A method and system for identifying faults in high-frequency transformers

Abstract

This invention discloses a method and system for high-frequency transformer fault identification, belonging to the field of high-frequency transformer fault detection technology. The method comprises: acquiring a characteristic gas vector sequence of the high-frequency transformer and converting the characteristic gas vector sequence into an actual fault recurrence graph; generating an expanded dataset based on the actual fault recurrence graph, a pre-acquired random noise vector, and a preset initial depth convolutional generative adversarial network model; training a preset initial MobileNetV3 model based on the expanded dataset and the characteristic gas vector sequence and constructing a cross-entropy loss function, then updating the model with the goal of minimizing the cross-entropy loss function until a preset convergence condition is met, thus obtaining the MobileNetV3 model; inputting the fault recurrence graph to be tested into the MobileNetV3 model to obtain the fault detection result of the high-frequency transformer to be detected. Therefore, by implementing this invention, the low accuracy of high-frequency transformer fault identification and the insufficient ability to identify complex faults can be avoided, achieving high-precision and high-reliability identification of various types of high-frequency transformer faults.

Description

Technical Field

[0001] This invention relates to the field of high-frequency transformer fault detection technology, and in particular to a high-frequency transformer fault identification method and system. Background Technology

[0002] As the core power conversion component of power electronic transformers, the high-frequency transformer (HF transformer) directly affects the safety and stability of the entire power electronic equipment and even the power grid system. Dissolved gas analysis (DGA) in oil is currently one of the most effective and commonly used techniques for monitoring the internal insulation condition of oil-immersed transformers and diagnosing potential faults. When faults such as partial discharge or overheating occur inside the transformer, the insulating oil and solid insulating materials decompose to produce gases including hydrogen (… ), methane ( ), ethane ( ), ethylene ( ), acetylene ( Characteristic gases such as ( ), the content, proportion and growth trend of these gases are closely related to the type and severity of the fault.

[0003] Currently, transformer fault identification methods based on DGA are mainly divided into three categories: The first category is the traditional ratio coding method, such as the three-ratio method and the Doernenburg ratio method recommended by the International Electrotechnical Commission (IEC). This type of method diagnoses faults by calculating the ratio of specific gas pairs and referring to a pre-set coding interval and rule table. The principle is simple and easy to implement. The second category is based on traditional machine learning methods, such as Support Vector Machine (SVM), Random Forest, and Decision Tree. These methods extract DGA data features (such as absolute gas content, relative gas content, ratio, etc.) manually as input and use a classifier for fault identification, which improves performance compared to the ratio method. The third category is based on deep learning methods, such as Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM). These methods can automatically learn deep features from raw or pre-processed data, and theoretically have high identification accuracy.

[0004] Under the current technological background, although each transformer fault identification method based on DGA has its own basic advantages, all methods have obvious drawbacks, as follows: Traditional ratio coding methods have inherent defects such as "coding defects" (i.e., the ratio cannot be diagnosed if it falls outside the preset range), "fuzzy threshold criteria", and weak ability to identify complex faults, resulting in limited diagnostic accuracy; Traditional machine learning-based methods rely heavily on the quality of manual feature engineering, and the model performance will rapidly decline and the generalization ability will be insufficient when the number of fault types increases and the data distribution becomes more complex; Deep learning-based methods rely on large-scale, high-quality labeled data, but the high reliability of power equipment and the extremely low probability of fault occurrence result in a scarcity of transformer fault samples that can be used for training, which can easily lead to overfitting of complex deep learning models and make them difficult to apply in practice; At the same time, deep learning-based methods also generally suffer from insufficient utilization of one-dimensional sequence information, making it difficult to efficiently and losslessly input the temporal dynamic information and numerical magnitude relationships contained in the one-dimensional sequence of DGA into the deep model. Existing data augmentation methods are ineffective. Traditional data augmentation methods generate samples of poor quality and insufficient diversity, failing to effectively simulate the distribution of real fault data. Furthermore, image-based methods have low feature discrimination. Existing one-dimensional sequence to two-dimensional image conversion methods (such as relative position transformation and Markov transfer fields) generate images with high visual feature similarity, and the differences between images of different fault categories are not obvious, making it difficult for the model to learn highly discriminative features. In addition, some high-performance deep learning models (such as VGG16) have a huge number of parameters, high computational resource consumption, and long training and inference times, making them difficult to deploy in field equipment with limited computing resources or systems that require rapid response. Summary of the Invention

[0005] The present invention aims to provide a method and system for identifying high-frequency transformer faults, so as to solve the above-mentioned technical problems, avoid low accuracy in identifying high-frequency transformer faults and insufficient ability to identify complex faults, and achieve high accuracy and high reliability in identifying various types of faults in high-frequency transformers.

[0006] To address the aforementioned technical problems, this invention provides a high-frequency transformer fault identification method, comprising: Obtain several characteristic gas vector sequences of several high-frequency transformers, and convert the several characteristic gas vector sequences into several actual fault recursion graphs; Based on several real-world fault recursion graphs, several pre-acquired random noise vectors, and a pre-set initial depth convolutional generative adversarial network model, an expanded dataset is generated. The preset initial MobileNetV3 model is trained based on the expanded dataset and the feature gas vector sequence to obtain several fault prediction probability vectors. A cross-entropy loss function is constructed based on the expanded dataset and the fault prediction probability vectors. The preset initial MobileNetV3 model is then updated with the goal of minimizing the cross-entropy loss function until the preset convergence condition is met, and the MobileNetV3 model is obtained. Input the fault recursion graph of the high-frequency transformer to be tested into the MobileNetV3 model to obtain the fault detection results of the high-frequency transformer to be tested.

[0007] In the above scheme, by acquiring the feature gas vector sequence and converting it into an actual fault recurrence graph, the fault features can be visualized in two dimensions, enhancing the identification of fault features and providing input data for subsequent model training. Simultaneously, from the feature representation level, this lays the foundation for avoiding low accuracy in high-frequency transformer fault identification and insufficient ability to identify complex faults. Next, by combining the actual fault recurrence graph, pre-acquired random noise vectors, and a pre-set initial depth convolutional generative adversarial network model to generate an expanded dataset, the fault training and verification samples can be expanded, effectively supplementing the samples of complex faults and addressing the problems of low fault identification accuracy and insufficient ability to identify complex faults from a data perspective. Then, by training the initial MobileNetV3 model using the expanded dataset and feature gas vector sequence, generating fault prediction probability vectors, and constructing a cross-entropy loss function to optimize the model until convergence, a MobileNetV3 model with high accuracy and complex fault identification capabilities can be obtained. Thus, the fault recurrence graph of the high-frequency transformer to be tested can be input into this MobileNetV3 model to obtain the fault detection results of the high-frequency transformer to be tested, achieving high-precision and high-reliability identification of various faults in the high-frequency transformer to be tested.

[0008] Further, the step of acquiring several characteristic gas vector sequences of several high-frequency transformers and converting the several characteristic gas vector sequences into several actual fault recursion graphs includes: Obtain historical dissolved gas analysis data of several high-frequency transformers, and extract several characteristic gas content values ​​from the historical dissolved gas analysis data to obtain several characteristic gas vector sequences; Pair any two characteristic gas vectors in the characteristic gas vector sequence to obtain several vector pairs, and calculate the Euclidean distance between the vector pairs, so as to convert several characteristic gas vector sequences into several actual fault recursion graphs based on the Euclidean distance between the vector pairs.

