System for generating model for detecting partial discharge defect in gas-insulated switchgear, apparatus for defect detection , and method for defect detection

Abstract

The present invention comprises: a memory for storing phase-resolved partial discharge (PRPD) datasets of a gas-insulated switchgear received from an ultra-high frequency (UHF) sensor; and a processor for generating a defect detection model by using the stored PRPD datasets. To detect a partial discharge defect in a gas-insulated switchgear with high accuracy even with a small amount of training data, the processor generates a defect detection model by performing supervised contrastive learning on the basis of pre-collected PRPD datasets, and determines the type of partial discharge defect upon receiving an input of the PRPD data.

Description

A system for generating a partial discharge defect detection model for a gas-insulated switchgear, a defect detection device, and a defect detection method

[0001] The disclosed invention relates to a system for generating a partial discharge defect detection model for a gas-insulated switchgear, a defect detection device, and a defect detection method capable of detecting a partial discharge defect in a gas-insulated switchgear with high accuracy even with a small amount of learning data.

[0002] As power systems continue to evolve, electricity has become an essential element not only in manufacturing but also in everyday life. However, the use of high currents increases the risk of equipment explosions and system damage, including fire hazards. To mitigate these risks, maintaining stable system operation is essential. Gas-insulated switchgear (GIS), which uses gas to insulate and protect power system components, offers numerous advantages, including compact installation, flexible installation options, and ease of maintenance. Therefore, ensuring and maintaining the safety of GIS is crucial.

[0003] Partial discharge (PD) detection is crucial for ensuring the reliability and safety of power infrastructure. PD occurring in gas-insulated switchgear (GIS) can lead to serious accidents and insulation degradation due to insulation failures, making early PD detection crucial. Several technologies, including loop antennas, acoustic emissions, and various internal and external sensors, have been developed to detect PD. Ultra-high frequency (UHF) sensors can capture a wide range of frequencies and effectively reduce noise. PD signals can be expressed in two forms: time-resolved partial discharges (TRPDs) and phase-resolved partial discharges (PRPDs). Because TRPDs are significantly affected by signal attenuation and noise interference, PRPD patterns are used in conjunction with UHF sensors for fault diagnosis in gas-insulated switchgear (GIS).

[0004] Recent artificial intelligence technologies are being effectively applied to fault detection in GIS, utilizing various machine learning algorithms such as Support Vector Machine (SVM), Random Forest (RF), and k-Nearest Neighbors (kNN). A GIS fault classification algorithm combining SVM and Principal Component Analysis (PCA), computational reduction using kNN, PD signal classification using RF, and machine learning techniques for identifying and classifying incipient discharges in GIS have been proposed. Furthermore, an ensemble stacking technique has been introduced for PD diagnosis. However, traditional machine learning techniques have limitations in extracting advanced features and difficulty in selecting appropriate parameters.

[0005] To overcome the limitations of machine learning, deep neural networks (DNNs) capable of handling large amounts of data and complex calculations have been developed. However, the large number of nonlinear activation functions, layers, and neurons causes slow convergence and complexity in the structure. To address this, various studies using convolutional neural networks (CNNs) have been proposed. CNN architectures composed of various convolutional blocks and fully connected layers have demonstrated superior learning capabilities compared to traditional machine learning methods. However, fault diagnosis requires a large training dataset, and in GIS, it is difficult to secure a large amount of fault data across various environments and fault severities.

[0006] For the above reasons, one aspect of the disclosed invention is to provide a system for generating a partial discharge fault detection model for a gas-insulated switchgear, a fault detection device, and a fault detection method capable of detecting a fault in a gas-insulated switchgear based on a data augmentation technique and supervised contrastive learning to solve the problem of limited fault data.

[0007] A system for generating a partial discharge defect detection model of a gas-insulated switchgear according to one aspect of the disclosed invention comprises: a memory for storing a PRPD (Phase Resolved Partial Discharge) dataset of a gas-insulated switchgear received from an ultra-high frequency (UHF) sensor; And a processor for generating a defect detection model using the stored PRPD data set, wherein the processor generates a pair of propagated PRPD data sets by applying different data propagation techniques to each PRPD data for the PRPD data set, labels each pair of propagated PRPD data with a corresponding partial discharge defect type for each pair of propagated PRPD data for each pair of propagated PRPD data sets, divides each pair of propagated PRPD data sets into a pair of propagated training PRPD data sets and a pair of propagated testing PRPD data sets, performs pre-learning for the defect detection model using the pair of propagated training PRPD data sets, inputs the pair of propagated testing PRPD data sets into the defect detection model to determine a defect type, and adjusts a weight of the defect detection model based on a determination result to complete the learning.

