A GIS device defect identification model training method and defect identification method

Abstract

This invention relates to the field of power equipment monitoring technology, specifically providing a training method and a defect identification method for a GIS equipment defect identification model. The training method includes: using source domain data and target domain data as training samples; extracting and fusing multimodal features from the training samples using a feature extraction module to obtain a fused feature vector; applying domain adversarial adaptation to the fused feature vector using a domain adversarial adaptation module to obtain the domain discrimination result of the training samples; using a prototype metric diagnosis module to measure the distance between the query set samples and multiple prototype centers in the training samples to obtain the defect category result of the query set samples; and determining the total loss function value based on the domain discrimination result and the defect category result to train the GIS equipment defect identification model. This invention solves the problems of scarce GIS field defect samples and domain offset between simulation data and measured data, achieving high-precision and strong generalization defect identification under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring technology, and in particular to a training method and a defect identification method for a GIS equipment defect identification model. Background Technology

[0002] Gas-insulated gas-insulated switchgear (GIS) has become a core component in modern high-voltage and ultra-high-voltage power transmission and transformation systems due to its compact structure, high reliability, and low maintenance requirements. It is widely used in urban substations, hydropower stations, and other locations with stringent requirements for land area and operational safety. Its operational stability directly affects the power supply security and reliability of the entire power grid. As smart grid construction evolves towards deeper digitalization and intelligence, higher demands are placed on the condition monitoring and fault diagnosis of power equipment. Against this backdrop, deep learning-based GIS partial discharge (PD) pattern recognition technology, with its powerful automatic feature extraction and complex pattern recognition capabilities, has gradually replaced traditional methods relying on manual experience and threshold judgments, becoming the mainstream trend in intelligent fault diagnosis.

[0003] However, applying this advanced deep learning technology to real-world engineering scenarios has encountered a severe data bottleneck. Due to the sophisticated manufacturing processes and extremely high insulation levels of GIS equipment, the probability of serious insulation failures is extremely low. This directly results in a very limited number of effective partial discharge samples that can be collected at the operating equipment site to characterize real core defects (such as surface discharge of insulators, tiny metal particles, and floating potential discharge), and the cost of acquiring these samples is extremely high. This objective reality of "scarce field samples" creates a sharp contradiction with the "massive amounts of labeled data" typically required for deep learning model training, constituting a typical "small sample" learning problem. To address the data shortage, the industry currently commonly uses physical simulation software (such as CST and COMSOL) for electromagnetic simulation, or builds scaled-down models in laboratories to simulate typical defects, generating a large amount of data for initial model training. However, this method has inherent limitations: the simulation or laboratory environment is highly simplified and idealized, while the real substation environment is complex and variable, filled with electromagnetic interference of various frequency bands, signal attenuation and distortion during propagation within the GIS cavity, and multipath effects. These factors cause a systematic difference in the statistical feature distribution between simulation data and field-measured data, i.e., a significant "domain shift". If a model trained on simulation data is directly deployed in the field, the diagnostic accuracy will drop sharply and the generalization ability will be severely insufficient because the learned feature distribution is inconsistent with the target scene.

[0004] Some researchers have attempted to alleviate these problems by introducing transfer learning or feature engineering methods. For example, patent application CN118468136A uses a ResNet-based transfer learning model, pre-trained on the ImageNet dataset and then transferred to the GIS partial discharge identification task, integrating an attention mechanism into the network to enhance feature focusing ability. However, this approach has the following shortcomings: First, the source domain ImageNet is a general image dataset with very weak correlation to the physical features of GIS partial discharge maps, belonging to "distant domain transfer," making it difficult to efficiently adapt pre-trained knowledge to the discharge diagnosis task; second, this approach does not include any algorithmic modules designed to actively reduce the difference in feature distribution between the source and target domains, failing to address the domain offset problem between simulation data and field data; third, its classifier still uses a traditional Softmax fully connected layer, which is prone to overfitting in small sample scenarios. Another example is patent CN112949497A, which uses an improved generalized regression neural network for partial discharge pattern recognition, manually extracting statistical features such as skewness and steepness from the PRPD map and performing dimensionality reduction through principal component analysis. This approach relies entirely on expert experience to design features, resulting in a fixed feature extraction process and shallow information dimensionality. It cannot adaptively learn deeper discriminative features, and the model is trained solely on historical field data without considering the use of simulation data to expand knowledge, thus limiting its recognition ability under small sample conditions. In summary, existing technologies have failed to systematically solve the two core coupling problems of "scarcity of field defect samples" and "significant domain shift between simulation and field data," leading to insufficient generalization ability of the model in practical engineering applications. Summary of the Invention

[0005] The purpose of this invention is to overcome the two major pain points commonly found in existing technologies: "scarcity of GIS field defect samples" and "domain offset between simulation data and field measured data", and to provide a training method and defect identification method for a GIS equipment defect identification model.

