Deep learning based distribution network cable partial discharge prpd phase pattern recognition system

Abstract

The application relates to the technical field of phase pattern recognition, and specifically discloses a PRPD phase pattern recognition system for partial discharge of a distribution network cable based on deep learning. The system comprises: a feature extraction module which extracts multi-scale features to obtain an effective prediction frame; a subordinate screening module which performs spatial overlap analysis, divides the prediction frame which is interfered into a dominant scale frame and a subordinate scale frame; a feature analysis module which calculates the phase domain distribution gradient of the subordinate scale frame and quantifies an independent saliency; a condition construction module which dynamically relaxes the global filtering criterion according to the saliency and generates a local reservation condition; and a constraint verification module which introduces the local reservation condition for deduplication verification, and outputs the final defect category and positioning result. The application is beneficial to isolating independent real defects and false associated noise under the condition of spatial nesting and overlapping of multiple discharge sources, and improves the positioning recognition accuracy and engineering reliability under the condition of multiple defect concurrency.

Description

Technical Field

[0001] This invention relates to the field of phase map recognition technology, and more specifically to a deep learning-based partial discharge (PRPD) phase map recognition system for distribution network cables. Background Technology

[0002] The insulation condition assessment of distribution network cables typically relies on partial discharge monitoring technology. Among them, PRPD (Phase Resolved Partial Discharge) phase maps can intuitively reflect the phase distribution and amplitude variation of discharge pulses within the power frequency cycle. Deep learning-based target detection algorithms (such as the YOLO series) have been widely introduced into the field of automatic recognition of PRPD maps. The algorithms rely on feature extraction networks to output prediction boxes with category attributes and spatial coordinates, thereby realizing the location and classification of discharge defects.

[0003] In complex industrial environments, distribution network cables often experience multiple types of insulation defects simultaneously (e.g., internal air gap discharge with localized surface flashover). Mapped onto the two-dimensional plane of the partial discharge phase pattern (PRPD), clusters of different discharge sources are prone to spatial nesting and overlap. Existing PRPD phase pattern recognition systems, when faced with overlapping multiple discharge clusters, rely solely on rigid two-dimensional geometric crossover thresholds for direct filtering. They lack underlying topological identification methods targeting the physical characteristics of partial discharges (such as density jump patterns and phase distribution polarity). Consequently, they cannot accurately separate independent real defects from spurious accompanying noise in nested conditions, leading to a significant decrease in the accuracy and reliability of localization and recognition in scenarios with multiple concurrent defects.

[0004] To address this, the present invention provides a deep learning-based partial discharge (PRPD) phase map identification system for distribution network cables. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based partial discharge (PRPD) phase map identification system for distribution network cables to solve the aforementioned background problems.

[0006] The objective of this invention can be achieved through the following technical solutions: A deep learning-based partial discharge (PRPD) phase map identification system for distribution network cables includes the following modules: Feature extraction module: It is used to input the phase map to be measured into a preset target detection network for multi-scale feature extraction to obtain an initial prediction box with spatial coordinates; based on the preset confidence conditions, the initial prediction box is cleaned to obtain an effective prediction box with discharge localization features. Subordinate filtering module: used to perform spatial overlap analysis on valid prediction boxes and extract interference prediction boxes with enclosing interference phenomena; based on the geometric inclusion relationship between interference prediction boxes, the interference prediction boxes are divided into dominant scale boxes on the periphery and subordinate scale boxes on the inside. Feature analysis module: used to extract the visual feature matrices of the dominant scale box and subordinate scale boxes; calculate the phase distribution gradient of the subordinate scale box relative to the dominant scale box using the visual feature matrix; perform local feature quantization on the phase distribution gradient to obtain the independent significance of the subordinate scale box independence; Condition construction module: used to evaluate the objective existence state of discharge characteristics within the subordinate scale box based on independent significance; dynamically relax the global filtering criteria faced by the subordinate scale box in the non-maximum suppression stage according to the objective existence state, and construct local retention conditions for the subordinate scale box.

[0007] Furthermore, the process of cleaning the parameters is as follows: Extract background hash parameters from the phase map of the test subject, detached from the dense texture feature region; The initial recognition lower limit of the target detection network during the validation phase is extracted and used as the confidence filtering benchmark; the bias compensation amount for benchmark compensation is extracted based on the background hash parameter. The confidence filtering benchmark is numerically superimposed and corrected using the bias compensation amount to obtain a dynamic adaptive value; a preset category addressing calibration is introduced to extract the initial confidence level bound to the candidate discharge category from the category probability parameter; A filtering logic for discharge location features is constructed based on the initial confidence level and dynamic adaptive value. The initial prediction boxes are then filtered to obtain effective prediction boxes with discharge location features.

[0008] Furthermore, the initial prediction box is obtained as follows: The phase map to be measured is divided into feature-sensing grids with multiple resolution levels according to spatial dimensions; Extract the pixel distribution parameters mapped inside the feature-aware mesh; based on the pixel distribution parameters and a preset local scanning window, calculate the texture density features of the feature-aware mesh at different resolution levels; Based on the dense texture features, the size regression calculation is performed on the preset prior anchor boxes of the target detection network to generate the mapping boundary; Extract the coordinates of the two-dimensional vertices anchored in the two-dimensional plane by the mapping boundary; extract the category probability parameters of classification tendency through the attribute classification layer of the object detection network; encapsulate the two-dimensional vertex coordinates and category probability parameters into low-level data to obtain the initial prediction box with spatial coordinates.