[0008] The above pre-learning may be supervised contrastive learning.

[0009] The processor may convert a pair of propagated training PRPD data sets into a pair of feature vectors based on an encoder network, convert the pair of feature vectors into a pair of space vectors by applying a nonlinear transformation based on a projection head, and adjust weights of the encoder network and weights of the projection head to minimize a supervised contrastive loss function for the pair of space vectors.

[0010] The above defect detection model is the encoder network, and the weights of the adjusted encoder network can be applied to the defect detection model as fixed values.

[0011] The weights of the above defect detection model can be adjusted to minimize the cross entropy loss function.

[0012] The weights of the encoder network, the weights of the projection head, and the weights of the defect detection model can be adjusted based on the Adam optimization technique (Adaptive Moment Estimation).

[0013] The above different data augmentation techniques may be any two techniques selected from among a Gaussian noise technique, a Gaussian scaling technique, a random crop technique, and a phase shift technique.

[0014] The above partial discharge defect type may be any one of Corona, Floating, Particle, Void, and Noise.

[0015] A method for detecting a partial discharge defect in a gas-insulated switchgear according to one aspect of the disclosed invention comprises the steps of: generating a defect detection model by performing supervised contrastive learning based on a pre-collected PRPD data set; and receiving the PRPD data and determining the partial discharge defect type.

[0016] The step of generating the defect detection model comprises the steps of: generating a pair of propagated PRPD data sets by applying different data propagation techniques to each PRPD data set collected in advance; labeling each pair of propagated PRPD data with a corresponding defect type for each pair of propagated PRPD data sets, and dividing the pair of propagated PRPD data sets into a pair of propagated training PRPD data sets and a pair of propagated test PRPD data sets; converting the pair of propagated training PRPD data for the pair of propagated training PRPD data sets into a pair of feature vectors based on an encoder network; converting the pair of feature vectors into a pair of space vectors by applying a nonlinear transformation based on a projection head; adjusting the weights of the encoder network and the weights of the projection head to minimize a supervised contrastive loss function for the pair of space vectors; The step may include inputting the above pair of proliferated test PRPD data sets into the above defect detection model to determine the defect type, and adjusting the weights of the defect detection model based on the determination result to complete learning.

[0017] The above defect detection model is the encoder network, and the weights of the adjusted encoder network can be applied to the defect detection model as fixed values.

[0018] The weights of the above defect detection model can be adjusted to minimize the cross entropy loss function.

[0019] The weights of the encoder network, the weights of the projection head, and the weights of the defect detection model can be adjusted based on the Adam optimization technique (Adaptive Moment Estimation).

[0020] The above different data augmentation techniques may be any two techniques selected from among a Gaussian noise technique, a Gaussian scaling technique, a random crop technique, and a phase shift technique.

[0021] The above partial discharge defect type may be any one of Corona, Floating, Particle, Void, and Noise.

[0022] A partial discharge fault detection device of a gas-insulated switchgear according to one aspect of the disclosed invention is a partial discharge fault detection device of a gas-insulated switchgear that analyzes PRPD (Phase Resolved Partial Discharge) data received from an ultra-high frequency (UHF) sensor to determine a partial discharge fault type, the device may include: an artificial intelligence modeler that generates a fault detection model by performing supervised contrastive learning based on a pre-collected PRPD dataset; and a fault detection model that receives the PRPD data and determines a partial discharge fault type.

[0023] The artificial intelligence modeler comprises: a data augmentation unit that applies different data augmentation techniques to each PRPD data for the pre-collected PRPD data set to generate a pair of augmented PRPD data sets, labels each pair of augmented PRPD data with a corresponding defect type for each pair of the augmented PRPD data sets, and divides each pair of the augmented PRPD data sets into a pair of augmented training PRPD data sets and a pair of augmented testing PRPD data sets; an encoder network that converts a pair of augmented training PRPD data for the pair of augmented training PRPD data sets into a pair of feature vectors; a projection head that applies a nonlinear transformation to the pair of feature vectors to convert them into a pair of space vectors; and a pre-learning unit that adjusts weights of the encoder network and weights of the projection head to minimize a supervised contrastive loss function for the pair of space vectors. And a model optimization unit that inputs the pair of proliferated test PRPD data sets into the defect detection model to determine the type of defect, and adjusts the weight of the defect detection model based on the determination result to complete learning, wherein the defect detection model is the encoder network, and the weight of the adjusted encoder network can be applied to the defect detection model as a fixed value.