[0006] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0007] This invention provides a method for training a GIS equipment defect identification model, comprising:

[0008] Source domain data and target domain data are used as training samples; wherein, the source domain data is a simulated signal of GIS equipment defects, the target domain data is a GIS equipment defect signal collected on-site at the substation, and the target domain data packet contains labeled data and unlabeled data;

[0009] The training samples are subjected to multimodal feature extraction and fusion using the feature extraction module in the GIS equipment defect identification model to obtain a fused feature vector.

[0010] Using the domain adversarial adaptation module in the GIS equipment defect identification model, the fused feature vector is subjected to domain adversarial adaptation through gradient inversion to obtain the domain discrimination result of the training sample;

[0011] Using the prototype measurement and diagnosis module in the GIS equipment defect identification model, based on the labeled data in the target domain data, prototype centers for multiple defect categories are constructed, and distance measurement is performed on the query set samples in the training samples to obtain the defect category results of the query set samples.

[0012] The total loss function value is determined based on the domain discrimination result and the defect category result, and the GIS equipment defect identification model is trained based on the total loss function value.

[0013] In some embodiments of the present invention, before performing multimodal feature extraction and fusion on the training samples through the feature extraction module in the GIS equipment defect identification model to obtain the fused feature vector, the method further includes:

[0014] The source domain data is used as pre-training samples;

[0015] The feature extraction module and the prototype measurement and diagnosis module are pre-trained using the pre-trained samples.

[0016] In some embodiments of the present invention, the training samples are subjected to multimodal feature extraction and fusion through the feature extraction module in the GIS equipment defect identification model to obtain a fused feature vector, including:

[0017] The training samples are divided into multiple phase windows according to the power frequency cycle. The discharge amplitude and discharge number in each phase window are counted to generate a phase-resolved partial discharge map.

[0018] The training samples are subjected to time-frequency transformation using continuous wavelet transform to generate a continuous wavelet transform time-frequency map.

[0019] The phase-resolved partial discharge map and the continuous wavelet transform time-frequency map are respectively input into one channel branch of a dual-channel convolutional neural network for feature extraction, and the two output first features are concatenated to obtain a concatenated feature vector; wherein, the dual-channel convolutional neural network shares weights;

[0020] The concatenated feature vector is input into a deep convolutional neural network for feature extraction, and the output second feature is subjected to global average pooling to obtain a fused feature vector.

[0021] In some embodiments of the present invention, the domain adversarial adaptation module includes a domain discriminator and a gradient inversion layer;

[0022] Using the domain adversarial adaptation module in the GIS equipment defect identification model, the fused feature vector is subjected to domain adversarial adaptation through gradient inversion to obtain the domain discrimination result of the training samples, including:

[0023] The fused feature vector is input into the domain discriminator to perform binary classification on the fused feature vector, and outputs the probability value of the training sample belonging to the target domain as the domain discrimination result;

[0024] The gradient inversion layer is provided on the back propagation path of the domain discriminator. The gradient inversion layer keeps the fused feature vector unchanged during forward propagation and multiplies the gradients it passes through by a negative constant during back propagation.

[0025] In some embodiments of the present invention, based on the labeled data in the target domain data, prototype centers for multiple defect categories are constructed, including:

[0026] In each round of training, at least one support set sample is randomly selected from the labeled data of the target domain data for each defect category;

[0027] Based on the arithmetic mean of the fused feature vectors of the support set samples for each defect category, construct the prototype center for the corresponding defect category.

[0028] In some embodiments of the present invention, the query set sample includes the source domain data and the remaining samples in the labeled data of the target domain data that were not selected as support set samples;

[0029] Perform distance metric calculations on the query set samples in the training samples to obtain the defect category results of the query set samples, including:

[0030] The Euclidean distance between the fused feature vector of the query set sample and the prototype center of each type of defect is calculated to obtain the distance value between the query set sample and each prototype center.

[0031] After taking the negative value of the distance, normalization is performed using the Softmax function to obtain the probability distribution of the query set samples belonging to various types of defects;

[0032] The defect category with the highest probability is selected as the defect category result of the query set sample.

[0033] In some embodiments of the present invention, the total loss function value is calculated as follows:

[0034] ;

[0035] in, Represents the total loss function; Indicates the loss from defect classification; Indicate the loss by representing the domain; This represents the balance coefficient.

[0036] Secondly, the present invention also provides a method for identifying defects in GIS equipment, comprising:

[0037] Real-time acquisition of UHF signal data from GIS equipment;

[0038] The feature extraction module in the GIS equipment defect identification model is used to extract and fuse multimodal features from the UHF signal data to obtain a fused feature vector.

[0039] Using the prototype measurement and diagnosis module in the GIS equipment defect identification model, the distance between the fused feature vector and the prototype centers of multiple pre-stored defect categories is measured to obtain the defect identification result of the GIS equipment.