[0009] In the above scheme, by acquiring historical dissolved gas analysis data from the oil of high-frequency transformers and extracting characteristic gas content values ​​to obtain a characteristic gas vector sequence, fault characteristics of high-frequency transformers can be extracted, providing data support for subsequent visualization conversion of fault characteristics. Next, by pairing any two characteristic gas vectors in the characteristic gas vector sequence and calculating the Euclidean distance, the one-dimensional characteristic gas vector sequence can be converted into an actual fault recurrence graph, achieving two-dimensional visualization of fault characteristics. Simultaneously, it can fully preserve the temporal dynamics and recurrence characteristics of the fault, thereby enhancing the identification of different fault characteristics and providing input for subsequent model training.

[0010] Furthermore, the step of acquiring historical dissolved gas analysis data in the oil of several high-frequency transformers, and extracting several characteristic gas content values ​​from the historical dissolved gas analysis data to obtain several characteristic gas vector sequences includes: Obtain historical dissolved gas analysis data in the oil of several high-frequency transformers; Data cleaning was performed on the historical dissolved gas analysis data in the oil to obtain dissolved gas analysis data in the pure oil. The dissolved gas analysis data in the pure oil are normalized to obtain the dissolved gas analysis data in the standard oil. Several characteristic gas content values ​​were extracted from the dissolved gas analysis data of standard oil, and several characteristic gas vector sequences were obtained.

[0011] In the above scheme, acquiring historical dissolved gas analysis data from several high-frequency transformer oils provides foundational data for subsequent data processing and feature extraction. Next, data cleaning is performed on the historical dissolved gas analysis data to obtain purified dissolved gas analysis data, which removes noise and outliers, ensuring data validity and accuracy. Then, normalization is performed on the purified dissolved gas analysis data to obtain standard dissolved gas analysis data, eliminating the dimensions of the purified data and standardizing its scale to avoid interference from data scale differences in subsequent feature processing. Finally, by extracting several characteristic gas content values ​​from the standard dissolved gas analysis data to obtain a characteristic gas vector sequence, fault features of the high-frequency transformer can be extracted, forming a one-dimensional fault feature sequence, providing suitable feature input for subsequent recursive graph conversion.

[0012] Furthermore, the process of generating an expanded dataset based on several actual fault recurrence graphs, several pre-acquired random noise vectors, and a pre-set initial depth convolutional generative adversarial network model includes: The preset initial depth convolutional generative adversarial network model includes an initial generator; Based on the initial generator, several pre-acquired random noise vectors are mapped to obtain several initial fake feature maps; The initial forged feature map is upsampled based on the initial generator to obtain an upgraded forged feature map. The upgraded forgery feature map is cropped based on the initial generator and the actual fault recursion map to obtain the forged fault recursion map; The pre-set initial depthwise convolutional generative adversarial network (GAN) model is trained based on the actual fault recurrence graph and the fake fault recurrence graph to obtain the standard depthwise convolutional GAN ​​model, and the augmented dataset is obtained based on the actual fault recurrence graph and the standard depthwise convolutional GAN ​​model.

[0013] In the above scheme, an initial fake feature map is obtained by mapping a pre-acquired random noise vector to an initial generator. This transforms disordered noise into feature representations, providing a basic feature carrier for generating fake fault recurrence maps. Next, the initial fake feature map is upsampled by the initial generator to obtain an upgraded fake feature map. This increases the dimension and size of the initial fake feature map, making its feature scale approximate the actual fault recurrence map, laying the foundation for subsequent steps. Then, the upgraded fake feature map is cropped using the initial generator and the actual fault recurrence map to obtain the fake fault recurrence map. This ensures that the size of the fake fault feature map matches the actual fault recurrence map, guaranteeing a consistent format between the fake and real samples. Finally, by combining the actual fault recurrence map, the fake fault recurrence map, and a pre-defined deep convolutional generative adversarial network model, an expanded dataset is obtained. This generates simulation samples highly consistent with the actual fault recurrence map, completing a high-quality expansion of the fault training samples.

[0014] Furthermore, the preset initial depthwise convolutional generative adversarial network model also includes an initial discriminator; the step of training the preset initial depthwise convolutional generative adversarial network model based on the actual fault recurrence graph and the fake fault recurrence graph to obtain a standard depthwise convolutional generative adversarial network model, and obtaining an expanded dataset based on the actual fault recurrence graph and the standard depthwise convolutional generative adversarial network model, includes: The preset initial depth convolutional generative adversarial network model also includes an initial discriminator; The initial discriminator and the initial generator are trained based on the actual fault recursion graph and the fake fault recursion graph respectively. The probability values ​​of the real recursion graph and the trained generator are obtained. If the probability values ​​of the real recursion graph satisfy the preset probability conditions and the trained generator satisfies Nash equilibrium, the training is completed and the fault generator and the fault discriminator are obtained to form a standard deep convolutional generative adversarial network model. An augmented dataset is obtained based on a real fault recursion graph and a standard deep convolutional generative adversarial network model.

[0015] In the above scheme, the initial discriminator and initial generator are trained using actual and simulated fault recurrence graphs. Training of the preset initial deep convolutional generative adversarial network (DBGAN) model is completed when the probability values ​​of the real recurrence graph meet a preset probability condition and the trained generator satisfies Nash equilibrium. At this point, the resulting standard DBGAN model reaches Nash equilibrium and can generate fault samples with high similarity to the actual fault recurrence graph. Then, using the actual fault recurrence graph and the trained standard DBGAN model, simulated fault recurrence graphs that are highly consistent with the actual fault recurrence graphs can be generated in batches, achieving high-quality augmentation of the dataset.

[0016] Further, the step of training a preset initial MobileNetV3 model based on the expanded dataset and the feature gas vector sequence to obtain several fault prediction probability vectors, constructing a cross-entropy loss function based on the expanded dataset and the fault prediction probability vectors, and then updating the preset initial MobileNetV3 model with the goal of minimizing the cross-entropy loss function until a preset convergence condition is met to obtain the MobileNetV3 model includes: The expanded dataset is divided according to a preset ratio to obtain a fault classification training set and a fault classification validation set; The feature gas vector sequence is matched to obtain several real fault classification labels; Based on the fault classification training set and the preset initial MobileNetV3 model, several fault prediction probability vectors are obtained. A cross-entropy loss function is constructed based on the real fault classification labels and fault prediction probability vectors. The preset initial MobileNetV3 model is updated with the goal of minimizing the cross-entropy loss function to obtain the MobileNetV3 training model. The MobileNetV3 trained model is validated based on the fault classification validation set to obtain the model accuracy. If the model accuracy meets the preset convergence condition, the training is completed and the MobileNetV3 model is obtained.

[0017] In the above scheme, the expanded dataset is divided into a fault classification training set and a fault classification validation set according to a preset ratio, providing data support for the subsequent training and convergence determination of the preset initial MobileNetV3 model. Next, several real fault classification labels are obtained by matching feature gas vector sequences, providing a real reference for the construction of the loss function and the optimization of model parameters. Then, several fault prediction probability vectors are obtained using the fault classification training set and the preset initial MobileNetV3 model, providing prediction data for loss calculation. Subsequently, a cross-entropy loss function is constructed using the real fault classification labels and fault prediction probability vectors, and the preset initial MobileNetV3 model is updated with its minimization as the objective. Backpropagation is used to iteratively optimize the model parameters, improving the accuracy of fault identification. Finally, the MobileNetV3 trained model is validated using the fault classification validation set, and it is determined whether the model accuracy meets the preset convergence condition. This verifies the model's generalization ability and determines whether to stop training, ultimately obtaining a MobileNetV3 model that meets the accuracy requirements.