[0024] The above partial discharge defect type may be any one of Corona, Floating, Particle, Void, and Noise.

[0025] The above different data augmentation techniques may be any two techniques selected from among a Gaussian noise technique, a Gaussian scaling technique, a random crop technique, and a phase shift technique.

[0026] The weights of the encoder network, the weights of the projection head, and the weights of the defect detection model can be adjusted based on the Adam optimization technique (Adaptive Moment Estimation).

[0027] According to one aspect of the disclosed invention, a partial discharge defect detection model generation system, a defect detection device, and a defect detection method can be provided, which can generate a high-accuracy partial discharge defect detection model by multiplying a small amount of learning data.

[0028] Figure 1 illustrates a system for generating a partial discharge defect detection model according to one embodiment.

[0029] Fig. 2 illustrates a partial discharge defect detection device according to one embodiment.

[0030] Figure 3 illustrates components of an artificial intelligence modeler in a partial discharge defect detection device according to one embodiment.

[0031] Figure 4 schematically illustrates a defect detection model based on map contrast learning according to one embodiment.

[0032] Figure 5 illustrates an encoder network according to one embodiment.

[0033] Fig. 6 is a flowchart showing a partial discharge defect detection method according to one embodiment.

[0034] Fig. 7 is a flowchart showing more specifically the defect detection model generation step in a partial discharge defect detection method according to one embodiment.

[0035] Figure 8 is a diagram comparing the detection performance of a defect detection model based on map contrast learning according to one embodiment with another detection model.

[0036] Figure 9 is a diagram comparing the precision, recall, and F-1 score of a defect detection model based on supervised contrastive learning according to one embodiment with other detection models.

[0037] Throughout the specification, the same reference numerals denote the same components. This specification does not describe all elements of the embodiments, and any content that is general in the technical field to which the disclosed invention belongs or that overlaps between the embodiments is omitted. The terms 'part, module, element, block' used in the specification may be implemented in software or hardware, and depending on the embodiments, multiple 'parts, modules, elements, blocks' may be implemented as a single component, or a single 'part, module, element, block' may include multiple components.

[0038] Throughout the specification, when a part is said to be 'connected' to another part, this includes not only direct connection but also indirect connection, and indirect connection includes connection via a wireless communication network.

[0039] Additionally, when a part is said to 'include' a component, this does not mean that it excludes other components, but rather that it may include other components, unless otherwise specifically stated.

[0040] Throughout the specification, when we say that an element is located 'on' another element, this includes not only cases where the element is in contact with the other element, but also cases where another element exists between the two elements.

[0041] The terms first, second, etc. are used to distinguish one component from another, and the components are not limited by the aforementioned terms.

[0042] Singular expressions include plural expressions unless the context clearly indicates otherwise.

[0043] The identification codes for each step are used for convenience of explanation and do not describe the order of each step. Each step may be performed in a different order than specified unless the context clearly indicates a specific order.

[0044] The operating principle and embodiments of the disclosed invention will be described with reference to the attached drawings below.

[0045] In the following description, the term referring to the 'partial discharge defect detection model generation system of a gas-insulated switchgear' of the present invention may be abbreviated as 'defect detection model generation system' for convenience of explanation.

[0046] FIG. 1 illustrates a partial discharge defect detection model generation system (100) according to one embodiment.

[0047] Referring to FIG. 1, a defect detection model generation system (100) may include a memory (130), a processor (120), and a communication module (110).

[0048] The memory (130) can store a PRPD (Phase Resolved Partial Discharge) data set of a gas-insulated switchgear received from an ultra-high frequency (UHF) sensor.

[0049] Additionally, the memory (130) may include various algorithms that can be used to create a defect detection model, a command set for processing the algorithm, defect types, etc.

[0050] The processor (120) can create a defect detection model using the stored PRPD dataset.