[0040] The GIS equipment defect identification model is pre-trained using the training method described in the above embodiments.

[0041] Thirdly, the present invention also provides an electronic device, comprising: a processor, and a memory storing a program, the program including instructions that, when executed by the processor, cause the processor to perform the training method or defect identification method described above.

[0042] Fourthly, the present invention also provides a non-transitory machine-readable medium storing computer instructions for causing the computer to perform the training method or defect identification method described above.

[0043] Compared with the prior art, the above-described technical solution of the present invention has the following advantages:

[0044] This invention systematically solves two core challenges in intelligent diagnosis of partial discharge in GIS: the scarcity of field defect samples and the domain shift between simulation and measured data. This is achieved by constructing an integrated end-to-end deep learning framework encompassing multimodal fusion, domain adversarial alignment, and prototype metric diagnosis. First, addressing the lack of field samples, this invention abandons the traditional Softmax classification paradigm that relies on massive amounts of labeled data. Instead, it introduces a prototype network architecture based on metric learning, transforming the classification task into a distance metric problem in the feature space. Only a very small number of typical defect samples are needed as a support set to establish a stable classification standard, fundamentally overcoming the data bottleneck of model training under small sample conditions. Second, addressing the domain shift caused by the inconsistent distribution of simulation and measured data, this invention embeds a gradient inversion layer between the feature extraction module and the domain discriminator. Through an adversarial training mechanism, it actively compares the feature distributions of simulation and real defect signals, forcing the feature extraction module to ignore non-essential factors such as environmental noise and propagation paths, extracting common physical features with domain invariance. This significantly improves the model's generalization ability and anti-interference performance in complex substation environments.

[0045] This invention overcomes the limitations of single-modal input by constructing a multimodal feature fusion network. It simultaneously extracts the phase statistical features of phase-resolved partial discharge (PD) spectra and the transient energy features of continuous wavelet transform time-frequency plots. This achieves deep complementarity and comprehensive characterization of the statistical laws of discharge signals and high-frequency transient details, effectively improving the model's accuracy in identifying similar defects. Furthermore, this invention creatively selects electromagnetic simulation data consistent with the physical mechanism of the target domain as the transfer source domain, realizing a paradigm shift from "far-domain transfer" to "near-domain transfer." This allows the pre-trained model to possess initial parameters more consistent with the physical essence of PD, effectively avoiding negative transfer and greatly improving the model's convergence speed and the targeting of feature extraction. Attached Figure Description

[0046] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other embodiments based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating a training method for a GIS equipment defect identification model provided in an embodiment of the present invention;

[0048] Figure 2 This is a schematic diagram of the structure of the GIS equipment defect identification model in an embodiment of the present invention;

[0049] Figure 3This is a schematic diagram of the feature extraction module of the GIS equipment defect identification model in this embodiment of the invention;

[0050] Figure 4 This is a schematic diagram of the domain adversarial adaptation module of the GIS equipment defect identification model in this embodiment of the invention;

[0051] Figure 5 This is a schematic diagram of the prototype measurement and diagnosis module of the GIS equipment defect identification model in this embodiment of the invention;

[0052] Figure 6 This is a flowchart illustrating a method for identifying defects in GIS equipment provided in an embodiment of the present invention;

[0053] Figure 7 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0054] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0055] like Figure 1 As shown, this embodiment of the invention provides a training method for a GIS equipment defect identification model. Figure 1 This is a flowchart illustrating the training method for a GIS equipment defect identification model. This flowchart only shows the logical sequence of the method in this embodiment; however, in other possible embodiments of the invention, different methods may be used, provided there are no conflicts. Figure 1 Complete the steps shown or described in the order indicated.

[0056] See Figure 1 The method of this invention specifically includes the following steps:

[0057] Step S101: Use source domain data and target domain data as training samples.

[0058] See Figure 2In the data input layer of the GIS equipment defect identification model, two types of data streams are received simultaneously as training samples. The first type is source domain data, which consists of simulated signals of GIS equipment defects. These signals mainly originate from GIS physical models constructed using electromagnetic simulation software such as CST or COMSOL, as well as simulated experimental data from high-voltage shielded rooms in laboratories. This data is large in volume and accurately labeled, covering ultra-high frequency (UHF) signals of various typical defects, such as insulator air gaps, metal particles, and floating potentials, at different voltage levels and locations. The second type is target domain data, which consists of GIS equipment defect signals collected at the substation site, i.e., real signals collected by the substation's on-site GIS online monitoring system. This data is usually mixed with complex environmental noise, and very few samples contain clear fault labels, falling into the category of small samples.

