A transformer equipment fault diagnosis method, system, device and medium

Abstract

The application discloses a transformer equipment fault diagnosis method, system, device and medium, relates to the electric power equipment detection technical field, and the method comprises the following steps: integrating various operation fault cases of transformer equipment, constructing an initial question and answer pair sample set and inputting a preset large model for training, introducing a preset dynamic knowledge extractor to perform text entity extraction on the initial question and answer pair sample set during the training process, and generating an equipment fault atlas; converting the equipment fault atlas vector and inputting it into an attention layer reconstructed by rotation position coding to capture the equipment coupling relationship; performing parameter fine-tuning optimization on the output result of the attention layer and iterating the training process to obtain a trained fault diagnosis large model; and collecting the operation data of the transformer equipment in real time and inputting the fault diagnosis large model to generate a high-precision fault diagnosis result and a maintenance strategy.

Description

Technical Field

[0001] This invention relates to the field of power equipment testing technology, and in particular to a method, system, equipment and medium for diagnosing transformer equipment faults. Background Technology

[0002] Transformers are the core hubs and stability pillars of power systems, and their operating status directly determines the safety boundaries of the power grid, power quality, and energy transmission efficiency. In the context of new power system construction, transformers, as key hubs for energy conversion and voltage regulation, have made their operational status monitoring and fault detection a crucial technical defense for ensuring the safety and resilience of large power grids.

[0003] Traditional fault diagnosis methods rely on human experience, which is not only time-consuming and labor-intensive, but also inefficient and highly subjective, resulting in large errors in the test results. This makes it difficult to formulate effective and reasonable maintenance strategies and affects the stable operation of power equipment.

[0004] Therefore, how to effectively diagnose transformer equipment and improve the accuracy of fault diagnosis has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] This invention provides a method, system, equipment, and medium for diagnosing transformer equipment faults, addressing how to improve the accuracy of fault diagnosis by optimizing the training process of a large model.

[0006] To address the aforementioned technical problems, embodiments of the present invention provide a method for diagnosing transformer equipment faults, including:

[0007] Integrate and analyze various operational failure cases of transformer equipment in the target power system to construct an initial question-and-answer pair sample set;

[0008] The initial question-and-answer pair sample set is input into a preset large model for training. During the training process, a preset dynamic knowledge extractor is introduced to extract text entities from the initial question-and-answer pair sample set to generate a device fault map. The dynamic knowledge extractor is composed of a distillation mechanism and a graph neural network.

[0009] The device fault map vector is transformed and then input into the attention layer reconstructed by rotational position encoding to capture device coupling relationships;

[0010] The output of the attention layer is fine-tuned and optimized by parameters, and the training process is iterated to obtain a well-trained fault diagnosis model.

[0011] Real-time acquisition of transformer equipment operation data is input into the fault diagnosis model to generate fault diagnosis results and maintenance strategies.

[0012] Furthermore, the integrated analysis of various operational fault cases of transformer equipment in the target power system constructs an initial question-and-answer pair sample set, including:

[0013] Collect transformer operation procedures, transformer test specifications data and historical transformer fault data, and annotate them with question-and-answer pairs to integrate them into operational fault cases;

[0014] Generative adversarial networks are used to augment the runtime failure cases to construct the initial question-answer pair sample set.

[0015] Furthermore, the process of generating the equipment fault map includes:

[0016] After the initial question and answer pairs are sequence-transformed, they are input into a preset large model. A distillation mechanism is introduced to dynamically optimize the weights of transformer equipment operation indicators and generate fault knowledge vectors.

[0017] The driving graph neural network extracts text entities from the fault knowledge vector to construct node features and edge features, thereby generating the equipment fault map.

[0018] Furthermore, the step of transforming the device fault map vector and inputting it into the attention layer reconstructed by rotational position encoding to capture device coupling relationships includes:

[0019] The vector representation of the device fault map is input into the attention layer and feature normalization is performed.

[0020] The normalization result is mapped to a rotation angle using a learnable frequency parameter; wherein the rotation angle is designed to dynamically correct the rotation angle weight using a gating network under a preset index threshold condition.

[0021] The rotation matrix is ​​generated by weighted analysis of the rotation angles.