[0051] More specifically, the processor (120) can apply different data augmentation techniques to each PRPD data set to generate a pair of augmented PRPD data sets.

[0052] Here, the different data augmentation techniques can be any two techniques selected randomly from among the Gaussian noise technique, the Gaussian scaling technique, the random cropping technique, and the phase shifting technique.

[0053] The processor (120) can label each pair of propagated PRPD data sets with a corresponding partial discharge defect type.

[0054] Here, the partial discharge defect type can be any one of Corona, Floating, Particle, Void, and Noise.

[0055] The processor (120) may also divide a pair of propagated PRPD datasets into a pair of propagated training PRPD datasets and a pair of propagated test PRPD datasets, and perform pre-training on a defect detection model using the pair of propagated training PRPD datasets.

[0056] Here, pre-learning can be supervised contrastive learning.

[0057] More specifically, the processor (120) can convert a pair of propagated training PRPD data sets into a pair of feature vectors based on an encoder network.

[0058] Next, the processor (120) can apply a nonlinear transformation to a pair of feature vectors based on the projection head (213) to convert them into a pair of space vectors.

[0059] Next, the processor (120) can adjust the weights of the encoder network and the weights of the projection head to minimize a supervised contrastive loss function for a pair of space vectors.

[0060] Here, the fault detection model can be an encoder network that has undergone pretraining. Furthermore, the weights of the tuned encoder network can be applied to the fault detection model as fixed values.

[0061] The processor (120) can input a pair of propagated test PRPD data sets into a defect detection model on which pre-learning has been performed to determine the type of defect, and complete learning by adjusting the weights of the defect detection model based on the determination result.

[0062] The weights of the defect detection model can be adjusted to minimize the cross-entropy loss function.

[0063] That is, the processor (120) processes data pairs that have been multiplied by a series of PRPD data as described above to perform pre-learning (supervised contrastive learning), and updates the pre-learned defect detection model to complete the final defect detection model.

[0064] Here, the processor (120) can adjust the weights of the encoder network, the weights of the projection head (213), and the weights of the defect detection model based on the Adam optimization technique (Adaptive Moment Estimation) during the pre-learning and additional testing process.

[0065] Through this, the performance of the defect detection model can be improved by performing data augmentation for efficient and accurate learning of partial discharge data characteristics using partial discharge data, which is relatively difficult to obtain as fault data.

[0066] A more detailed explanation of this will be provided later.

[0067] FIG. 2 illustrates a partial discharge defect detection device (200) according to one embodiment, FIG. 3 illustrates components of an artificial intelligence modeler (210) in a partial discharge defect detection device (200) according to one embodiment, and FIG. 4 schematically illustrates a defect detection model (220) based on supervised contrastive learning according to one embodiment.

[0068] Referring to FIGS. 2 to 4, a defect detection device (200) can determine a partial discharge defect type by analyzing PRPD (Phase Resolved Partial Discharge) data received from an ultra-high frequency (UHF) sensor.

[0069] For this purpose, the defect detection device (200) may include an artificial intelligence modeler (210) and a defect detection model (220).

[0070] The artificial intelligence modeler (210) is ready to detect partial discharge defects in actual gas-insulated switchgear by creating and updating a defect detection model (220).

[0071] An artificial intelligence modeler (210) can create a defect detection model (220) by performing supervised contrastive learning based on a pre-collected PRPD dataset.

[0072] The artificial intelligence modeler (210) may include a data augmentation unit (211), an encoder network (212), a projection head (213), a pre-learning unit (214), and a model optimization unit (215), as illustrated in FIG. 3.

[0073] The data augmentation unit (211) can generate a pair of augmented PRPD data sets by applying different data augmentation techniques to each PRPD data set collected in advance.

[0074] Here, different data augmentation techniques can be applied to augment / augment the insufficient sample data as illustrated in Fig. 4, and two techniques can be arbitrarily selected from among the Gaussian noise technique, the Gaussian scaling technique, the random crop technique, and the phase shift technique.

[0075] The data augmentation unit (211) labels each pair of propagated PRPD data with a corresponding defect type for each pair of propagated PRPD data sets, and can divide each pair of propagated PRPD data sets into a pair of propagated training PRPD data sets and a pair of propagated test PRPD data sets.

[0076] The encoder network (212) can convert a pair of propagated training PRPD data sets into a pair of feature vectors.