[0009] Furthermore, the subordinate scale box is divided as follows: Extract the lateral and longitudinal limit coordinates of the interferometric prediction box; perform same-direction limit constraint comparison on the lateral and longitudinal limit coordinates to divide the interferometric prediction box into the dominant scale box on the periphery and the subordinate scale box on the inside.

[0010] Furthermore, the process of obtaining the interference prediction box is as follows: Extract the coordinates of the diagonal vertices of the valid prediction boxes; reduce the dimensionality of the diagonal vertex coordinates to an axial boundary sequence; By comparing the staggered distribution results of the axial boundary sequences, interference prediction boxes with enveloping interference phenomena are selected.

[0011] Furthermore, the method for performing the local feature quantization is as follows: Extract the gradient polarity of the phase domain distribution gradient; perform trend alignment analysis between the gradient polarity and the prior polarity rule to obtain the independent significance.

[0012] Furthermore, the method for performing the trend alignment analysis is as follows: The source-derived feature frequency difference value is used to calculate the gradient of the phase domain distribution. The sign of the numerical value carried by the frequency difference of the feature is analyzed, and the sign of the numerical value is used as the gradient polarity representing the direction of frequency jump. Extract the pre-set inner density and outer sparseness rules of distribution network cables, and take the inner density and outer sparseness rules as the a priori polarity rules; Perform a logical consistency check between the gradient polarity and the prior polarity rule to determine whether the gradient polarity exhibits convergent polarity or divergent polarity: If the sign of the numerical value is positive, then it is determined that a linear convergence operation is performed on the gradient distribution in the phase domain, and the numerical result after the linear convergence operation is used as the independent significance. If the sign of the numerical value is negative or absolute zero, the independent significance will be forcibly set to zero.

[0013] Furthermore, the local preservation condition is constructed as follows: Analyze the numerical activation features of independent significance in numerical space; evaluate the objective existence of discharge features within the subordinate scale frame based on the numerical activation features; The target detection network is extracted to determine the spatial overlap ratio of overlapping redundancy in the non-maximum suppression stage. The spatial overlap ratio is used as the regular suppression benchmark and the regular suppression benchmark is used as the global filtering criterion. Based on the objectively existing state, the execution rules are reconstructed to extract the suppression exemption factor for the global filtering criteria; By using the suppression exemption factor to relax the filtering scale of the global filtering criteria, local retention conditions are obtained.

[0014] Furthermore, it also includes the following modules: Constraint verification module: This module is used to introduce the interference prediction box into the local retention condition for deduplication constraint verification, and to filter out the target dependent boxes; it summarizes the target dependent boxes retained after constraint verification and the independent prediction boxes that have not interfered in the effective prediction boxes, and outputs the defect category and location results of the phase map.

[0015] Furthermore, the method for outputting the defect category and location result is as follows: Extract the overlap and encroachment ratio between the dominant scale box and the subordinate scale boxes; compare the overlap and encroachment ratio with the local preservation conditions under extreme value constraints, and filter to obtain the target subordinate boxes; The target's dependent boxes, dominant scale boxes, and independent prediction boxes are summarized as the final detected target; the category tendency carried by the final detected target is confirmed as the final classification result, and the final detected target is encapsulated as a global defect topology that presents the spatial distribution pattern of partial discharge; The global defect topology is used as the final defect category and location result for terminal output.

[0016] The beneficial effects of this invention are as follows: 1. This invention extracts hash parameters from the non-discharge background region to dynamically offset the initial identification lower limit during pre-deployed network convergence, adaptively correcting the confidence filtering benchmark to clean up the prediction boxes, which helps reduce the false alarm probability. In the subordinate selection stage, for the selected interference prediction boxes, the horizontal and vertical limit coordinates are extracted for same-direction constraint comparison and peripheral state verification, clearly dividing the interference objects into dominant and subordinate scale boxes. This facilitates the decoupling of data between dominant and subordinate targets, establishing a clear topological relationship, and reducing the bottleneck of false merging caused by a lack of hierarchical judgment when dealing with dense scattered points.

[0017] 2. In the feature analysis stage, the spatial pixel matrix corresponding to the master and subordinate scale boxes is separated, and the vector polarity of the phase domain distribution gradient is extracted. Trend alignment verification is performed by combining the prior rule of dense inner and sparse outer points presented by the inherent scattered points of partial discharge, quantifying the independent salience of subordinate targets. This allows the system to remove accompanying scattered noise points attached to the edge of the master discharge. In the condition construction module, the system determines the objective existence state of nested discharge regions by analyzing the activation pattern of independent salience in the numerical space. For targets exhibiting an independent existence state, a suppression exemption factor is extracted. Using the exemption factor, the system performs targeted scale relaxation, generating local retention conditions for evaluating subordinate boxes. This achieves adaptive boundary reconstruction based on the physical independence features within candidate boxes, enhancing the partial discharge detection rate of the algorithm in highly nested environments.

[0018] 3. The constraint verification module extracts the actual overlap and encroachment ratio between the dominant and subordinate scale boxes and performs extreme value constraint comparison by introducing the aforementioned relaxed local retention conditions. This filters out target subordinate boxes that break through the spatial encroachment limit, which helps to increase the effectiveness of partial discharge identification without interfering with the conventional identification process of isolated scattered points. The system finally encapsulates the retained targets and independent predicted boxes into a unified global defect topology for output. Attached Figure Description

[0019] The invention will now be further described with reference to the accompanying drawings.