[0059] In this embodiment of the invention, during the data preparation stage, source domain data generation relies on CST Microwave Studio electromagnetic simulation software to construct a three-dimensional model of a typical 252kV GIS bay, setting four operating conditions including "insulator internal air gap," "busbar particles," "floating potential," and "normal state." For each operating condition, the defect size (e.g., air gap diameter 0.5mm-2mm) and location coordinates are changed to perform ultra-high frequency signal propagation simulation, generating a total of 2000 sets of tagged clean simulation signals. Target domain data is sourced from the historical database of a substation GIS online monitoring system. After manual review and screening by experts, only 5 typical field-measured signals are selected for each defect category, totaling 20 sets of data, as a minimal sample support set.

[0060] Step S102: Through the feature extraction module in the GIS equipment defect identification model, multimodal feature extraction and fusion are performed on the training samples to obtain a fused feature vector.

[0061] Source domain data and target domain data are used as training samples and fed in parallel into the feature extraction module of the GIS equipment defect identification model. See also... Figure 3 The feature extraction module performs multimodal feature extraction and fusion on the training samples to obtain a fused feature vector, specifically including the following steps S1021 to S1024:

[0062] Step S1021: Divide the training samples into multiple phase windows according to the power frequency cycle, count the discharge amplitude and discharge number in each phase window, and generate a phase-resolved partial discharge map.

[0063] Step S1022: Perform time-frequency transformation on the training samples using continuous wavelet transform to generate a continuous wavelet transform time-frequency graph.

[0064] The training samples are one-dimensional UHF time series signals. In order to fully capture the physical characteristics of the signal data, this invention transforms the training samples into phase-resolved partial discharge (PRPD) maps and continuous wavelet transform (CWT) time-frequency maps, respectively, to extract defect features from two complementary dimensions of statistics and transients, providing complementary feature inputs for the subsequent neural network.

[0065] In this embodiment of the invention, all training samples are uniformly truncated to a 20ms time window (corresponding to a 50Hz power frequency cycle). The statistical analysis module calculates the discharge amplitude and number of discharges for each phase, generating a resolution of [resolution value missing]. Phase-resolved partial discharge grayscale image; simultaneously, continuous wavelet transform is performed using Morlet wavelets, with a scale sequence from 1 to 64, generating a size of... RGB time-frequency graph.

[0066] Construction of Phase-Resolved Partial Discharge Map: Partial discharge exhibits significant power frequency phase correlation. A power frequency cycle (20ms) is divided into 360 phase windows, each corresponding to 1 degree of power frequency phase. The discharge amplitude and number of discharges within each phase window are statistically analyzed. By accumulating signals from multiple cycles, a PRPD grayscale map with a resolution of 360×128 is generated. This map clearly reflects the statistical differences between different defect types. For example, metal particle discharge is mainly concentrated near the power frequency peak phase, while air gap discharge inside the insulator exhibits a wider phase distribution.

[0067] Construction of the time-frequency plot using continuous wavelet transform: To capture the frequency components and energy decay characteristics of a single discharge pulse over time, the original UHF signal is processed using continuous wavelet transform (CWT). The Morlet wavelet is selected as the mother wavelet, and its transform formula is as follows:

[0068] ;

[0069] in, This represents the training samples, i.e., the time series of the original UHF signal; Represents the Morlet wavelet function; The scaling factor is used to control the frequency resolution of the wavelet. In this embodiment of the invention, the scaling sequence is set to 1 to 64. This represents the translation factor, used to control the position of the wavelet on the time axis to locate the transient pulse moments in the signal; Represents the conjugate complex number; The amplitude of the wavelet coefficients reflects the energy intensity of the signal at that time-frequency position.

[0070] The time-frequency matrix obtained after transformation Normalized and mapped to an RGB image, generating a size of The RGB time-frequency plot. Compared to the short-time Fourier transform, the continuous wavelet transform has better multi-resolution analysis capabilities when processing sudden, non-stationary signals such as partial discharge in GIS, and can simultaneously preserve the temporal details of high-frequency signals and the frequency characteristics of low-frequency signals.

[0071] Step S1023: Input the phase-resolved partial discharge map and the continuous wavelet transform time-frequency map into one channel branch of the dual-channel convolutional neural network for feature extraction, and concatenate the two output first features to obtain the concatenated feature vector; wherein, the dual-channel convolutional neural network shares weights.

[0072] Phase-resolved partial discharge maps and continuous wavelet transform time-frequency maps are fed into a weighted dual-channel convolutional neural network, with each channel branch used for feature extraction. This dual-channel convolutional neural network serves as a general feature encoder, responsible for mapping high-dimensional image data into low-dimensional, compact feature vectors.

[0073] The two output features are concatenated, and the resulting tensor combines statistical regularities with time-frequency details.

[0074] Step S1024: Input the concatenated feature vector into a deep convolutional neural network for feature extraction, and perform global average pooling on the output second feature to obtain a fused feature vector.