[0077] The projection head (213) can apply a nonlinear transformation to a pair of feature vectors to transform them into a pair of space vectors.

[0078] The pre-learning unit (214) can adjust the weights of the encoder network (212) and the weights of the projection head (213) to minimize the supervised contrastive loss function for a pair of space vectors.

[0079] The model optimization unit (215) inputs a pair of proliferated test PRPD data sets into a defect detection model (220) to determine the defect type, and adjusts the weights of the defect detection model (220) based on the determination result to complete learning.

[0080] Here, the weights of the encoder network (212), the weights of the projection head (213), and the weights of the defect detection model (220) can be adjusted based on the Adam optimization technique (Adaptive Moment Estimation).

[0081] The defect detection model (220) can receive PRPD data and determine the partial discharge defect type.

[0082] Here, the partial discharge defect type can be any one of Corona, Floating, Particle, Void, and Noise.

[0083] Referring to FIG. 4, a defect detection model (220) generated by performing supervised contrastive learning based on a pre-collected PRPD (Phase Resolved Partial Discharge) dataset can be confirmed.

[0084] It should be noted that before performing supervised contrastive learning, a pair of augmented PRPD datasets is generated by applying a data augmentation technique to the pre-collected PRPD dataset. In other words, the pre-collected PRPD dataset is not directly used for supervised contrastive learning, but two data augmentation techniques are randomly selected from the four data augmentation techniques and applied to one of the pre-collected PRPD datasets. By applying to a pair of propagated PRPD data , can be created.

[0085] At this time, a pair of propagated PRPD data as shown in Fig. 4 , Labeling of the same defect type can be performed. This is , Despite these different characteristics, they are derived from the same PRPD data and are therefore given the same label for contrastive learning.

[0086] Here, a pair of propagated PPD datasets can be divided into training and testing sets. For example, 80% of the dataset can be used for training and 20% for testing.

[0087] Supervised contrastive learning based on the pre-collected PRPD dataset is performed as follows.

[0088] 1) Encoder network (212) is a pair of input propagated training PRPD data , It receives input and converts it into a single 1-dimensional feature vector. That is, the encoder network (212) A two-dimensional matrix Converts to a one-dimensional vector. This is can be expressed as . Here represents the output shape of the last layer of the network. class From, class A pair of feature vectors can be obtained. These feature vectors represents the features of the image. The encoder network (212) is mainly composed of convolution layers and flatten layers, and extracts high-dimensional features and then flattens them.

[0089] 2) Projection head (213) Is can be defined as, It consists of a single linear layer of units and a nonlinear activation ReLU function. In addition, is the index of an arbitrary augmented sample within the multiviewer batch.

[0090] The projection head (213) outputs the vector from the encoder network (212). It takes as input another space vector Convert to .

[0091] That is, the projection head (213) is a component that follows the encoder network (212), and has the role of receiving the feature vector generated by the encoder network (212) and converting it into another space.

[0092] Through this transformation, the feature vectors can be more useful in the contrastive learning process.

[0093] The projection head (213) is primarily composed of linear layers, which transform feature vectors into new dimensions. The linear layers perform this transformation using weights and biases.

[0094] Next, the projection head (213) uses a nonlinear activation function, for example, the ReLU (Rectified Linear Unit) function, which makes negative values ​​0 but leaves positive values ​​as they are.

[0095] 3) Space vector is input to the contrastive loss function and the loss value is calculated. The contrastive loss function is defined as follows.

[0096]

[0097] Here is as follows.

[0098]

[0099] Here Is from until In the element This is a set that has been removed, Is is the index set of positive samples with the same label as the th label. Here is a scalar temperature parameter, represents the total number of elements in the set.

[0100] in other words, is a data point represents a set of positive samples, where positive samples mean samples with the same class label, is a data point in the set of all possible samples. It means excluding .

[0101] represents the loss for individual data points, which is the overall loss Contribute to.

[0102] 4) Model training proceeds toward minimizing a contrastive loss function. That is, the model adjusts weights to maximize the similarity between positive samples and minimize the similarity with negative samples through this loss function. As a result, the encoder network (212) The weights are frozen by the model optimization unit (215) and used in the defect detection model (220).

[0103] The model optimization unit (215) is an encoder network (212). cast Fixed weights are used before classifying the defect detection model (220). Is And, It consists of a hidden layer, two dropout layers, and a softmax.