[0022] During attention calculation, the rotation matrix is ​​used to perform encoding reconstruction on the query matrix and key matrix to generate a reconstructed attention matrix to capture the device coupling relationship.

[0023] Furthermore, the process of fine-tuning and optimizing the parameters includes:

[0024] An expert routing network is deployed to predict the probability of each feature vector in the reconstructed attention matrix, and feature vectors with probability values ​​exceeding preset conditions are used as activated experts.

[0025] The DoRA fine-tuning mechanism is introduced to decompose the original weights of the activated expert into magnitude and direction components;

[0026] The direction component is updated in low rank, and the updated direction component is merged with the amplitude component to optimize the training.

[0027] Furthermore, the process of fine-tuning and decoding the output of the attention layer and iteratively training iteratively also includes:

[0028] During iterative training, a greedy algorithm is integrated into the output layer of a pre-defined large model to select the result with the highest predicted probability from the output probability distribution as the target output result.

[0029] Furthermore, after real-time acquisition of transformer equipment operating data and input into the fault diagnosis model to generate fault diagnosis results and maintenance strategies, the process also includes:

[0030] The execution status of the maintenance strategy is fed back to the monitoring terminal;

[0031] The equipment fault map is updated based on the feedback results, and the fault diagnosis model is iteratively optimized.

[0032] Another embodiment of the present invention provides a transformer equipment fault diagnosis system, comprising:

[0033] The training sample construction module is used to integrate and analyze various operational failure cases of transformer equipment in the target power system and construct an initial question-answer pair sample set;

[0034] The text entity extraction module is used to input the initial question-and-answer pair sample set into a preset large model for training. During the training process, a preset dynamic knowledge extractor is introduced to extract text entities from the initial question-and-answer pair sample set and generate a device fault map. The dynamic knowledge extractor is composed of a distillation mechanism and a graph neural network.

[0035] The attention reconstruction module is used to transform the device fault map vector and input it into the attention layer reconstructed by rotation position encoding to capture device coupling relationships;

[0036] The iterative training module is used to fine-tune and optimize the output of the attention layer and iteratively train the model to obtain a well-trained fault diagnosis model.

[0037] The fault diagnosis module is used to collect real-time operating data of transformer equipment and input it into the fault diagnosis model to generate fault diagnosis results and maintenance strategies.

[0038] Another embodiment of the present invention provides a computer device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the transformer equipment fault diagnosis method as described above.

[0039] In another embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, wherein when the device containing the computer-readable storage medium executes the computer program, the transformer equipment fault diagnosis method described above is implemented.

[0040] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following:

[0041] This invention introduces generative adversarial networks to generate rare fault samples, constructing a question-answering sample database containing a large amount of professional knowledge in transformer fault diagnosis. This database can more comprehensively cover the fault types of transformer equipment and provide sufficient data input for model training. During model training, a dynamic knowledge distillation mechanism is adopted to construct a domain knowledge extractor based on graph neural networks, which improves knowledge transfer efficiency while ensuring the accuracy of model recognition. A hybrid optimization mechanism of multi-head self-attention, rotation coding, and expert routing + fine-tuning is introduced to greatly reduce the computational complexity and false alarm rate of model training. Finally, the trained fault diagnosis model is used to diagnose transformer equipment, thereby formulating accurate maintenance strategies and maintaining the stable operation of the power grid. Attached Figure Description

[0042] Figure 1 This is a schematic flowchart of a transformer equipment fault diagnosis method in one embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the encoding reconstruction architecture of the attention mechanism in one embodiment of the present invention;

[0044] Figure 3 This is a schematic diagram of the parameter fine-tuning and optimization architecture in one embodiment of the present invention;

[0045] Figure 4 This is a schematic diagram of the component decomposition architecture in parameter fine-tuning optimization in one embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram of the evaluation results of the model maintenance strategy generation capability in one embodiment of the present invention;

[0047] Figure 6 This is a schematic diagram of the structure of a transformer equipment fault diagnosis system in one embodiment of the present invention;

[0048] Figure 7 A structural block diagram of a preferred embodiment of a computer device provided by the present invention;

[0049] Figure labels: Training sample construction module M1, text entity extraction module M2, attention reconstruction module M3, iterative training module M4, fault diagnosis module M5. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0051] In the description of this application, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0052] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0053] One embodiment of the present invention provides a method for diagnosing transformer equipment faults. For details, please refer to [link to relevant documentation]. Figure 1 , Figure 1 The embodiment shown includes the following steps:

[0054] S1. Integrate and analyze various operational failure cases of transformer equipment in the target power system to construct an initial question-and-answer pair sample set.