[0075] The deep convolutional neural network in this embodiment of the invention employs a ResNet-18 or ResNet-50 structure. ResNet (Residual Network) addresses the gradient vanishing and degradation problems in deep network training by introducing skip connections, enabling the network to stably reach deeper layers. ResNet-18 contains 18 convolutional layers, a relatively lightweight structure suitable for edge deployment scenarios with limited computing resources; ResNet-50 contains 50 convolutional layers, employing a bottleneck structure to control computation while maintaining high feature extraction capabilities, making it suitable for scenarios with higher accuracy requirements. This embodiment of the invention uses a Global Average Pooling (GAP) layer at the network's end instead of a traditional fully connected layer, reducing each feature map to a single numerical value, significantly reducing the number of model parameters, effectively preventing overfitting in small-sample training, while preserving global information of spatial features, making feature extraction more robust. Finally, the feature extraction module outputs a fused feature vector, which serves as the common input basis for subsequent domain adversarial and classification diagnostics.

[0076] Step S103: Using the domain adversarial adaptation module in the GIS equipment defect identification model, the fused feature vector is subjected to domain adversarial adaptation through gradient inversion to obtain the domain discrimination result of the training samples.

[0077] To address the "domain offset" problem caused by the inconsistency between simulation data and field data distribution, embodiments of this invention have designed the following... Figure 4 The domain adversarial adaptation module is shown. The core idea of ​​this module is to construct a "zero-sum game," ensuring that the features extracted by the feature extraction module contain sufficient classification information while minimizing domain information, thus preventing the domain discriminator from distinguishing between source and target domain data. The module consists of a domain discriminator and a gradient reversal layer. The discriminator is a multilayer perceptron (MLP), whose input is the fused feature vector, and whose output is the probability that a training sample comes from the source or target domain, serving as the domain discrimination result.

[0078] In this embodiment of the invention, the loss function of the domain discriminator The binary cross-entropy loss is used, and the calculation formula is as follows:

[0079] ;

[0080] in, Indicates the first The true domain labels of the training samples are: source domain data with a value of 0 and target domain data with a value of 1. This indicates the number of training samples in a training batch. Indicates the first One training sample; Indicates the first The fusion feature vector corresponding to each training sample; Representation domain discriminator function.

[0081] To achieve adversarial training, this embodiment of the invention embeds a gradient reversal layer (GRL) between the feature extraction module and the domain discriminator. See also Figure 4 The gradient reversal layer performs an identity transformation during the forward propagation phase, meaning the input equals the output, without any further processing.

[0082] ;

[0083] During the backpropagation phase, the gradient inversion layer multiplies the gradient flowing through it by a negative constant. , represented as:

[0084] ;

[0085] in, Hyperparameters representing the control of the intensity of the confrontation; Represents the total loss function; This represents the fused feature vector output by the feature extraction module.

[0086] When the optimizer updates network parameters, the domain discriminator optimizes to minimize the domain classification error, while the gradient received by the feature extraction module is inverted, and its update direction is opposite to the optimization direction of the domain discriminator, that is, it adjusts the parameters in the direction of maximizing the domain classification error. Through this adversarial training, the fused feature vector finally extracted by the feature extraction module will eliminate "domain-specific features" that are strongly correlated with the environment and simulation parameters, and retain only "domain-invariant features" that reflect the physical nature of the defects, thereby effectively improving the generalization ability of the GIS equipment defect identification model in complex field environments.

[0087] Step S104: Using the prototype measurement and diagnosis module in the GIS equipment defect identification model, construct prototype centers for multiple defect categories based on the labeled data in the target domain data, and perform distance measurement on the query set samples in the training samples to obtain the defect category results of the query set samples.

[0088] To address the problem of extremely limited labeled samples in the field, this invention abandons the traditional Softmax fully connected classification layer, which has a large number of parameters and is prone to overfitting with small sample sizes. Instead, it employs a method such as... Figure 5 The prototype metric diagnostic module shown performs metric learning.

[0089] Based on the labeled data in the target domain, prototype centers for multiple defect categories are constructed, including: defining the task as... The problem is that in each round of training, from the labeled data of the target domain, for Defect categories are randomly selected A support set of samples is used; based on the arithmetic mean of the fused feature vectors of the support set samples for each defect category, a prototype center for the corresponding defect category is constructed. This process does not involve parameter updates, but only simple vector operations, thus it is extremely efficient and stable.

[0090] All source domain data is labeled data. The remaining samples from the labeled data of both the source and target domains that were not selected as support set samples are used as query set samples to form the query set.

[0091] Distance metrics are applied to the query set samples in the training samples to obtain the defect category results for the query set samples, including:

[0092] The Euclidean distance between the fused feature vectors of the query set samples and the prototype centers of various defects is calculated to obtain the distance values ​​between the query set samples and each prototype center. The calculation formula is as follows:

[0093] ;

[0094] in, Represents a query set sample With the Prototype Center of Defect Categories The Euclidean distance between them.