[0020] Figure 1 This is a functional block diagram of the deep learning-based partial discharge PRPD phase map identification system for distribution network cables in this invention. Figure 2 This is a flowchart of parameter cleaning for the initial prediction box in this invention; Figure 3 This is a flowchart of the condition construction and constraint verification of the present invention. Detailed Implementation

[0021] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0022] Example 1:

[0023] like Figure 1 As shown, the deep learning-based partial discharge (PRPD) phase map identification system for distribution network cables includes the following modules: Feature extraction module: It is used to input the phase map to be measured into a preset target detection network for multi-scale feature extraction to obtain an initial prediction box with spatial coordinates; based on the preset confidence conditions, the initial prediction box is cleaned to obtain an effective prediction box with discharge localization features. The process of inputting the phase map to be measured into a preset target detection network for multi-scale feature extraction to obtain an initial prediction box with spatial coordinates is as follows: S101. Divide the phase map to be measured into feature sensing grids with multiple resolution levels according to spatial dimensions; Specifically, the method for dividing the feature-aware grid is as follows: Extract the original pixel scale of the phase map to be measured; Establish a downsampling step size that includes scaling rules; The original pixel scale of the phase map to be tested is continuously dimensionally shrunk according to the scaling reduction rule to obtain a feature perception grid with decreasing scale. S102. Extract the pixel distribution parameters mapped inside the feature-aware mesh; based on the pixel distribution parameters and the preset local scanning window, calculate the texture density features of the feature-aware mesh at different resolution levels. Specifically, the method for calculating the texture density features of feature-aware meshes at different resolution levels is as follows: The two-dimensional phase-amplitude coordinates mapped by the feature-sensing grid in the phase spectrum to be measured are analyzed, and the discharge statistical frequencies corresponding to the two-dimensional phase-amplitude coordinates are extracted. The frequency of discharge statistics is used as a parameter for the pixel distribution mapped within the feature-aware grid. Extract the phase-aware receptive field of the local scanning window in the horizontal dimension, and the amplitude-aware receptive field of the local scanning window in the vertical dimension; By utilizing the phase-sensing receptive field and the amplitude-sensing receptive field, an asymmetric local feature scan is performed within the feature-sensing grid to obtain a frequency distribution array of the local region. Extract the global frequency mean of the phase spectrum to be measured, and use the global frequency mean as the preset response intensity; The effective frequency extreme points in the statistical frequency distribution array that exceed the preset response strength; The total number of pixels contained in the feature-aware grid in two-dimensional space is extracted as the theoretical carrying capacity extremum. The texture density feature of the local discharge accumulation state is obtained by calculating the ratio of the effective frequency extremum point to the theoretical carrying capacity extremum of the feature-aware grid. S103. Based on the dense texture features, the size regression calculation is performed on the preset prior anchor boxes of the target detection network to generate the mapping boundary; Specifically, the method for generating mapping boundaries is as follows: Extract the preset prior anchor frames in the target detection network based on the size distribution characteristics of the distribution network cable pattern; Input dense texture features into the location regression layer of the object detection network; The location regression layer outputs the center coordinate offset and scale factor for the preset prior anchor frame; Using the center coordinate offset and the scale scaling factor, the preset prior anchor frame is spatially decoded and mapped to obtain a closed mapping boundary. S104. Extract the coordinates of the two-dimensional vertices anchored in the two-dimensional plane by the mapping boundary; extract the category probability parameters of classification tendency through the attribute classification layer of the object detection network; encapsulate the two-dimensional vertex coordinates and category probability parameters into low-level data to obtain the initial prediction box with spatial coordinates. Specifically, the method for extracting the class probability parameter of classification tendency through the attribute classification layer of the object detection network is as follows: Extract the initial classification tensor output by the attribute classification layer of the object detection network; Extract the pre-defined exponential mapping operator within the attribute classification layer; The initial classification tensor is compressed in numerical intervals using the exponential mapping operator to obtain a probability distribution sequence that exhibits a discrete distribution. Extract the pre-set candidate discharge categories for distribution network cables; The probability distribution sequence is parsed with the candidate discharge category using channel alignment, and the maximum probability extremum within the probability distribution sequence is extracted. The candidate discharge category mapped by the maximum probability extreme value is used as the initial category tendency; The maximum probability extreme value is used as the category probability parameter characterizing the initial category tendency; Among them, such as Figure 2 As shown, the process of cleaning the parameters of the initial prediction box based on the preset confidence level to obtain an effective prediction box with discharge localization features is as follows: S111. Extract the background hash parameters from the phase spectrum to be measured, which are detached from the dense texture feature region. Specifically, the method for extracting background hash parameters is as follows: The global grayscale mean of all pixels in the phase map to be tested is statistically analyzed; the preset noise tolerance coefficient is extracted, and the global grayscale mean is multiplied with the noise tolerance coefficient to dynamically generate a preset background threshold. Extract inactive discrete pixels with gray values ​​below a preset background threshold from the phase map to be tested; The phase fluctuation variance of discrete pixels within a periodic time window is statistically analyzed, and the phase fluctuation variance is used as a background hash parameter. Preferably, the noise tolerance coefficient is set according to the proportion of the noise floor distribution in the partial discharge historical calibration data, for example, it can be set to a floating-point constant between 1.1 and 1.5; S112. Extract the initial recognition lower limit of the target detection network during the verification phase, and use the initial recognition lower limit as the confidence filtering benchmark; extract the bias compensation amount for benchmark compensation based on the background hash parameters. Specifically, the method for extracting the initial recognition lower bound for convergence of the object detection network during the validation phase is as follows: Extract the predicted confidence samples output by the object detection network during the validation phase; Statistical prediction of confidence samples: identification recall parameters and identification precision parameters; The recall and precision parameters are cross-correlated and evaluated to obtain a comprehensive evaluation curve. Extract the confidence node parameters corresponding to the highest evaluation extreme value of the comprehensive evaluation curve; The confidence node parameters are used as the initial recognition lower bound for the convergence of the target detection network during the validation phase. Preferably, the method for extracting the bias compensation amount for benchmark compensation based on background hash parameters is as follows: Based on historical calibration data of partial discharge, an offset mapping table is established that maps background noise to offset values. Input the background hash parameters into the bias mapping table for data addressing, and match to obtain the corresponding bias compensation amount. S113. The bias compensation amount is used to perform numerical superposition correction on the confidence filtering benchmark to obtain a dynamic adaptive value; a preset category addressing calibration is introduced to extract the initial confidence level bound to the candidate discharge category from the category probability parameter. Specifically, the method for obtaining a dynamic adaptive value by performing numerical superposition correction on the confidence filtering benchmark using the bias compensation amount is as follows: Extract the initial reference constants that are fixed in the numerical space for the confidence filtering reference; The initial reference constant and the bias compensation amount are numerically superimposed to obtain the dynamic adaptive value after threshold boundary adjustment. Specifically, a pre-defined category addressing calibration is introduced, and the initial confidence level bound to the candidate discharge category is extracted from the category probability parameters as follows: Extract the system's underlying pre-defined category addressing calibration for candidate discharge categories; Analyze the multidimensional probability sequence encapsulated within the category probability parameters; By using category addressing calibration to perform node data matching in a multidimensional probability sequence, specific probability values ​​specific to candidate discharge categories are located and extracted. A specific probability value is used as the initial confidence level to represent that the initial prediction box belongs to a suspected defective target; S114. Based on the initial confidence level and dynamic adaptive value, construct the filtering logic of discharge location features, filter the initial prediction box, and obtain the effective prediction box with discharge location features. Preferably, the filtering logic is constructed as follows: If the initial confidence level is greater than or equal to the dynamic adaptive value, the initial prediction box is determined to meet the effective positioning criteria. The parameter sequence of the initial prediction box is retained to obtain an effective prediction box with discharge positioning characteristics. If the initial confidence level is less than the dynamic adaptive value, the initial predicted box is determined to be an artifact interference target. The parameter sequences of the initial predicted box are stripped, and the initial predicted box is marked as a failure filtering state, preventing the failure filtering state from entering the structure analysis process.