[0104] To prevent overfitting, two dropout layers were applied before and after the hidden layer.

[0105] Model optimization unit (215) defect detection model (220) The cross entropy loss function used to optimize is as follows.

[0106]

[0107] represents the index of the mini-batch, Index is 1 when it is an index to the ground truth, and 0 otherwise, that is, am.

[0108]

[0109] Figure 5 illustrates an encoder network (212) according to one embodiment.

[0110] The structure of the encoder network (212) consists of five convolutional blocks (each including a 3x3 convolutional layer, a rectified linear unit activation, and a max pooling layer). It consists of an in-platon layer.

[0111] After the encoder network (212), the projection head (213) It consists of a single hidden layer, is used to apply nonlinear transformations, Minimize the loss of supervised contrast by projecting.

[0112] According to one embodiment, a fault detection model (220) whose weights are adjusted by a model optimization unit (215) has a hidden layer of 900 nodes and five fault types to classify. It consists of a Softmax layer.

[0113]

[0114] FIG. 6 is a flowchart showing a partial discharge defect detection method according to one embodiment, and FIG. 7 is a flowchart showing in more detail a defect detection model (220) generation step (1100) in a partial discharge defect detection method according to one embodiment.

[0115] Referring to FIG. 6, a method for detecting a partial discharge defect in a gas-insulated switchgear may include a step (1100) of generating a defect detection model (220) by performing supervised contrastive learning based on a pre-collected PRPD data set and a step (1200) of inputting PRPD data and determining a partial discharge defect type.

[0116] Referring to FIG. 7, the step (1100) of generating a defect detection model (220) includes a step (1110) of generating a pair of propagated PRPD data sets by applying different data propagation techniques to PRPD data for a pre-collected PRPD data set, a step (1120) of labeling each pair of propagated PRPD data with a corresponding defect type for each pair of propagated PRPD data sets, and dividing each pair of propagated PRPD data sets into a pair of propagated training PRPD data sets and a pair of propagated test PRPD data sets, a step (1130) of converting a pair of propagated training PRPD data for a pair of propagated training PRPD data sets into a pair of feature vectors based on an encoder network (212), a step (1140) of applying a nonlinear transformation to a pair of feature vectors based on a projection head (213) to convert them into a pair of space vectors, and a step (1150) of minimizing a supervised contrastive loss function for a pair of space vectors using an encoder network. It may include a step (1150) of adjusting the weights of the network (212) and the weights of the projection head (213), and a step (1160) of inputting a pair of propagated test PRPD data sets into a defect detection model (220) to determine the type of defect, and adjusting the weights of the defect detection model (220) based on the determination result to complete learning.

[0117]

[0118] Fig. 8 is a diagram comparing the detection performance of a defect detection model (220) based on map contrast learning according to one embodiment with another detection model.

[0119] The defect detection model (220) based on supervised contrastive learning according to an embodiment of the present invention is superior to SVM due to the difference in particle failure performance (100% vs. 53.85% in the test).

[0120] In addition, the defect detection model (220) shows 4.28% higher performance than MLP in terms of overall classification performance, and shows high performance in particle, cavity, and noise classification in particular.

[0121] Although the corona class accuracy of the defect detection model (220) is lower than that of CNN, the overall performance (97.28% vs. 95.24% for CNN) is still superior, mainly due to differences in particle and noise classification.

[0122] Therefore, the diagram in Fig. 8 shows that the defect detection model (220) is the best among the four methods in terms of accuracy.

[0123]

[0124] FIG. 9 is a diagram comparing the precision, recall, and F-1 score of a defect detection model (220) based on map contrast learning according to one embodiment with other detection models.

[0125] TP stands for the number of samples correctly predicted as “positive”, FP stands for the number of samples incorrectly predicted as “positive”, and FN stands for the number of samples incorrectly predicted as “negative”.

[0126] Precision represents the quality of positive predictions made by the model, while recall represents the proportion of positive instances that are correctly classified.

[0127] The F1-score represents the balance between precision and recall.

[0128] The F1-score ranges between 0 and 1, with values ​​closer to 1 indicating a better model, and conversely, values ​​closer to 0 indicating a worse model.

[0129] Referring to Fig. 9, the defect detection model (220) shows high performance in terms of precision and recall, and in particular, shows values ​​of 0.968 and 1 in the noise class.