[0095] The smaller the distance, the higher the similarity between the query set sample and the defect of that category. To output standardized probability results, the distance is negatively denoted and then subjected to Softmax normalization to obtain the probability distribution of the query set sample belonging to each category of defects. The defect category with the highest probability is selected as the defect category result of the query set sample.

[0096] The formula for calculating Softmax normalization is as follows:

[0097] ;

[0098] in, Represents a query set sample Belongs to the The probability of a defect category.

[0099] This invention also provides an online update strategy. When maintenance personnel confirm a new defect sample and its true label on-site, for example, by disassembly and inspection confirming it to be a high-frequency signal of metal particles, the GIS equipment defect identification model can dynamically update the prototype center of that category without retraining the entire neural network. The update formula uses a moving average method, and the update formula is as follows:

[0100] ;

[0101] in, This indicates the updated prototype center; This indicates the prototype center before the update; This represents a new defect sample; Indicating the update rate, embodiments of the present invention will... Set it to 0.1.

[0102] This update strategy endows the model with the ability to continuously learn, allowing it to accumulate more field experience over time, resulting in more accurate prototypes and higher diagnostic accuracy. Without retraining, defect prototype centers can be dynamically added or updated simply by calculating the mean of new sample features. This enables the system to continuously improve itself with the accumulation of field operation data, quickly adapting to new defects or new voltage level equipment, significantly reducing the total lifecycle maintenance cost and model iteration cycle.

[0103] Step S105: Determine the total loss function value based on the domain discrimination result and the defect category result, and train the GIS equipment defect identification model based on the total loss function value.

[0104] The training of the GIS equipment defect identification model of the present invention is divided into two stages: the first stage is source domain pre-training, and the second stage is cross-domain joint training. Among them, steps S102 to S105 belong to the cross-domain joint training stage, and source domain pre-training is included before step S102.

[0105] Source domain pre-training uses only a large amount of source domain data as pre-training samples to pre-train the feature extraction module and the prototype measurement and diagnosis module. At this stage, the domain adversarial adaptation module is not enabled; only the defect classification loss is used to optimize the feature extraction module, giving it a preliminary ability to identify different partial discharge waveforms. The support set is randomly drawn from the source domain data, and the query set is the remaining portion of the source domain data that was not included in the support set.

[0106] In this embodiment of the invention, the defect classification loss is defined as the negative log-likelihood loss of the pre-trained samples:

[0107] ;

[0108] in, Represents a query set; Indicates a sample of the query set; This represents the true defect category label of the query set sample.

[0109] Cross-domain joint training inputs data from both the source and target domains. For a small number of labeled support samples in both the source and target domains, a defect classification loss is calculated. For all data, compute domain discriminant loss. Overall objective function for:

[0110] ;

[0111] in, This represents the balance coefficient, used to adjust the weight of the domain discrimination loss in the total loss.

[0112] The optimizer updates the network parameters through backpropagation based on the total loss. This is due to the presence of the gradient reversal layer. The decrease in the value actually drives the parameters of the feature extraction module to evolve towards the "confusion domain distribution", thereby achieving active alignment of the feature distributions of the source domain and the target domain.

[0113] The embodiments of the present invention also provide a GIS equipment defect identification method. The GIS equipment defect identification method is based on the GIS equipment defect identification model trained by the above embodiments, and realizes real-time and accurate diagnosis of newly acquired signals.

[0114] See Figure 6 The method of this invention specifically includes the following steps:

[0115] Step S201: Real-time acquisition of UHF signal data from GIS equipment.

[0116] UHF sensors deployed at substation sites collect UHF signal data in real time during the operation of GIS equipment. When an abnormal pulse is detected, a diagnostic process is triggered, and the captured signal segment is input into the pre-trained GIS equipment defect identification model.

[0117] Step S202: Using the feature extraction module in the GIS equipment defect identification model, perform multimodal feature extraction and fusion on the UHF signal data to obtain a fused feature vector.

[0118] A feature extraction module is used to extract and fuse multimodal features from the acquired UHF signals. This module employs the same preprocessing method as the training phase, synchronously converting the original signal into a phase-resolved partial discharge spectrum and a continuous wavelet transform time-frequency map, and then extracting a deep fusion feature vector through a convolutional neural network. This process achieves a comprehensive characterization of the statistical regularity and transient details of the discharge signal, providing an information-rich and highly discriminative feature foundation for subsequent classification.

[0119] Step S203: Using the prototype measurement and diagnosis module in the GIS equipment defect identification model, the distance between the fused feature vector and the prototype center of multiple pre-stored defect categories is measured to obtain the defect identification result of the GIS equipment.

[0120] The prototype metric diagnostic module uses Euclidean distance to measure the extracted fused feature vectors with pre-stored prototype centers of various defects, and then converts them into a probability distribution using Softmax to output the defect category and confidence score. This classification method requires no retraining or fine-tuning; diagnosis can be completed solely through distance calculation, resulting in high computational efficiency and adaptability to classification needs in small sample scenarios.