[0024] Subordinate filtering module: used to perform spatial overlap analysis on valid prediction boxes and extract interference prediction boxes with enclosing interference phenomena; based on the geometric inclusion relationship between interference prediction boxes, the interference prediction boxes are divided into dominant scale boxes on the periphery and subordinate scale boxes on the inside. The process of performing spatial overlap analysis on the effective prediction boxes to extract the interference prediction boxes exhibiting enclosing interference is as follows: S201. Extract the coordinates of the diagonal vertices of the valid prediction boxes; reduce the dimensionality of the diagonal vertex coordinates into an axial boundary sequence; Specifically, the process of dimensionality reduction to an axial boundary sequence is as follows: Extract the coordinates of the diagonal vertices of the valid prediction bounding boxes in the two-dimensional plane; By decomposing the coordinates of the diagonal vertices into a dimension-reducing sequence, we obtain the horizontal extremum sequence and the vertical extremum sequence. The horizontal and vertical extreme value sequences are recombined at the lowest level to obtain the axial boundary sequence. S202. By comparing the staggered distribution results of the axial boundary sequences, interference prediction boxes with enveloping interference phenomena are selected. Specifically, the method for extracting interference prediction boxes exhibiting encapsulation interference is as follows: If the pairwise valid prediction boxes exhibit numerical interleaving in both the horizontal and vertical extreme value sequences, it is determined that the pairwise valid prediction boxes substantially overlap and interfere in physical space, and the combination of the pairwise valid prediction boxes is marked as an interference prediction box. If two pairs of valid prediction boxes exhibit an absolutely isolated interval in the horizontal or vertical extreme value sequence, then the two pairs of valid prediction boxes are determined to be in a spatially isolated environment, and the two pairs of valid prediction boxes are marked as independent prediction boxes, while maintaining the original feature attributes of the independent prediction boxes unchanged. For example, the method for determining whether a horizontal extreme value sequence exhibits numerical interleaving or absolute isolation intervals is as follows: Extract the upper and lower horizontal limits of the first prediction box from each pair of valid prediction boxes, and extract the upper and lower horizontal limits of the second prediction box. Verify whether the upper horizontal limit of the first prediction box is greater than the lower horizontal limit of the second prediction box, and simultaneously verify whether the upper horizontal limit of the second prediction box is greater than the lower horizontal limit of the first prediction box. If both sets of verification conditions are met, it is determined that the horizontal extreme value sequences of the first and second prediction boxes are in a state of numerical interleaving. If any set of verification conditions is not met, it is determined that the horizontal extreme value sequences of the first prediction box and the second prediction box present an absolutely isolated interval. The logic for interleaving the vertical extremum sequence is consistent with that for the horizontal extremum sequence; Based on the geometric inclusion relationship between interferometric prediction boxes, the method for dividing the interferometric prediction boxes into dominant scale boxes on the periphery and subordinate scale boxes on the inside is as follows: S203. Extract the lateral and longitudinal limit coordinates of the interferometric prediction box; perform same-direction limit constraint comparison on the lateral and longitudinal limit coordinates to divide the interferometric prediction box into the dominant scale box on the periphery and the subordinate scale box on the inside. Preferably, the method for dividing the interferometric prediction box into a dominant scale box on the periphery and a subordinate scale box on the inside is as follows: For two target prediction objects that are combined and marked as interferometric prediction boxes, extract the horizontal and vertical limit coordinates of the target prediction objects respectively. By integrating the horizontal and vertical limit coordinates into spatial boundaries, the outer envelope interval corresponding to the target prediction object is obtained. Compare the outer envelope intervals of the two target prediction objects using limit constraints in the same direction: If the horizontal and vertical limit coordinates of one object are completely contained within the outer envelope of the other object, then the two target prediction objects are determined to have an absolute nesting feature. The target prediction object that presents an outer containment form is defined as the dominant scale box, while the target prediction object that presents an inner embedding form is defined as the subordinate scale box. If the lateral or longitudinal limit coordinates of two target prediction objects show a state of mutual breakthrough in the outer envelope interval, the two target prediction objects are determined to belong to a non-subordinate sticky discharge cluster. Both target prediction objects are downgraded and marked as parallel overlapping boxes, and the parallel overlapping boxes are directly output to the regular fusion process.