[0018] Furthermore, based on the fault classification training set and the preset initial MobileNetV3 model, several fault prediction probability vectors are obtained, including: The preset initial MobileNetV3 model includes a dimensionality increase module, a feature extraction module, an attention enhancement module, and a dimensionality reduction module; Based on the dimension-upgrading module, the fault classification training set is mapped to a high dimension to obtain several initial high-dimensional feature maps; Based on the feature extraction module, spatial features are extracted from the initial high-dimensional feature map to obtain the basic fault feature map; An enhanced fault feature map is obtained by adaptively optimizing the basic fault feature map based on the attention enhancement module. The enhanced fault feature map is processed by channel dimensionality reduction module to obtain the dimensionality-reduced fault feature map. The reduced-dimensional fault feature map and the fault classification training set are residually concatenated to obtain a global fusion feature map. The global fusion feature map is sequentially subjected to global average pooling, fully connected layer feature mapping, and Softmax activation function calculation to obtain several fault prediction probability vectors.

[0019] In the above scheme, the initial high-dimensional feature map is obtained by mapping the fault classification training set to a high dimension through the dimensionality-up module. This enables channel-level dimensionality upscaling and high-dimensional spatial representation of fault features, laying the foundation for the extraction of deep fault features. Next, the basic fault feature map is obtained by spatial feature extraction from the initial high-dimensional feature map through the feature extraction module. This accurately captures the core spatial texture features of the initial high-dimensional feature map, achieving a concrete extraction of fault features. Then, the enhanced fault feature map is adaptively optimized through the attention enhancement module to obtain an enhanced fault feature map. This strengthens key fault features and suppresses redundant features, improving the discriminative power of fault features. Subsequently, the enhanced fault feature map is channel-level dimensionality-reduced through the dimensionality-reduction module to obtain a dimensionality-reduced fault feature map. This simplifies the feature dimension while retaining the core fault features, reducing the computational complexity of the model. Finally, a global fusion feature map is obtained by residually connecting the dimensionality-reduced fault feature map and the fault classification training set. This fuses shallow and deep fault features, improving the overall feature representation capability. Finally, by sequentially performing global average pooling, fully connected layer feature mapping, and Softmax activation function calculation on the global fused feature map, the fused fault features can be transformed into a fault category probability distribution, resulting in a fault prediction probability vector corresponding to the fault classification training set.

[0020] Furthermore, the step of inputting the recursive graph of the high-frequency transformer under test into the MobileNetV3 model to obtain the fault detection results of the high-frequency transformer under test includes: The dissolved gas analysis data in the oil of the high-frequency transformer under test is converted to obtain the fault recursion graph and input into the MobileNetV3 model to obtain the fault detection results of the high-frequency transformer under test.

[0021] In the above scheme, by converting the dissolved gas analysis data of the oil in the high-frequency transformer under test into a fault recursion graph, the fault characteristics of the high-frequency transformer under test can be visualized in two dimensions. Simultaneously, this conversion provides suitable feature input for MobileNetV3 model inference. Then, by inputting the fault recursion graph into the MobileNetV3 model, the fault type of the high-frequency transformer under test can be inferred and determined, resulting in fault detection results. This achieves high-precision and high-reliability identification of various faults in the high-frequency transformer under test.

[0022] This invention provides a high-frequency transformer fault identification system, comprising a fault recurrence graph acquisition unit, a fault sample expansion unit, a MobileNetV3 model training and optimization unit, and a fault detection unit, specifically: The fault recursion graph acquisition unit is used to acquire several characteristic gas vector sequences of several high-frequency transformers and convert the several characteristic gas vector sequences into several actual fault recursion graphs. The fault sample expansion unit is used to generate an expanded dataset based on several actual fault recursion graphs, several pre-acquired random noise vectors, and a preset initial depth convolutional generative adversarial network model. The MobileNetV3 model training and optimization unit is used to train a preset initial MobileNetV3 model based on the expanded dataset and the feature gas vector sequence, obtain several fault prediction probability vectors, construct a cross-entropy loss function based on the expanded dataset and the fault prediction probability vectors, and then update the preset initial MobileNetV3 model with the goal of minimizing the cross-entropy loss function until the preset convergence condition is met, and obtain the MobileNetV3 model. The fault detection unit is used to input the fault recursion graph of the high-frequency transformer under test into the MobileNetV3 model to obtain the fault detection result of the high-frequency transformer under test.

[0023] This invention provides a high-frequency transformer fault identification system. In practical applications, it only requires a fault recurrence graph acquisition unit. By acquiring feature gas vector sequences and converting them into an actual fault recurrence graph, it can visualize fault features in two dimensions, enhancing the identification accuracy of fault features and providing input data for subsequent model training. Simultaneously, from the feature representation level, it lays the foundation for avoiding low accuracy and insufficient ability to identify complex faults in high-frequency transformers. Next, a fault sample expansion unit is employed. By combining the actual fault recurrence graph, pre-acquired random noise vectors, and a pre-set initial depth convolutional generative adversarial network model, it generates an expanded dataset. This expands the fault training and verification samples, effectively supplementing samples of complex faults and addressing the problems of low fault identification accuracy and insufficient ability to identify complex faults from a data perspective. Then, the MobileNetV3 model is used to train the optimization unit. The initial MobileNetV3 model is trained by expanding the dataset and the feature gas vector sequence. The fault prediction probability vector is generated and the cross-entropy loss function is constructed to optimize the model until convergence. This yields a MobileNetV3 model with high accuracy and complex fault identification capabilities. The fault detection unit can then input the fault recurrence graph of the high-frequency transformer under test into the MobileNetV3 model to obtain the fault detection results of the high-frequency transformer under test. This achieves high-precision and high-reliability identification of various faults in the high-frequency transformer under test.

[0024] Furthermore, the fault recurrence graph acquisition unit is used to acquire several characteristic gas vector sequences of several high-frequency transformers, and convert the several characteristic gas vector sequences into several actual fault recurrence graphs, including: Obtain historical dissolved gas analysis data of several high-frequency transformers, and extract several characteristic gas content values ​​from the historical dissolved gas analysis data to obtain several characteristic gas vector sequences; Pair any two characteristic gas vectors in the characteristic gas vector sequence to obtain several vector pairs, and calculate the Euclidean distance between the vector pairs, so as to convert several characteristic gas vector sequences into several actual fault recursion graphs based on the Euclidean distance between the vector pairs.

[0025] In the above scheme, by acquiring historical dissolved gas analysis data from the oil of high-frequency transformers and extracting characteristic gas content values ​​to obtain a characteristic gas vector sequence, fault characteristics of high-frequency transformers can be extracted, providing data support for subsequent visualization conversion of fault characteristics. Next, by pairing any two characteristic gas vectors in the characteristic gas vector sequence and calculating the Euclidean distance, the one-dimensional characteristic gas vector sequence can be converted into an actual fault recurrence graph, achieving two-dimensional visualization of fault characteristics. Simultaneously, it can fully preserve the temporal dynamics and recurrence characteristics of the fault, thereby enhancing the identification of different fault characteristics and providing input for subsequent model training. Attached Figure Description