[0130] As a result, the F1-score also reaches 0.984, outperforming other methods.

[0131] The defect detection model (220) exhibits a precision value of 1 for floating faults and a recall value of 1 for particle faults.

[0132] The defect detection model (220) shows the highest F1-score in three of the five classes.

[0133] As shown in Fig. 9, it can be seen that the defect detection model (220) is superior to SVM, MLP, and CNN in terms of F1-score performance.

[0134]

[0135] The above description is merely an illustrative illustration of the technical idea of ​​the present invention, and those skilled in the art will appreciate that various modifications and variations can be made without departing from the essential characteristics of the present invention. Therefore, the embodiments disclosed in the present invention are intended to illustrate, rather than limit, the technical idea of ​​the present invention, and the scope of the technical idea of ​​the present invention is not limited by these embodiments. The scope of protection of the present invention should be interpreted by the following claims, and all technical ideas within a scope equivalent thereto should be interpreted as being included in the scope of the rights of the present invention.

[0136]

[0137] The present invention can be used in a technology for detecting defects such as partial discharge in gas-insulated switchgear, etc.

Claims

1. A memory for storing a PRPD (Phase Resolved Partial Discharge) data set of a gas-insulated switchgear received from an ultra-high frequency (UHF) sensor; and A processor for generating a defect detection model using the stored PRPD dataset, The above processor, For the above PRPD dataset, a pair of augmented PRPD datasets is created by applying different data augmentation techniques to each PRPD data, For each pair of propagated PRPD data sets above, label each pair of propagated PRPD data with the corresponding partial discharge defect type, Divide the above pair of propagated PRPD datasets into a pair of propagated training PRPD datasets and a pair of propagated test PRPD datasets, respectively. Pre-training for the defect detection model is performed using the above pair of propagated learning PRPD datasets, A system for generating a partial discharge fault detection model for a gas-insulated switchgear, which inputs the above pair of proliferated test PRPD data sets into the above fault detection model to determine a fault type, and completes learning by adjusting the weights of the fault detection model based on the determination result.

2. In paragraph 1, The above pre-learning is a model generation system for detecting partial discharge defects in gas-insulated switchgear using supervised contrastive learning.

3. In paragraph 2, The above processor, Based on the encoder network, a pair of propagated training PRPD data sets is converted into a pair of feature vectors, Based on the projection head, a nonlinear transformation is applied to the above pair of feature vectors to transform them into a pair of space vectors, A system for generating a partial discharge fault detection model of a gas-insulated switchgear, which adjusts the weights of the encoder network and the weights of the projection head to minimize a supervised contrast loss function for the above pair of space vectors.

4. In paragraph 3, A system for generating a partial discharge fault detection model for a gas-insulated switchgear, wherein the above fault detection model is the encoder network, and the weights of the adjusted encoder network are applied to the fault detection model as fixed values.

5. In paragraph 3, A system for generating a partial discharge fault detection model of a gas-insulated switchgear, wherein the weights of the above fault detection model are adjusted to minimize a cross entropy loss function.

6. In paragraph 1, A system for generating a partial discharge fault detection model for a gas-insulated switchgear, wherein the weights of the encoder network, the weights of the projection head, and the weights of the fault detection model are adjusted based on the Adam optimization technique (Adaptive Moment Estimation).

7. In paragraph 1, The above different data augmentation techniques are a partial discharge defect detection model generation system for gas-insulated switchgear, wherein two techniques are randomly selected from among Gaussian noise technique, Gaussian scaling technique, random crop technique and phase shift technique.

8. In paragraph 1, A system for generating a partial discharge defect detection model for a gas-insulated switchgear, wherein the above partial discharge defect type is one of Corona, Floating, Particle, Void, and Noise.

9. A method for detecting a partial discharge defect in a gas-insulated switchgear, which determines the type of partial discharge defect by analyzing PRPD (Phase Resolved Partial Discharge) data received from an ultra-high frequency (UHF) sensor, A step of generating a defect detection model by performing supervised contrastive learning based on a pre-collected PRPD dataset; and A method for detecting a partial discharge defect in a gas-insulated switchgear, comprising the step of inputting the above PRPD data and determining a partial discharge defect type.