[0025] Example 2:

[0026] Please see Figure 1 and Figure 3 As shown, the deep learning-based partial discharge (PRPD) phase map identification system for distribution network cables also includes the following modules: Feature analysis module: used to extract the visual feature matrices of the dominant scale box and subordinate scale boxes; calculate the phase distribution gradient of the subordinate scale box relative to the dominant scale box using the visual feature matrix; perform local feature quantization on the phase distribution gradient to obtain the independent significance of the subordinate scale box independence; The process of extracting the visual feature matrices of the dominant and subordinate scale boxes is as follows: The spatial extremum coordinates of the dominant scale box and the subordinate scale box are mapped to the phase map to be measured, and the first pixel matrix corresponding to the dominant scale box and the second pixel matrix corresponding to the subordinate scale box are extracted. The overlapping area occupied by the second pixel matrix is ​​removed from the first pixel matrix to obtain the unique outer background matrix of the dominant scale box; The second pixel matrix is ​​used as the core inline matrix corresponding to the subordinate scale box; The outer background matrix and the core embedded matrix are collectively referred to as the visual feature matrix; The process of calculating the phase domain distribution gradient of the subordinate scale box relative to the dominant scale box using the visual feature matrix is ​​as follows: S301. Extract the frequency density of the outer background matrix and the core embedded matrix, as well as the phase span feature of the core embedded matrix; Specifically, the method for extracting the frequency density of the outer background matrix and the core embedded matrix, as well as the phase span feature of the core embedded matrix, is as follows: The discharge frequency is accumulated for both the outer background matrix and the core embedded matrix to obtain the global total frequency value of the outer background matrix and the local total frequency value of the core embedded matrix. The background frequency density is obtained by performing a unit area mapping allocation on the global total frequency value of the outer background matrix using the spatial area occupied by the outer background matrix; The core frequency density is obtained by performing a unit area mapping allocation on the total local frequency value of the core embedded matrix using the spatial area occupied by the core embedded matrix. Extract the horizontal axis of the power frequency phase of the phase spectrum to be measured; Project the core embedded matrix onto the horizontal axis of the power frequency phase, and statistically analyze the absolute physical range covered by the core embedded matrix on the horizontal axis of the power frequency phase. Use the absolute physical range as the phase span feature of the core embedded matrix. S302. Perform phase domain orthogonal mapping between frequency density and phase span characteristics to obtain the phase domain distribution gradient; Specifically, the method for performing phase domain orthogonal mapping is as follows: The difference between the core frequency density and the background frequency density is compared to obtain the characteristic frequency difference value that characterizes the degree of frequency jump. The frequency difference of features is taken as the primary feature dimension, and the phase span feature is taken as the secondary feature dimension; The primary and secondary feature dimensions are normalized and transformed into dimensionless features. Then, the underlying sequence is concatenated to construct a two-dimensional topological vector. Extract the Euclidean distance parameter of a two-dimensional topological vector in the vector space; The Euclidean distance parameter is used as the phase domain distribution gradient, which includes visual density and phase constraint attributes. The process of performing local feature quantization on the phase domain distribution gradient to obtain the independent significance of the dependent scale box independence is as follows: S303. Extract the gradient polarity of the phase domain distribution gradient; perform trend alignment analysis between the gradient polarity and the prior polarity rule to obtain the independent significance. Specifically, the method for obtaining the independent significance of the dependent scale box by combining local feature quantization with the phase domain distribution gradient is as follows: The source-derived feature frequency difference value is used to calculate the gradient of the phase domain distribution. The sign of the numerical value carried by the frequency difference of the feature is analyzed, and the sign of the numerical value is used as the gradient polarity representing the direction of frequency jump. Extract the pre-set inner density and outer sparseness rules of distribution network cables, and take the inner density and outer sparseness rules as the a priori polarity rules; Perform a logical consistency check between the gradient polarity and the prior polarity rule to determine whether the gradient polarity exhibits convergent polarity or divergent polarity: If the sign of the numerical value is positive, the gradient polarity is determined to be an inward convergence polarity that increases from the outer background matrix to the core embedded matrix. The core embedded matrix is ​​then identified as an independent and compact state that conforms to the prior polarity law. A linear convergence operation is performed on the gradient distribution in the phase domain, and the numerical result after the linear convergence operation is used as the independent significance. If the sign of the numerical value is negative or the absolute zero value, the gradient polarity is determined to be divergent and diffuse. The core embedded matrix is ​​identified as a scattered edge of the surrounding background matrix, and the independent significance is forcibly assigned to zero. For example, the method for performing a linear convergence operation on the phase domain distributed gradient is as follows: Extract the theoretical distance upper limit and absolute zero value calculated based on normalized features, and use them as the upper limit and lower limit of the distribution extreme values, respectively; The global span parameter is obtained by performing a numerical difference calculation between the upper limit and the lower limit of the distribution extreme values. The local bias parameter is obtained by numerically subtracting the phase domain distribution gradient from the distribution extreme lower limit. The normalized coefficients are obtained by performing a proportional mapping assignment on the local bias parameters using the global span parameter. The normalized coefficients are used as the numerical results after linear convergence.