[0026] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0027] Figure 1 A flowchart of a high-frequency transformer fault identification method provided in an embodiment of the present invention; Figure 2 This is an architecture diagram of a MobileNetV3 model provided in an embodiment of the present invention; Figure 3 This is an architecture diagram of a high-frequency transformer fault identification system provided in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0030] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0031] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0032] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0033] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0034] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0035] See Figure 1To avoid low accuracy in log correlation analysis and poor efficiency in attack event assessment due to data heterogeneity in multi-source security logs of power information networks, and to achieve efficient and accurate correlation analysis of multi-source security logs, this embodiment provides a high-frequency transformer fault identification method. The flowchart of this method can be found in [link to flowchart]. Figure 1 ,include: Step S1: Obtain several characteristic gas vector sequences of several high-frequency transformers, and convert the several characteristic gas vector sequences into several actual fault recursion diagrams; Step S2: Generate an expanded dataset based on several actual fault recursion graphs, several pre-acquired random noise vectors, and a pre-set initial depth convolutional generative adversarial network model; Step S3: Train the preset initial MobileNetV3 model based on the expanded dataset and the feature gas vector sequence to obtain several fault prediction probability vectors. Construct a cross-entropy loss function based on the expanded dataset and the fault prediction probability vectors. Update the preset initial MobileNetV3 model with the goal of minimizing the cross-entropy loss function until the preset convergence condition is met, and obtain the MobileNetV3 model. Step S4: Input the fault recursion graph of the high-frequency transformer to be tested into the MobileNetV3 model to obtain the fault detection results of the high-frequency transformer to be tested.

[0036] In this embodiment, the characteristic gas vector sequence is obtained. The recurrence plot (RP) algorithm is used to convert the fault into a real fault recurrence plot, which enables two-dimensional visualization of fault features, enhances the identification of fault features, and provides input data for subsequent model training. This also lays the foundation for avoiding low accuracy in high-frequency transformer fault identification and insufficient ability to identify complex faults from the feature representation level. Next, by combining the real fault recurrence plot, pre-acquired random noise vectors, and a pre-set initial depth convolutional generative adversarial network model to generate an expanded dataset, the fault training and verification samples are expanded, effectively supplementing the samples of complex faults and addressing the problems of low fault identification accuracy and insufficient ability to identify complex faults from the data level. Then, the initial MobileNetV3 model is trained using the expanded dataset and feature gas vector sequences. Fault prediction probability vectors are generated, and a cross-entropy loss function is constructed to optimize the model until convergence. This yields a MobileNetV3 model with high accuracy and complex fault identification capabilities. It maintains a high recognition rate when facing various typical faults (such as low-energy discharge, high-energy discharge, partial discharge, and overheating at different temperatures), and has good robustness and generalization ability, making it applicable to real data collected from the field. Finally, by inputting the fault recursion graph of the high-frequency transformer under test into the MobileNetV3 model, the fault detection results of the high-frequency transformer under test can be obtained, achieving high-precision and high-reliability identification of various faults in the high-frequency transformer under test. The average fault identification accuracy of this embodiment reaches 93.46%, and the accuracy of each fault type can reach over 90.63%, demonstrating the powerful generalization ability of the MobileNetV3 model in this embodiment to real complex data and possessing extremely high engineering application value.

[0037] Further, the step of acquiring several characteristic gas vector sequences of several high-frequency transformers and converting the several characteristic gas vector sequences into several actual fault recursion graphs includes: Obtain historical dissolved gas analysis data of several high-frequency transformers, and extract several characteristic gas content values ​​from the historical dissolved gas analysis data to obtain several characteristic gas vector sequences; Pair any two characteristic gas vectors in the characteristic gas vector sequence to obtain several vector pairs, and calculate the Euclidean distance between the vector pairs, so as to convert several characteristic gas vector sequences into several actual fault recursion graphs based on the Euclidean distance between the vector pairs.

[0038] In this embodiment, by acquiring historical dissolved gas analysis data from the oil of a high-frequency transformer and extracting characteristic gas content values ​​to obtain a characteristic gas vector sequence, fault characteristics of the high-frequency transformer can be extracted, providing data support for subsequent visualization and conversion of fault characteristics. Next, the characteristic gas vector sequence is mapped onto a two-dimensional image. By pairing any two characteristic gas vectors in the sequence and calculating their Euclidean distance, points with an Euclidean distance less than a preset distance threshold are considered... In the image, pixels marked as black (or dark) indicate that the two states are "recursive"; otherwise, they are marked as white (or light). This allows the one-dimensional feature gas vector sequence to be converted into an actual fault recursion graph, achieving two-dimensional visualization of fault features. It clearly displays the internal structure, periodicity, and abrupt change points of the sequence, while fully preserving the temporal dynamics and recursive characteristics of the fault. Furthermore, the actual fault recursion graphs converted from feature gas vector sequences of different fault types (such as intermediate-temperature superheating and low-energy discharge) exhibit significant visual differences in texture, patch structure, and distribution, demonstrating extremely high discriminative power. This enhances the identification of different fault features and provides input for subsequent model training. To increase the information dimension and visual effect of the image, the Euclidean distance value can be further mapped to grayscale or color gradients to generate grayscale or true-color recursion graphs. The Recurrence Plot (RP) algorithm used in this embodiment can identify faults with an accuracy of 96.7%, which is far superior to the relative position transformation method (91.9%) and the Markov transfer field method (94.1%). In particular, for some easily confused fault types, the RP image shows better distinguishing ability.

[0039] Furthermore, the step of acquiring historical dissolved gas analysis data in the oil of several high-frequency transformers, and extracting several characteristic gas content values ​​from the historical dissolved gas analysis data to obtain several characteristic gas vector sequences includes: Obtain historical dissolved gas analysis data in the oil of several high-frequency transformers; Data cleaning was performed on the historical dissolved gas analysis data in the oil to obtain dissolved gas analysis data in the pure oil. The dissolved gas analysis data in the pure oil are normalized to obtain the dissolved gas analysis data in the standard oil. Several characteristic gas content values ​​were extracted from the dissolved gas analysis data of standard oil, and several characteristic gas vector sequences were obtained.

[0040] In this embodiment, historical dissolved gas analysis data of several high-frequency transformers are obtained through an online transformer monitoring system or a historical database, including five key characteristic gases: hydrogen ( ), methane ( ), ethane ( ), ethylene ( ), acetylene ( The content value of ) (usually in units of This provides foundational data for subsequent data processing and feature extraction. Next, by cleaning historical dissolved gas analysis data in oil to obtain pure dissolved gas analysis data, noise reduction and outlier removal are performed, ensuring data validity and accuracy. Then, by normalizing the pure dissolved gas analysis data to obtain standard dissolved gas analysis data, the dimensions of the pure dissolved gas analysis data are eliminated and the scale is standardized, avoiding interference from data scale differences in subsequent feature processing. Finally, by extracting several characteristic gas content values ​​from the standard dissolved gas analysis data to obtain a characteristic gas vector sequence, high-frequency transformer fault features can be extracted, forming a one-dimensional fault feature sequence, providing suitable feature input for subsequent recursive graph conversion.

[0041] Furthermore, the process of generating an expanded dataset based on several actual fault recurrence graphs, several pre-acquired random noise vectors, and a pre-set initial depth convolutional generative adversarial network model includes: The preset initial depth convolutional generative adversarial network model includes an initial generator; Based on the initial generator, several pre-acquired random noise vectors are mapped to obtain several initial fake feature maps; The initial forged feature map is upsampled based on the initial generator to obtain an upgraded forged feature map. The upgraded forgery feature map is cropped based on the initial generator and the actual fault recursion map to obtain the forged fault recursion map; The pre-set initial depthwise convolutional generative adversarial network (GAN) model is trained based on the actual fault recurrence graph and the fake fault recurrence graph to obtain the standard depthwise convolutional GAN ​​model, and the augmented dataset is obtained based on the actual fault recurrence graph and the standard depthwise convolutional GAN ​​model.