[0121] The GIS equipment defect identification method proposed in this invention achieves rapid and accurate identification of small sample defects through a pre-set prototype center, overcoming the constraint of scarce field samples. The model has eliminated the distribution difference between simulation and field data through a domain adversarial mechanism during the training phase, thus possessing excellent generalization ability in actual deployment. Based on the open architecture of the prototype measurement and diagnosis module, it supports online dynamic updates of the prototype center, enabling the system to continuously improve itself as field operation data accumulates without retraining.

[0122] This invention breaks down the data barriers between physical simulation and on-site monitoring through algorithmic means, and constructs a low-cost engineering application paradigm of "simulation-assisted measurement". It provides an efficient, implementable and self-evolving technical path for intelligent operation and maintenance of GIS equipment, and has significant engineering practical value and promotion prospects.

[0123] The embodiments of the present invention also provide a non-transitory machine-readable medium storing a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform a training method for a GIS equipment defect identification model or a GIS equipment defect identification method according to the embodiments of the present invention.

[0124] The embodiments of the present invention also provide a computer program product, including a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform a training method for a GIS equipment defect identification model or a GIS equipment defect identification method according to embodiments of the present invention.

[0125] Embodiments of this invention also provide an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, which, when executed by the at least one processor, causes the electronic device to perform a training method for a GIS device defect identification model or a GIS device defect identification method according to embodiments of this invention.

[0126] refer to Figure 7The present invention will now describe a structural block diagram of an electronic device that can serve as an embodiment of the present invention, serving as an example of a hardware device applicable to various aspects of the present invention. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present invention described and / or claimed herein.

[0127] like Figure 7 As shown, the electronic device includes a computing unit 101, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 102 or a computer program loaded from a storage unit 108 into a random access memory (RAM) 103. The RAM 103 may also store various programs and data required for the operation of the electronic device. The computing unit 101, ROM 102, and RAM 103 are interconnected via a bus 104. An input / output (I / O) interface 105 is also connected to the bus 104.

[0128] Multiple components in the electronic device are connected to I / O interface 105, including: input unit 106, output unit 107, storage unit 108, and communication unit 109. Input unit 106 can be any type of device capable of inputting information into the electronic device. Input unit 106 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of the electronic device. Output unit 107 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 108 may include, but is not limited to, disks and optical discs. Communication unit 109 allows the electronic device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, and / or wireless communication transceivers, such as Bluetooth devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0129] The computing unit 101 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 101 include, but are not limited to, CPUs, graphics processing units (GPUs), various special-purpose artificial intelligence (AI) computing units, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. The computing unit 101 performs the various methods and processes described above. For example, in some embodiments, the method embodiments of the present invention can be implemented as computer programs tangibly contained in a machine-readable medium, such as storage unit 108. In some embodiments, part or all of the computer program can be loaded and / or installed on an electronic device via ROM 102 and / or communication unit 109. In some embodiments, the computing unit 101 can be configured to perform the methods described above by any other suitable means (e.g., by means of firmware).

[0130] Computer programs for implementing the methods of embodiments of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0131] In the context of embodiments of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, or infrared systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0132] It should be noted that the term "comprising" and its variations used in the embodiments of this invention are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The modifications of "one" and "a plurality" mentioned in the embodiments of this invention are illustrative and not restrictive, and those skilled in the art should understand that unless explicitly indicated otherwise in the context, they should be understood as "one or more".

[0133] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this invention are subject to strict compliance with relevant laws, regulations, and regulatory requirements in their collection, storage, use, processing, transmission, provision, and disclosure, and adhere to the principles of legality, legitimacy, necessity, and good faith. The acquisition of relevant information and data is premised on the user's explicit consent or other legitimate reasons, and a clear and convenient authorization management approach is provided to the user, allowing the user to independently choose to consent, withdraw consent, or refuse to provide relevant information. For functions that rely on user information, if the user does not authorize or withdraws authorization, the corresponding technical function cannot be implemented, and the technical solution of this invention is not applicable in this scenario.

[0134] The steps described in the method embodiments provided by the present invention can be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of protection of the present invention is not limited in this respect.

[0135] The term "embodiment" in this specification refers to a specific feature, structure, or characteristic described in connection with an embodiment that may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily imply the same embodiment, nor does it imply independence or alternativeity from other embodiments. The various embodiments in this specification are described in a related manner, with reference to each other for similar or identical parts. In particular, for apparatus, device, and system embodiments, since they are substantially similar to method embodiments, the description is relatively simple, and relevant details are referred to in the description of the method embodiments.