[0027] Condition construction module: used to evaluate the objective existence state of discharge characteristics within the subordinate scale box based on independent significance; dynamically relax the global filtering criteria faced by the subordinate scale box in the non-maximum suppression stage according to the objective existence state, and construct local retention conditions for the subordinate scale box. The method for assessing the objective existence of discharge characteristics within the subordinate scale frame based on independent significance is as follows: S401. Analyze the numerical activation features of independent significance in the numerical space; evaluate the objective existence of discharge features within the subordinate scale frame based on the numerical activation features; Specifically, the method for assessing the objective existence of discharge characteristics within the subordinate scale frame based on independent significance is as follows: Extract the specific feature values ​​of independent significance at the data flow stage; The specific feature value is compared with the absolute zero value preset at the system's underlying layer to perform a numerical extremum check, determining whether the specific feature value exhibits a non-zero activation state or an absolute zero state: If the specific feature value is greater than the absolute zero value, it is determined that the specific feature value presents a non-zero activation state, and the discharge feature within the subordinate scale box is identified as having a real physical discharge source, thus defining the objective existence state as an independent existence state. If the specific feature value is equal to the absolute zero value, it is determined that the specific feature value is in an absolute zero state, and the discharge feature within the subordinate scale frame is identified as a false data interference imaging, and the objectively existing state is defined as an invalid associated state. Specifically, the global filtering criteria faced by the subordinate scale box in the non-maximum suppression stage are dynamically relaxed based on the objective existence state, and the local retention conditions for the subordinate scale box are constructed as follows: S402. Extract the spatial overlap ratio used by the target detection network in the non-maximum suppression stage to determine the overlapping redundancy, use the spatial overlap ratio as the regular suppression benchmark, and use the regular suppression benchmark as the global filtering criterion. Specifically, the method for extracting the conventional suppression benchmark is as follows: Extract the spatial overlap ratio of candidate bounding boxes during the non-maximum suppression stage of the object detection network; Set the spatial overlap ratio to the maximum allowable boundary that allows the prediction boxes to physically overlap; The maximum allowable boundary is defined as the conventional suppression benchmark; S403. Execute rule reconstruction logic based on the objective existing state and extract the suppression exemption factor for the global filtering criteria; Specifically, the method for extracting the suppression exemption factor for the global filtering criteria, based on the rule reconstruction logic executed according to the objectively existing state, is as follows: Introduce the objectively existing state into the rule reconstruction logic to determine whether the objectively existing state is an independent state or an invalid associated state: If the objectively existing state is an independent state, then extract the extreme value range of the independent significance. Extract the number of interval divisions preset by the system, and perform equidistant cutting operation on the span of extreme values ​​according to the number of interval divisions to obtain multiple continuous numerical intervals arranged from smallest to largest. Extract a set of compensation step size parameters that are in an increasing sequence from the system's preset parameters, sort them in ascending order, and map and assign the compensation step size parameters one by one to the corresponding continuous value intervals. Preferably, to ensure the above dynamic relaxation logic has clear engineering reproducibility, a specific numerical allocation strategy is provided: for example, the number of interval divisions is set to 3, and the range of the extreme values ​​of independent significance is equally divided into the first interval (0, 0.33), the second interval [0.33, 0.66), and the third interval [0.66, 1.0]; at the same time, the system's preset compensation step size parameters in an increasing sequence are set to: 0.05, 0.10, and 0.15; By pairing continuous numerical intervals with compensation step size parameters, a mapping rule array is obtained. The independent significance is determined one by one by the continuous numerical intervals within the mapping rule array, and the target interval that completely encompasses the independent significance is extracted. Extract the compensation step size parameter of the paired encapsulation within the target interval, and use the compensation step size parameter as the suppression exemption factor for the global filtering criterion; S404. By using the suppression exemption factor to relax the filtering scale of the global filtering criteria, local retention conditions are obtained. Specifically, the method of relaxing the filtering scale of the global filtering criteria by using the suppression exemption factor to obtain the local retention conditions is as follows: Extract the initial baseline values ​​for the global filtering criteria; The initial baseline value is numerically superimposed with the suppression exemption factor to obtain the tolerance filtering baseline for expanding the constraint boundary. Use the tolerance filtering criterion as a local retention condition for subordinate scale boxes; If the objectively existing state is an invalid associated state, then the extraction of the suppression exemption factor is abandoned, the initial baseline value of the global filtering criterion remains unchanged, and the global filtering criterion is directly used as the local retention condition for the subordinate scale box.