[0042] In this embodiment, an initial generator samples a high-dimensional noise vector conforming to a standard normal distribution from a pre-constructed random noise space as the seed input for negative samples, obtaining a pre-acquired random noise vector. Next, the initial generator maps and reshapes the pre-acquired random noise vector through a fully connected layer to obtain an initial forged feature map, transforming disordered noise into feature representations and providing a basic feature carrier for generating the forged fault recurrence graph. Then, the initial generator upsamples the initial forged feature map through a series of transposed convolutional layers, followed by batch normalization layers and ReLU activation functions after each convolutional layer, obtaining an upgraded forged feature map. This increases the dimension and size of the initial forged feature map, making its feature scale approximate the actual fault recurrence graph, laying the foundation for subsequent steps. Finally, the output layer uses the Tanh activation function to crop the upgraded forged feature map using the actual fault recurrence graph, obtaining the forged fault recurrence graph. This ensures that the size of the forged fault feature map matches the actual fault recurrence graph, guaranteeing a consistent format between the forged and real samples. Finally, by combining the actual fault recurrence graph, the fake fault recurrence graph, and the preset initial deep convolutional generative adversarial network model to obtain the augmented dataset, it is possible to generate simulation samples that are highly consistent with the actual fault recurrence graph, thus completing the high-quality augmentation of the fault training samples. In this embodiment, after augmentation using the standard deep convolutional generative adversarial network model (DCGAN), the model achieves an accuracy of 97.1% on the fault classification validation set, which is very close to the accuracy on the fault classification training set (98.9%), and outperforms the methods without augmentation, autoencoder augmentation, and traditional GAN ​​augmentation, proving that the overfitting problem has been effectively solved.

[0043] Furthermore, the preset initial depthwise convolutional generative adversarial network model also includes an initial discriminator; the step of training the preset initial depthwise convolutional generative adversarial network model based on the actual fault recurrence graph and the fake fault recurrence graph to obtain a standard depthwise convolutional generative adversarial network model, and obtaining an expanded dataset based on the actual fault recurrence graph and the standard depthwise convolutional generative adversarial network model, includes: The preset initial depth convolutional generative adversarial network model also includes an initial discriminator; The initial discriminator and the initial generator are trained based on the actual fault recursion graph and the fake fault recursion graph respectively. The probability values ​​of the real recursion graph and the trained generator are obtained. If the probability values ​​of the real recursion graph satisfy the preset probability conditions and the trained generator satisfies Nash equilibrium, the training is completed and the fault generator and the fault discriminator are obtained to form a standard deep convolutional generative adversarial network model. An augmented dataset is obtained based on a real fault recursion graph and a standard deep convolutional generative adversarial network model.

[0044] In this embodiment, the initial discriminator and the initial generator are trained using actual and fake fault recurrence maps. The initial discriminator alternately receives actual and fake fault recurrence maps as input, and performs downsampling through a series of convolutional layers to extract texture features. Each layer uses the LeakyReLU activation function to avoid gradient sparsity. Finally, the true recurrence map probability value is output through a fully connected layer and a sigmoid function. During the training process, the entire adversarial training loop adopts a Min-Max game strategy. First, the parameter weights of the initial generator are fixed. The true recurrence map probability value is obtained by inputting the actual and fake fault recurrence maps into the initial discriminator. The parameter weights of the initial discriminator are updated with the goal of minimizing the true recurrence map probability value. After the update, the parameter weights of the initial generator are updated through backpropagation to obtain the trained generator. When the probability values ​​of the real recurrence graph meet the preset probability conditions and the generator satisfies Nash equilibrium, the training of the preset initial deep convolutional generative adversarial network (DBGAN) model is completed. At this point, the DBGAN model has reached Nash equilibrium and can generate fault samples with high similarity to the actual fault recurrence graph. Nash equilibrium means that the fake fault recurrence graph generated by the fault generator is indistinguishable from the actual fault recurrence graph in terms of texture distribution; the preset probability conditions mean that the probability value of the real recurrence graph of the fault discriminator is stable around 0.5. Then, using the actual fault recurrence graph and the trained standard DBGAN model, simulated fault recurrence graphs that are highly consistent with the actual fault recurrence graphs can be generated in batches, completing a high-quality expansion of the augmented dataset.

[0045] Further, the step of training a preset initial MobileNetV3 model based on the expanded dataset and the feature gas vector sequence to obtain several fault prediction probability vectors, constructing a cross-entropy loss function based on the expanded dataset and the fault prediction probability vectors, and then updating the preset initial MobileNetV3 model with the goal of minimizing the cross-entropy loss function until a preset convergence condition is met to obtain the MobileNetV3 model includes: The expanded dataset is divided according to a preset ratio to obtain a fault classification training set and a fault classification validation set; The feature gas vector sequence is matched to obtain several real fault classification labels; Based on the fault classification training set and the preset initial MobileNetV3 model, several fault prediction probability vectors are obtained. A cross-entropy loss function is constructed based on the real fault classification labels and fault prediction probability vectors. The preset initial MobileNetV3 model is updated with the goal of minimizing the cross-entropy loss function to obtain the MobileNetV3 training model. The MobileNetV3 trained model is validated based on the fault classification validation set to obtain the model accuracy. If the model accuracy meets the preset convergence condition, the training is completed and the MobileNetV3 model is obtained.

[0046] In this embodiment, a fault classification training set and a fault classification validation set are obtained by dividing the expanded dataset according to a preset ratio, providing data support for the subsequent training and convergence determination of the preset initial MobileNetV3 model. Next, several real fault classification labels are obtained by matching the feature gas vector sequences, providing a real reference for the construction of the loss function and the optimization of model parameters. Each feature gas vector sequence corresponds to a fault label, such as "normal," "low-energy discharge," "partial discharge," "high-energy discharge," "low-temperature overheating," "medium-temperature overheating," or "high-temperature overheating." Then, the fault classification training set data is shuffled and batch-packaged using a data loader, converted into tensor format, and input into the preset initial MobileNetV3 model to obtain several fault prediction probability vectors, providing prediction data for loss calculation. Subsequently, a cross-entropy loss function is constructed using the real fault classification labels and fault prediction probability vectors, and the initial MobileNetV3 model is updated with its minimization as the objective. Specifically, the gradient of the cross-entropy loss function with respect to all weight parameters in the model is calculated using an automatic differentiation mechanism, and the Adam optimizer is used to update the parameters based on the calculated gradients. This allows for iterative optimization of the model parameters through backpropagation, improving the accuracy of fault identification. The Adam optimizer combines the advantages of momentum and RMSProp, enabling adaptive adjustment of the learning rate. The cross-entropy loss function used in this embodiment is more suitable for multi-class classification problems than mean squared error, accelerating the convergence of gradient descent. Finally, after completing one training cycle, the model parameters are immediately frozen. The MobileNetV3 training model is validated using a fault classification validation set, and it is determined whether the model accuracy meets the preset convergence condition. This allows for testing the model's generalization ability and determining whether to stop training, ultimately resulting in a MobileNetV3 model that meets the accuracy requirements. The preset convergence condition specifically refers to the early stopping mechanism being triggered if the model accuracy no longer increases within a certain number of consecutive cycles. The improved MobileNetV3 model in this embodiment has only 1.3 × 10^6 parameters, making it about half the size of the traditional MobileNetV3 and two orders of magnitude smaller than AlexNet. Its training time is only 2376 seconds, demonstrating extremely high efficiency. Therefore, it is highly suitable for embedding into transformer online monitoring devices, edge computing gateways, or mobile inspection equipment to achieve real-time, localized intelligent fault diagnosis. This embodiment achieved an accuracy of 97.1% in benchmark tests, higher than other mainstream deep learning models (AlexNet 90.3%, VGG16 94.4%, and the original MobileNetV3 95.3%).