10. In paragraph 9, The steps of generating the above defect detection model are: A step of generating a pair of augmented PRPD datasets by applying different data augmentation techniques to each PRPD data set collected in advance; A step of labeling each pair of propagated PRPD data with a corresponding defect type for each pair of propagated PRPD data sets, and dividing each pair of propagated PRPD data sets into a pair of propagated training PRPD data sets and a pair of propagated test PRPD data sets; A step of converting a pair of propagated training PRPD data into a pair of feature vectors for the pair of propagated training PRPD data sets based on the encoder network; A step of applying a nonlinear transformation to the above pair of feature vectors based on the projection head to convert them into a pair of space vectors; A step of adjusting the weights of the encoder network and the weights of the projection head to minimize a supervised contrastive loss function for the pair of space vectors; and A method for detecting a partial discharge fault in a gas-insulated switchgear, comprising the steps of inputting the above pair of proliferated test PRPD data sets into the above fault detection model to determine a fault type, and completing learning by adjusting the weights of the above fault detection model based on the determination result.

11. In paragraph 10, A method for detecting a partial discharge fault in a gas-insulated switchgear, wherein the above fault detection model is the encoder network, and the weights of the adjusted encoder network are applied to the fault detection model as fixed values.

12. In paragraph 10, A method for detecting a partial discharge fault in a gas-insulated switchgear, wherein the weights of the above fault detection model are adjusted to minimize a cross entropy loss function.

13. In paragraph 10, A method for detecting a partial discharge fault in a gas-insulated switchgear, wherein the weights of the encoder network, the weights of the projection head, and the weights of the fault detection model are adjusted based on the Adam optimization technique (Adaptive Moment Estimation).

14. In paragraph 10, The above different data augmentation techniques are a method for detecting partial discharge defects in gas-insulated switchgear, wherein two techniques are randomly selected from among Gaussian noise technique, Gaussian scaling technique, random crop technique and phase shift technique.

15. In paragraph 9, A method for detecting a partial discharge defect in a gas-insulated switchgear, wherein the above partial discharge defect type is one of Corona, Floating, Particle, Void and Noise.

16. In a partial discharge fault detection device of a gas-insulated switchgear, which analyzes PRPD (Phase Resolved Partial Discharge) data received from an ultra-high frequency (UHF) sensor to determine the partial discharge fault type, An artificial intelligence modeler that generates a defect detection model by performing supervised contrastive learning based on a pre-collected PRPD dataset; and A partial discharge fault detection device for a gas-insulated switchgear, comprising a fault detection model that receives the above PRPD data and determines a partial discharge fault type.

17. In paragraph 16, The above artificial intelligence modeler, A data augmentation unit that applies different data augmentation techniques to each PRPD data for the previously collected PRPD data sets to generate a pair of augmented PRPD data sets, labels each pair of augmented PRPD data with a corresponding defect type for each of the pair of augmented PRPD data sets, and divides each of the pair of augmented PRPD data sets into a pair of augmented training PRPD data sets and a pair of augmented testing PRPD data sets; An encoder network that converts a pair of propagated training PRPD data sets into a pair of feature vectors for the above pair of propagated training PRPD data sets; A projection head that applies a nonlinear transformation to the above pair of feature vectors to transform them into a pair of space vectors; A pre-learning unit that adjusts the weights of the encoder network and the weights of the projection head to minimize a supervised contrastive loss function for the pair of space vectors; and A model optimization unit is included that inputs the above pair of proliferated test PRPD data sets into the above defect detection model to determine the defect type, and completes learning by adjusting the weights of the defect detection model based on the determination result. A partial discharge fault detection device of a gas-insulated switchgear, wherein the above fault detection model is the encoder network, and the weights of the adjusted encoder network are applied to the fault detection model as fixed values.

18. In paragraph 17, A device for detecting a partial discharge defect in a gas-insulated switchgear, wherein the above partial discharge defect type is one of Corona, Floating, Particle, Void and Noise.

19. In Article 17, The above different data augmentation techniques are two techniques randomly selected from among Gaussian noise technique, Gaussian scaling technique, random crop technique and phase shift technique. A device for detecting partial discharge defects in a gas-insulated switchgear.

20. In paragraph 17, A partial discharge fault detection device for a gas-insulated switchgear, wherein the weights of the encoder network, the weights of the projection head, and the weights of the fault detection model are adjusted based on the Adam optimization technique (Adaptive Moment Estimation).

Images