[0028] Constraint verification module: This module is used to introduce the interference prediction box into the local retention condition for deduplication constraint verification, and to filter out the target dependent box; it summarizes the target dependent boxes retained after constraint verification and the independent prediction boxes that have not interfered in the effective prediction boxes, and outputs the defect category and localization result of the phase map. The process of introducing local retention conditions into the interferometric prediction box for deduplication constraint verification and filtering out the target dependent boxes is as follows: S501. Extract the overlap and encroachment ratio between the dominant scale box and the subordinate scale boxes; compare the overlap and encroachment ratio with the local preservation conditions under extreme value constraints, and filter to obtain the target subordinate boxes. Specifically, the method of introducing local preservation conditions into the interferometric prediction boxes for deduplication constraint verification, and filtering out target subordinate boxes that are not swallowed up by the dominant scale box, is as follows: Extract the background spatial boundary corresponding to the dominant scale frame and the core spatial boundary corresponding to the subordinate scale frame; Perform an intersection spatial operation between the background space boundary and the core space boundary to obtain the spatial overlap region; By using the overlapping area of ​​the spatially overlapping region to perform spatial proportion calculation on the basic area occupied by the subordinate scale frame, the overlapping encroachment ratio is obtained. Extraction criteria to construct local retention criteria output by the condition building module; By comparing the overlap encroachment ratio with the local preservation condition using extreme value constraints, the relationship between the overlap encroachment ratio and the local preservation condition is determined: If the overlap and encroachment ratio is strictly less than the local preservation condition, it is determined that the subordinate scale box has broken through the deduplication and swallowing limit of the dominant scale box, the spatial coordinate sequence of the subordinate scale box is preserved, and the subordinate scale box is marked as the target subordinate box. If the overlap and encroachment ratio is greater than or equal to the local retention condition, it is determined that the subordinate scale box is physically swallowed by the dominant scale box. The spatial coordinate sequence of the subordinate scale box is stripped, the subordinate scale box is marked as a swallowed and removed box, and the swallowed and removed box is prevented from entering the summary output process. The process of summarizing the target dependent boxes retained after constraint verification and the independent prediction boxes that did not interfere with the effective prediction boxes, and outputting the defect category and localization result of the phase map is as follows: S502. The target subordinate boxes, dominant scale boxes, and independent prediction boxes are summarized as the final detected targets; the category tendency carried by the final detected targets is confirmed as the final classification result, and the final detected targets are encapsulated as a global defect topology that presents the spatial distribution pattern of partial discharge. Specifically, the final defect category and location results are summarized and output as follows: Extract the target's subordinate bounding boxes, dominant scale boxes, and independent predicted bounding boxes. The target's subordinate bounding boxes, dominant scale boxes, and independent predicted bounding boxes are collectively referred to as the final detected target. The initial class tendency and extreme coordinates of the finally detected target are analyzed. Once it is determined that the final detected target has passed the deduplication constraint verification, the initial category tendency is formally confirmed as the defect classification label for the final detected target; The defect classification label and the target extreme value coordinates are encapsulated at the underlying data level to obtain a single defect node; Global spatial correlation is performed on all individual defect nodes contained in the phase map to generate a global defect topology that presents the spatial distribution pattern of partial discharge. S503. Output the global defect topology as the final defect category and location result to the terminal.

[0029] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A deep learning-based partial discharge (PRPD) phase map identification system for distribution network cables, characterized in that, Includes the following modules: Feature extraction module: It is used to input the phase map to be measured into a preset target detection network for multi-scale feature extraction to obtain an initial prediction box with spatial coordinates; based on the preset confidence conditions, the initial prediction box is cleaned to obtain an effective prediction box with discharge localization features. Subordinate filtering module: used to perform spatial overlap analysis on valid prediction boxes and extract interference prediction boxes with enclosing interference phenomena; based on the geometric inclusion relationship between interference prediction boxes, the interference prediction boxes are divided into dominant scale boxes on the periphery and subordinate scale boxes on the inside. Feature analysis module: used to extract the visual feature matrices of the dominant and subordinate scale boxes; utilizing... The visual feature matrix is ​​used to calculate the phase domain distribution gradient of the subordinate scale box relative to the dominant scale box; the phase domain distribution gradient is quantized locally to obtain the independent significance of the subordinate scale box independence. Condition construction module: used to evaluate the objective existence state of discharge characteristics within the subordinate scale box based on independent significance; dynamically relax the global filtering criteria faced by the subordinate scale box in the non-maximum suppression stage according to the objective existence state, and construct local retention conditions for the subordinate scale box.