[0047] Furthermore, based on the fault classification training set and the preset initial MobileNetV3 model, several fault prediction probability vectors are obtained, including: The preset initial MobileNetV3 model includes a dimensionality increase module, a feature extraction module, an attention enhancement module, and a dimensionality reduction module; Based on the dimension-upgrading module, the fault classification training set is mapped to a high dimension to obtain several initial high-dimensional feature maps; Based on the feature extraction module, spatial features are extracted from the initial high-dimensional feature map to obtain the basic fault feature map; An enhanced fault feature map is obtained by adaptively optimizing the basic fault feature map based on the attention enhancement module. The enhanced fault feature map is processed by channel dimensionality reduction module to obtain the dimensionality-reduced fault feature map. The reduced-dimensional fault feature map and the fault classification training set are residually concatenated to obtain a global fusion feature map. The global fusion feature map is sequentially subjected to global average pooling, fully connected layer feature mapping, and Softmax activation function calculation to obtain several fault prediction probability vectors.

[0048] In this embodiment, please refer to Figure 2An initial high-dimensional feature map is obtained by using a dimensionality-upgrading module with a 1x1 convolutional kernel and an h-swish function to perform high-dimensional mapping on the fault classification training set. This enables channel-wise dimensionality upgrading and high-dimensional spatial representation of fault features, laying the foundation for the extraction of deep fault features. The h-swish function is used to balance accuracy and speed. Next, a feature extraction module using a 5x5 convolutional kernel and an h-swish function extracts spatial features from the initial high-dimensional feature map to obtain a basic fault feature map. This accurately captures the core spatial texture features of the initial high-dimensional feature map, achieving a concrete extraction of fault features. Then, an attention enhancement module adaptively optimizes the basic fault feature map to obtain an enhanced fault feature map. This enhances key fault features and suppresses redundant features, improving the discriminative power of fault features. In this embodiment, the attention enhancement module replaces or enhances the original network's attention mechanism with CBAM, a lightweight and general-purpose module that infers attention maps sequentially along two independent dimensions: channel dimension and spatial dimension. These maps are then multiplied onto the input feature map for adaptive feature optimization. Specifically, adding CBAM after the feature extraction module helps the model capture more global and discriminative features. Subsequently, the enhanced fault feature map is subjected to channel dimensionality reduction by the dimensionality reduction module to obtain a dimensionality-reduced fault feature map. This reduces the feature dimension while preserving the core fault features, thus lowering the model's computational complexity. Next, a global fusion feature map is obtained by residually connecting the dimensionality-reduced fault feature map and the fault classification training set. This fuses shallow and deep fault features, improving the overall feature representation capability. Finally, by sequentially performing global average pooling, fully connected layer feature mapping, and Softmax activation function calculation on the global fusion feature map, the fused fault features are transformed into a fault category probability distribution, yielding a fault prediction probability vector corresponding to the fault classification training set.

[0049] Furthermore, the step of inputting the recursive graph of the high-frequency transformer under test into the MobileNetV3 model to obtain the fault detection results of the high-frequency transformer under test includes: The dissolved gas analysis data in the oil of the high-frequency transformer under test is converted to obtain the fault recursion graph and input into the MobileNetV3 model to obtain the fault detection results of the high-frequency transformer under test.

[0050] In this embodiment, by converting the dissolved gas analysis data in the oil of the high-frequency transformer under test into a fault recurrence graph, the fault characteristics of the high-frequency transformer under test can be visualized in two dimensions. Simultaneously, the converted fault recurrence graph provides suitable feature input for MobileNetV3 model inference. Next, by inputting the fault recurrence graph into the MobileNetV3 model, the fault type of the high-frequency transformer under test can be inferred and determined. The MobileNetV3 model outputs a probability distribution of the fault type, and the category with the highest probability is taken as the final fault detection result, achieving high-precision and high-reliability identification of various faults in the high-frequency transformer under test.

[0051] Please see Figure 3 This embodiment provides a high-frequency transformer fault identification system, including a fault recursion graph acquisition unit, a fault sample expansion unit, a MobileNetV3 model training and optimization unit, and a fault detection unit, specifically: The fault recursion graph acquisition unit is used to acquire several characteristic gas vector sequences of several high-frequency transformers and convert the several characteristic gas vector sequences into several actual fault recursion graphs. The fault sample expansion unit is used to generate an expanded dataset based on several actual fault recursion graphs, several pre-acquired random noise vectors, and a preset initial depth convolutional generative adversarial network model. The MobileNetV3 model training and optimization unit is used to train a preset initial MobileNetV3 model based on the expanded dataset and the feature gas vector sequence, obtain several fault prediction probability vectors, construct a cross-entropy loss function based on the expanded dataset and the fault prediction probability vectors, and then update the preset initial MobileNetV3 model with the goal of minimizing the cross-entropy loss function until the preset convergence condition is met, and obtain the MobileNetV3 model. The fault detection unit is used to input the fault recursion graph of the high-frequency transformer under test into the MobileNetV3 model to obtain the fault detection result of the high-frequency transformer under test.

[0052] This embodiment provides a high-frequency transformer fault identification system. In practical applications, it only requires a fault recursion graph acquisition unit to obtain a characteristic gas vector sequence. The recurrence plot (RP) algorithm is used to convert the fault into a real fault recurrence plot, which enables two-dimensional visualization of fault features, enhances the identification of fault features, and provides input data for subsequent model training. Simultaneously, from the feature representation level, it lays the foundation for avoiding low accuracy in high-frequency transformer fault identification and insufficient ability to identify complex faults. Next, a fault sample augmentation unit is employed. By combining the real fault recurrence plot, pre-acquired random noise vectors, and a pre-set initial depth convolutional generative adversarial network model, an augmented dataset is generated. This expands the fault training and verification samples, effectively supplementing the samples of complex faults and addressing the problems of low fault identification accuracy and insufficient ability to identify complex faults from a data perspective. Then, the MobileNetV3 model training optimization unit is used. The initial MobileNetV3 model is trained by expanding the dataset and using feature gas vector sequences. Fault prediction probability vectors are generated, and a cross-entropy loss function is constructed to optimize the model until convergence. This yields a MobileNetV3 model with high accuracy and complex fault identification capabilities. It maintains high accuracy in identifying various typical faults (such as low-energy discharge, high-energy discharge, partial discharge, and overheating at different temperatures), exhibiting good robustness and generalization ability, making it applicable to real-world data collected from the field. Finally, a fault detection unit is used. The recursive graph of the fault to be tested in the high-frequency transformer is input into the MobileNetV3 model to obtain the fault detection results. This achieves high-precision and high-reliability identification of various faults in the high-frequency transformer under test. The average fault identification accuracy of this embodiment reaches 93.46%, and the accuracy for each fault type reaches over 90.63%, demonstrating the powerful generalization ability of the MobileNetV3 model in this embodiment for real and complex data, and possessing extremely high engineering application value.

[0053] Furthermore, the fault recurrence graph acquisition unit is used to acquire several characteristic gas vector sequences of several high-frequency transformers, and convert the several characteristic gas vector sequences into several actual fault recurrence graphs, including: Obtain historical dissolved gas analysis data of several high-frequency transformers, and extract several characteristic gas content values ​​from the historical dissolved gas analysis data to obtain several characteristic gas vector sequences; Pair any two characteristic gas vectors in the characteristic gas vector sequence to obtain several vector pairs, and calculate the Euclidean distance between the vector pairs, so as to convert several characteristic gas vector sequences into several actual fault recursion graphs based on the Euclidean distance between the vector pairs.

