed to obtain the final total model loss. In step S62, the phase offset angle of the sample to be processed and the number of signal superposition layers are analyzed to determine whether it belongs to an extreme working condition. For extreme working conditions, a weight matrix with an absolute value greater than the corresponding value of non-extreme working conditions is assigned as a dynamic mask to enhance the extreme features.

10. The GIS partial discharge type identification method according to any one of claims 1-9, characterized in that, It also includes step S7: constructing a data indicator system and dynamically optimizing the lightweight residual network model; Step S7 specifically includes the following sub-steps: S71. Establish a quantitative index library for evaluating the data cleaning stage, model training stage, and application identification stage. The quantitative index library includes the invalid map misjudgment rate and valid map retention rate for cleaning quality, the identification accuracy and convergence speed for model performance, and the extreme working condition recall rate and working condition adaptability index for adapting to complex environments. S72. Calculate the index deviation based on the actual value of the above index in each iteration cycle and the preset target value. Substitute the index deviation into the proportional-integral-derivative algorithm and combine the proportional coefficient, integral coefficient and derivative coefficient tuned on site to calculate the dynamic adjustment amount of the model parameters. S73. Adjust the corresponding key parameters according to the indicator deviation. The parameter adjustment is carried out at a fixed period and the boundary range is set to form a collaborative closed-loop optimization until the indicator deviation is stable within an acceptable range.

Images