2. The deep learning-based partial discharge (PRPD) phase map identification system for distribution network cables according to claim 1, characterized in that: The process of cleaning the parameters is as follows: Extract background hash parameters from the phase map of the test subject, detached from the dense texture feature region; Extract the initial recognition lower bound of the object detection network during the validation phase, and use the initial recognition lower bound as the confidence filtering benchmark. The bias compensation amount for benchmark compensation is extracted based on the background hash parameters. The bias compensation amount is used to perform numerical superposition correction on the confidence filtering benchmark to obtain a dynamic adaptive value; A pre-defined category addressing calibration is introduced to extract the initial confidence level bound to the candidate discharge category from the category probability parameters; A filtering logic for discharge location features is constructed based on the initial confidence level and dynamic adaptive value. The initial prediction boxes are then filtered to obtain effective prediction boxes with discharge location features.

3. The deep learning-based partial discharge (PRPD) phase map identification system for distribution network cables according to claim 2, characterized in that: The initial prediction box is obtained as follows: The phase map to be measured is divided into feature-sensing grids with multiple resolution levels according to spatial dimensions; Extract the pixel distribution parameters mapped inside the feature-aware grid; Based on pixel distribution parameters and a preset local scanning window, the texture density features of the feature-aware mesh at different resolution levels are calculated. Based on the dense texture features, the size regression calculation is performed on the preset prior anchor boxes of the target detection network to generate the mapping boundary; Extract the coordinates of the two-dimensional vertices anchored in the two-dimensional plane by the mapping boundary; extract the category probability parameters of classification tendency through the attribute classification layer of the object detection network; encapsulate the two-dimensional vertex coordinates and category probability parameters into low-level data to obtain the initial prediction box with spatial coordinates.

4. The deep learning-based partial discharge (PRPD) phase map identification system for distribution network cables according to claim 1, characterized in that: The method for dividing the subordinate scale box is as follows: Extract the lateral and longitudinal limit coordinates of the interferometric prediction box; perform same-direction limit constraint comparison on the lateral and longitudinal limit coordinates to divide the interferometric prediction box into the dominant scale box on the periphery and the subordinate scale box on the inside.

5. The deep learning-based partial discharge (PRPD) phase map identification system for distribution network cables according to claim 1, characterized in that: The process of obtaining the interference prediction box is as follows: Extract the coordinates of the diagonal vertices of the valid prediction boxes; reduce the dimensionality of the diagonal vertex coordinates to an axial boundary sequence; By comparing the staggered distribution results of the axial boundary sequences, interference prediction boxes with enveloping interference phenomena are selected.

6. The deep learning-based partial discharge (PRPD) phase map identification system for distribution network cables according to claim 1, characterized in that: The method for performing the local feature quantization is as follows: Extract the gradient polarity of the phase domain distribution gradient; perform trend alignment analysis between the gradient polarity and the prior polarity rule to obtain the independent significance.

7. The deep learning-based partial discharge (PRPD) phase map identification system for distribution network cables according to claim 1, characterized in that: The method for performing the trend alignment analysis is as follows: The source-derived feature frequency difference value is used to calculate the gradient of the phase domain distribution. The sign of the numerical value carried by the frequency difference of the feature is analyzed, and the sign of the numerical value is used as the gradient polarity representing the direction of frequency jump. Extract the pre-set inner density and outer sparseness rules of distribution network cables, and take the inner density and outer sparseness rules as the a priori polarity rules; Perform a logical consistency check between the gradient polarity and the prior polarity rule to determine whether the gradient polarity exhibits convergent polarity or divergent polarity: If the sign of the numerical value is positive, then it is determined that a linear convergence operation is performed on the gradient distribution in the phase domain, and the numerical result after the linear convergence operation is used as the independent significance. If the sign of the numerical value is negative or absolute zero, the independent significance will be forcibly set to zero.

8. The deep learning-based partial discharge (PRPD) phase map identification system for distribution network cables according to claim 1, characterized in that: The local preservation condition is constructed as follows: Analyze the numerical activation features of independent significance in the numerical space; The objective existence status of discharge characteristics within the subordinate scale frame is evaluated based on numerical activation features; The target detection network is extracted to determine the spatial overlap ratio of overlapping redundancy in the non-maximum suppression stage. The spatial overlap ratio is used as the regular suppression benchmark and the regular suppression benchmark is used as the global filtering criterion. Based on the objectively existing state, the execution rules are reconstructed to extract the suppression exemption factor for the global filtering criteria; By using the suppression exemption factor to relax the filtering scale of the global filtering criteria, local retention conditions are obtained.

9. The deep learning-based partial discharge (PRPD) phase map identification system for distribution network cables according to claim 1, characterized in that: It also includes the following modules: Constraint verification module: This module is used to introduce the interference prediction box into the local retention condition for deduplication constraint verification, and to filter out the target dependent boxes; it summarizes the target dependent boxes retained after constraint verification and the independent prediction boxes that have not interfered in the effective prediction boxes, and outputs the defect category and location results of the phase map.

10. The deep learning-based partial discharge (PRPD) phase map identification system for distribution network cables according to claim 9, characterized in that: The method for outputting the defect category and location result is as follows: Extract the overlap and encroachment ratio between the dominant scale box and the subordinate scale boxes; compare the overlap and encroachment ratio with the local preservation conditions under extreme value constraints, and filter to obtain the target subordinate boxes; The target's dependent boxes, dominant scale boxes, and independent prediction boxes are combined as the final detected target. The category tendency carried by the final detected target is confirmed as the final classification result, and the final detected target is encapsulated as a global defect topology that presents a partial discharge spatial distribution pattern. The global defect topology is used as the final defect category and location result for terminal output.

Images