[0054] In this embodiment, by acquiring historical dissolved gas analysis data from the oil of a high-frequency transformer and extracting characteristic gas content values ​​to obtain a characteristic gas vector sequence, fault characteristics of the high-frequency transformer can be extracted, providing data support for subsequent visualization and conversion of fault characteristics. Next, the characteristic gas vector sequence is mapped onto a two-dimensional image. By pairing any two characteristic gas vectors in the sequence and calculating their Euclidean distance, points with an Euclidean distance less than a preset distance threshold are considered... In the image, pixels marked as black (or dark) indicate that the two states are "recursive"; otherwise, they are marked as white (or light). This allows the one-dimensional feature gas vector sequence to be converted into an actual fault recursion graph, achieving two-dimensional visualization of fault features. It clearly displays the internal structure, periodicity, and abrupt change points of the sequence, while fully preserving the temporal dynamics and recursive characteristics of the fault. Furthermore, the actual fault recursion graphs converted from feature gas vector sequences of different fault types (such as intermediate-temperature superheating and low-energy discharge) exhibit significant visual differences in texture, patch structure, and distribution, demonstrating extremely high discriminative power. This enhances the identification of different fault features and provides input for subsequent model training. To increase the information dimension and visual effect of the image, the Euclidean distance value can be further mapped to grayscale or color gradients to generate grayscale or true-color recursion graphs. The Recurrence Plot (RP) algorithm used in this embodiment can identify faults with an accuracy of 96.7%, which is far superior to the relative position transformation method (91.9%) and the Markov transfer field method (94.1%). In particular, for some easily confused fault types, the RP image shows better distinguishing ability.

[0055] Furthermore, the step of acquiring historical dissolved gas analysis data in the oil of several high-frequency transformers, and extracting several characteristic gas content values ​​from the historical dissolved gas analysis data to obtain several characteristic gas vector sequences includes: Obtain historical dissolved gas analysis data in the oil of several high-frequency transformers; Data cleaning was performed on the historical dissolved gas analysis data in the oil to obtain dissolved gas analysis data in the pure oil. The dissolved gas analysis data in the pure oil are normalized to obtain the dissolved gas analysis data in the standard oil. Several characteristic gas content values ​​were extracted from the dissolved gas analysis data of standard oil, and several characteristic gas vector sequences were obtained.

[0056] In this embodiment, historical dissolved gas analysis data of several high-frequency transformers are obtained through an online transformer monitoring system or a historical database, including five key characteristic gases: hydrogen ( ), methane ( ), ethane ( ), ethylene ( ), acetylene ( The content values ​​of dissolved gases (usually in μL / L) provide the foundation for subsequent data processing and feature extraction. Next, by cleaning historical dissolved gas analysis data in oil to obtain pure oil dissolved gas analysis data, noise reduction and outlier removal are performed, ensuring data validity and accuracy. Then, by normalizing the pure oil dissolved gas analysis data to obtain standard oil dissolved gas analysis data, the dimensions of the pure oil dissolved gas analysis data are eliminated and the scale is standardized, avoiding interference from data scale differences in subsequent feature processing. Finally, by extracting several characteristic gas content values ​​from the standard oil dissolved gas analysis data to obtain a characteristic gas vector sequence, the fault features of high-frequency transformers can be extracted, forming a one-dimensional fault feature sequence, providing suitable feature input for subsequent recursive graph conversion.

[0057] Furthermore, the fault sample augmentation unit is used to generate an augmented dataset based on several actual fault recurrence graphs, several pre-acquired random noise vectors, and a preset initial depth convolutional generative adversarial network model, including: The preset initial depth convolutional generative adversarial network model includes an initial generator; Based on the initial generator, several pre-acquired random noise vectors are mapped to obtain several initial fake feature maps; The initial forged feature map is upsampled based on the initial generator to obtain an upgraded forged feature map. The upgraded forgery feature map is cropped based on the initial generator and the actual fault recursion map to obtain the forged fault recursion map; The pre-set initial depthwise convolutional generative adversarial network (GAN) model is trained based on the actual fault recurrence graph and the fake fault recurrence graph to obtain the standard depthwise convolutional GAN ​​model, and the augmented dataset is obtained based on the actual fault recurrence graph and the standard depthwise convolutional GAN ​​model.

[0058] In this embodiment, an initial generator samples a high-dimensional noise vector conforming to a standard normal distribution from a pre-constructed random noise space as the seed input for negative samples, obtaining a pre-acquired random noise vector. Next, the initial generator maps and reshapes the pre-acquired random noise vector through a fully connected layer to obtain an initial forged feature map, transforming disordered noise into feature representations and providing a basic feature carrier for generating the forged fault recurrence graph. Then, the initial generator upsamples the initial forged feature map through a series of transposed convolutional layers, followed by batch normalization layers and ReLU activation functions after each convolutional layer, obtaining an upgraded forged feature map. This increases the dimension and size of the initial forged feature map, making its feature scale approximate the actual fault recurrence graph, laying the foundation for subsequent steps. Finally, the output layer uses the Tanh activation function to crop the upgraded forged feature map using the actual fault recurrence graph, obtaining the forged fault recurrence graph. This ensures that the size of the forged fault feature map matches the actual fault recurrence graph, guaranteeing a consistent format between the forged and real samples. Finally, by combining the actual fault recurrence graph, the fake fault recurrence graph, and the preset initial deep convolutional generative adversarial network model to obtain the augmented dataset, it is possible to generate simulation samples that are highly consistent with the actual fault recurrence graph, thus completing the high-quality augmentation of the fault training samples. In this embodiment, after augmentation using the standard deep convolutional generative adversarial network model (DCGAN), the model achieves an accuracy of 97.1% on the fault classification validation set, which is very close to the accuracy on the fault classification training set (98.9%), and outperforms the methods without augmentation, autoencoder augmentation, and traditional GAN ​​augmentation, proving that the overfitting problem has been effectively solved.

[0059] This embodiment also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the functions of the system as described above.

[0060] It is understood that the above system embodiments correspond to the method embodiments of the present invention, and can implement the high-frequency transformer fault identification method provided by any of the above method embodiments of the present invention.

[0061] It should be noted that the system embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0062] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for identifying faults in a high-frequency transformer, characterized in that, include: Obtain several characteristic gas vector sequences of several high-frequency transformers, and convert the several characteristic gas vector sequences into several actual fault recursion graphs; Based on several real-world fault recursion graphs, several pre-acquired random noise vectors, and a pre-set initial depth convolutional generative adversarial network model, an expanded dataset is generated. The preset initial MobileNetV3 model is trained based on the expanded dataset and the feature gas vector sequence to obtain several fault prediction probability vectors. A cross-entropy loss function is constructed based on the expanded dataset and the fault prediction probability vectors. The preset initial MobileNetV3 model is then updated with the goal of minimizing the cross-entropy loss function until the preset convergence condition is met, and the MobileNetV3 model is obtained. Input the fault recursion graph of the high-frequency transformer to be tested into the MobileNetV3 model to obtain the fault detection results of the high-frequency transformer to be tested.