[0136] The above embodiments merely illustrate several implementation methods of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of protection. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A training method for a GIS equipment defect identification model, characterized in that, include: Source domain data and target domain data are used as training samples; wherein, the source domain data is a simulated signal of GIS equipment defects, the target domain data is a GIS equipment defect signal collected on-site at the substation, and the target domain data packet contains labeled data and unlabeled data; The training samples are subjected to multimodal feature extraction and fusion using the feature extraction module in the GIS equipment defect identification model to obtain a fused feature vector. Using the domain adversarial adaptation module in the GIS equipment defect identification model, the fused feature vector is subjected to domain adversarial adaptation through gradient inversion to obtain the domain discrimination result of the training sample; Using the prototype measurement and diagnosis module in the GIS equipment defect identification model, based on the labeled data in the target domain data, prototype centers for multiple defect categories are constructed, and distance measurement is performed on the query set samples in the training samples to obtain the defect category results of the query set samples. The total loss function value is determined based on the domain discrimination result and the defect category result, and the GIS equipment defect identification model is trained based on the total loss function value.

2. The training method for the GIS equipment defect identification model according to claim 1, characterized in that, Before performing multimodal feature extraction and fusion on the training samples through the feature extraction module in the GIS equipment defect identification model to obtain the fused feature vector, the following steps are also included: The source domain data is used as pre-training samples; The feature extraction module and the prototype measurement and diagnosis module are pre-trained using the pre-trained samples.

3. The training method for the GIS equipment defect identification model according to claim 1, characterized in that, The feature extraction module in the GIS equipment defect identification model performs multimodal feature extraction and fusion on the training samples to obtain a fused feature vector, including: The training samples are divided into multiple phase windows according to the power frequency cycle. The discharge amplitude and discharge number in each phase window are counted to generate a phase-resolved partial discharge map. The training samples are subjected to time-frequency transformation using continuous wavelet transform to generate a continuous wavelet transform time-frequency map. The phase-resolved partial discharge map and the continuous wavelet transform time-frequency map are respectively input into one channel branch of a dual-channel convolutional neural network for feature extraction, and the two output first features are concatenated to obtain a concatenated feature vector; wherein, the dual-channel convolutional neural network shares weights; The concatenated feature vector is input into a deep convolutional neural network for feature extraction, and the output second feature is subjected to global average pooling to obtain a fused feature vector.

4. The training method for the GIS equipment defect identification model according to claim 1, characterized in that, The domain adversarial adaptation module includes a domain discriminator and a gradient inversion layer; Using the domain adversarial adaptation module in the GIS equipment defect identification model, the fused feature vector is subjected to domain adversarial adaptation through gradient inversion to obtain the domain discrimination result of the training samples, including: The fused feature vector is input into the domain discriminator to perform binary classification on the fused feature vector, and outputs the probability value of the training sample belonging to the target domain as the domain discrimination result; The gradient inversion layer is provided on the back propagation path of the domain discriminator. The gradient inversion layer keeps the fused feature vector unchanged during forward propagation and multiplies the gradients it passes through by a negative constant during back propagation.

5. The training method for the GIS equipment defect identification model according to claim 1, characterized in that, Based on the labeled data in the target domain data, prototype centers for multiple defect categories are constructed, including: In each round of training, at least one support set sample is randomly selected from the labeled data of the target domain data for each defect category; Based on the arithmetic mean of the fused feature vectors of the support set samples for each defect category, construct the prototype center for the corresponding defect category.

6. The training method for the GIS equipment defect identification model according to claim 5, characterized in that, The query set sample includes the source domain data and the remaining samples in the labeled data of the target domain data that were not selected as support set samples; Perform distance metric calculations on the query set samples in the training samples to obtain the defect category results of the query set samples, including: The Euclidean distance between the fused feature vector of the query set sample and the prototype center of each type of defect is calculated to obtain the distance value between the query set sample and each prototype center. After taking the negative value of the distance, normalization is performed using the Softmax function to obtain the probability distribution of the query set samples belonging to various types of defects; The defect category with the highest probability is selected as the defect category result of the query set sample.

7. The training method for the GIS equipment defect identification model according to claim 1, characterized in that, The total loss function value is calculated as follows: ； in, Represents the total loss function; Indicates the loss from defect classification; Indicate the loss by representing the domain; This represents the balance coefficient.

8. A method for identifying defects in GIS equipment, characterized in that, include: Real-time acquisition of UHF signal data from GIS equipment; The feature extraction module in the GIS equipment defect identification model is used to extract and fuse multimodal features from the UHF signal data to obtain a fused feature vector. Using the prototype measurement and diagnosis module in the GIS equipment defect identification model, the distance between the fused feature vector and the prototype centers of multiple pre-stored defect categories is measured to obtain the defect identification result of the GIS equipment. The GIS equipment defect identification model is pre-trained using the training method described in any one of claims 1 to 7.

9. An electronic device, comprising: A processor and a memory storing a program, characterized in that the program includes instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1 to 8.

10. A non-transitory machine-readable medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1 to 8.

Images