2. The high-frequency transformer fault identification method according to claim 1, characterized in that, The process of acquiring several characteristic gas vector sequences of several high-frequency transformers and converting these sequences into several actual fault recursion graphs includes: Obtain historical dissolved gas analysis data of several high-frequency transformers, and extract several characteristic gas content values ​​from the historical dissolved gas analysis data to obtain several characteristic gas vector sequences; Pair any two characteristic gas vectors in the characteristic gas vector sequence to obtain several vector pairs, and calculate the Euclidean distance between the vector pairs, so as to convert several characteristic gas vector sequences into several actual fault recursion graphs based on the Euclidean distance between the vector pairs.

3. The high-frequency transformer fault identification method according to claim 2, characterized in that, The process involves acquiring historical dissolved gas analysis data from several high-frequency transformer oils, extracting several characteristic gas content values ​​from the historical dissolved gas analysis data, and obtaining several characteristic gas vector sequences, including: Obtain historical dissolved gas analysis data in the oil of several high-frequency transformers; Data cleaning was performed on the historical dissolved gas analysis data in the oil to obtain dissolved gas analysis data in the pure oil. The dissolved gas analysis data in the pure oil are normalized to obtain the dissolved gas analysis data in the standard oil. Several characteristic gas content values ​​were extracted from the dissolved gas analysis data of standard oil, and several characteristic gas vector sequences were obtained.

4. The high-frequency transformer fault identification method according to claim 1, characterized in that, The process of generating an expanded dataset based on several actual fault recurrence graphs, several pre-acquired random noise vectors, and a pre-set initial depth convolutional generative adversarial network model includes: The preset initial depth convolutional generative adversarial network model includes an initial generator; Based on the initial generator, several pre-acquired random noise vectors are mapped to obtain several initial fake feature maps; The initial forged feature map is upsampled based on the initial generator to obtain an upgraded forged feature map. The upgraded forgery feature map is cropped based on the initial generator and the actual fault recursion map to obtain the forged fault recursion map; The pre-set initial depthwise convolutional generative adversarial network (GAN) model is trained based on the actual fault recurrence graph and the fake fault recurrence graph to obtain the standard depthwise convolutional GAN ​​model, and the augmented dataset is obtained based on the actual fault recurrence graph and the standard depthwise convolutional GAN ​​model.

5. The high-frequency transformer fault identification method according to claim 4, characterized in that, The preset initial depthwise convolutional generative adversarial network (GAN) model further includes an initial discriminator; the process of training the preset initial depthwise convolutional GAN ​​model based on actual fault recurrence graphs and fake fault recurrence graphs to obtain a standard depthwise convolutional GAN ​​model, and obtaining an expanded dataset based on the actual fault recurrence graphs and the standard depthwise GAN model, includes: The preset initial depth convolutional generative adversarial network model also includes an initial discriminator; The initial discriminator and the initial generator are trained based on the actual fault recursion graph and the fake fault recursion graph respectively. The probability values ​​of the real recursion graph and the trained generator are obtained. If the probability values ​​of the real recursion graph satisfy the preset probability conditions and the trained generator satisfies Nash equilibrium, the training is completed and the fault generator and the fault discriminator are obtained to form a standard deep convolutional generative adversarial network model. An augmented dataset is obtained based on a real fault recursion graph and a standard deep convolutional generative adversarial network model.

6. The high-frequency transformer fault identification method according to claim 1, characterized in that, The process involves training a preset initial MobileNetV3 model based on the expanded dataset and feature gas vector sequences to obtain several fault prediction probability vectors. A cross-entropy loss function is then constructed based on the expanded dataset and the fault prediction probability vectors. The preset initial MobileNetV3 model is updated with the goal of minimizing the cross-entropy loss function until a preset convergence condition is met, thus obtaining the MobileNetV3 model. This process includes: The expanded dataset is divided according to a preset ratio to obtain a fault classification training set and a fault classification validation set; The feature gas vector sequence is matched to obtain several real fault classification labels; Based on the fault classification training set and the preset initial MobileNetV3 model, several fault prediction probability vectors are obtained. A cross-entropy loss function is constructed based on the real fault classification labels and fault prediction probability vectors. The preset initial MobileNetV3 model is updated with the goal of minimizing the cross-entropy loss function to obtain the MobileNetV3 training model. The MobileNetV3 trained model is validated based on the fault classification validation set to obtain the model accuracy. If the model accuracy meets the preset convergence condition, the training is completed and the MobileNetV3 model is obtained.

7. The high-frequency transformer fault identification method according to claim 6, characterized in that, Based on the fault classification training set and a preset initial MobileNetV3 model, several fault prediction probability vectors are obtained, including: The preset initial MobileNetV3 model includes a dimensionality increase module, a feature extraction module, an attention enhancement module, and a dimensionality reduction module; Based on the dimension-upgrading module, the fault classification training set is mapped to a high dimension to obtain several initial high-dimensional feature maps; Based on the feature extraction module, spatial features are extracted from the initial high-dimensional feature map to obtain the basic fault feature map; An enhanced fault feature map is obtained by adaptively optimizing the basic fault feature map based on the attention enhancement module. The enhanced fault feature map is processed by channel dimensionality reduction module to obtain the dimensionality-reduced fault feature map. The reduced-dimensional fault feature map and the fault classification training set are residually concatenated to obtain a global fusion feature map. The global fusion feature map is sequentially subjected to global average pooling, fully connected layer feature mapping, and Softmax activation function calculation to obtain several fault prediction probability vectors.

8. The high-frequency transformer fault identification method according to claim 1, characterized in that, The step of inputting the recursive graph of the high-frequency transformer under test into the MobileNetV3 model to obtain the fault detection results of the high-frequency transformer under test includes: The dissolved gas analysis data in the oil of the high-frequency transformer under test is converted to obtain the fault recursion graph and input into the MobileNetV3 model to obtain the fault detection results of the high-frequency transformer under test.

9. A high-frequency transformer fault identification system, characterized in that, It includes a fault recursion graph acquisition unit, a fault sample expansion unit, a MobileNetV3 model training and optimization unit, and a fault detection unit, specifically: The fault recursion graph acquisition unit is used to acquire several characteristic gas vector sequences of several high-frequency transformers and convert the several characteristic gas vector sequences into several actual fault recursion graphs. The fault sample expansion unit is used to generate an expanded dataset based on several actual fault recursion graphs, several pre-acquired random noise vectors, and a preset initial depth convolutional generative adversarial network model. The MobileNetV3 model training and optimization unit is used to train a preset initial MobileNetV3 model based on the expanded dataset and the feature gas vector sequence, obtain several fault prediction probability vectors, construct a cross-entropy loss function based on the expanded dataset and the fault prediction probability vectors, and then update the preset initial MobileNetV3 model with the goal of minimizing the cross-entropy loss function until the preset convergence condition is met, and obtain the MobileNetV3 model. The fault detection unit is used to input the fault recursion graph of the high-frequency transformer under test into the MobileNetV3 model to obtain the fault detection result of the high-frequency transformer under test.

10. A high-frequency transformer fault identification system according to claim 9, characterized in that, The fault recurrence graph acquisition unit is used to acquire several characteristic gas vector sequences of several high-frequency transformers and convert the several characteristic gas vector sequences into several actual fault recurrence graphs, including: Obtain historical dissolved gas analysis data of several high-frequency transformers, and extract several characteristic gas content values ​​from the historical dissolved gas analysis data to obtain several characteristic gas vector sequences; Pair any two characteristic gas vectors in the characteristic gas vector sequence to obtain several vector pairs, and calculate the Euclidean distance between the vector pairs, so as to convert several characteristic gas vector sequences into several actual fault recursion graphs based on the Euclidean distance between the vector pairs.

Images