[0055] Data on transformer operation procedures, transformer testing specifications, and historical transformer faults from the target power system are collected and annotated into question-and-answer pairs. These pairs are then integrated into operational fault cases required for model training. Specifically, this transformer operation procedure data, transformer testing specification data, and historical transformer fault data can include records of transformer defects and abnormal parameter values ​​from the past three years, or oil chromatography reports, etc. In other words, the integrated operational fault cases can reflect the description of the type, cause, and state of transformer faults.

[0056] For example, in some embodiments of the present invention, the number of characters acquired from the transformer operation and maintenance manual is 1.4 × The data includes transformer operation procedures such as daily operation, monitoring, maintenance, and overhaul; transformer test specification data from the IEC 60076 series standards or the GB 1094 series power transformer test specifications; and 2,318 typical transformer fault cases and defect records selected from power systems of multiple power supply bureaus, including historical transformer fault data such as 521 bushing faults, 732 transformer body faults, 496 protection device faults, 276 voltage regulating device faults, and 251 cooling device faults.

[0057] Bushing-related faults can include: water seepage and contamination caused by moisture or other pollutants entering the bushing's interior or surface, or insulation breakdown caused by insulation material failure creating a conductive path; transformer-related faults include mechanical damage or corrosion of the oil tank leading to shell damage; protection device faults commonly include relay malfunctions or false trips; voltage regulating device faults include mechanical failures of tap changer components such as switch levers or drive mechanisms; and cooling device faults include oil temperature rise caused by oil pump failure in the oil cooling system.

[0058] This embodiment also provides a specific example of the following question-and-answer pair (QA pair):

[0059] Q: What standards should be followed in maintaining the on-load tap changer of a transformer?

[0060] A: Maintenance of the on-load tap changer of the transformer should be carried out according to the manufacturer's specifications. If there are no manufacturer's specifications, the relevant standard DUT574 can be used as a reference for maintenance. Maintenance work includes removing oxide film and oil stains, measuring the DC resistance of the windings, and recording the tap change status. By maintaining according to the standard, the normal operation and use of the transformer tap changer can be ensured.

[0061] Generative Adversarial Networks (CGANs) are used to augment operational failure cases to construct an initial question-answer pair sample set. This step essentially utilizes CGANs to synthesize rare failure samples to expand the question-answer pairs. For example, in this embodiment, the number of question-answer samples related to cooling device failures is the lowest, so CGANs can be used to augment their data. The generator and discriminator of the CGAN include the following formulas:

[0062]

[0063]

[0064]

[0065] In the formula, Here, x represents the positive slope coefficient, y represents the input sample to the discriminator, z represents the explicit transformer fault type condition label, and z represents the noise vector. For the Sigmoid function, W k Let be the weight matrix of the k-th layer. Here, is the normalization function, Concat is the concatenation function, G is the mathematical expression formula for the generator, and D is the mathematical expression formula for the discriminator.

[0066] Furthermore, CGAN will introduce conditional constraints, namely the objective function V, as follows:

[0067]

[0068] In the formula, P data P represents actual transformer fault data. z For noise distribution data, It is the expectation of the actual data distribution. It is the expectation of the noise distribution.

[0069] In this embodiment, the conditional information "y" introduced by CGAN can be a specific transformer fault type label, such as bushing faults, transformer body faults, cooling device faults, voltage regulating device faults, protection device faults, etc. During the generation phase, the generator G concatenates random noise z with y to generate fault text, such as a simulated cooling system fault sample: "A 220kV transformer's fan failed to start during operation, causing the oil temperature to rise by 5°C, triggering a cooling system fault alarm." The discriminator D then evaluates whether the input sample meets the specified conditional label, thereby training the generator to generate high-fidelity question-and-answer data with specific fault directions.

[0070] S2. Input the initial question-and-answer pair sample set into a pre-defined large model for training. During the training process, a pre-defined dynamic knowledge extractor is introduced to extract text entities from the initial question-and-answer pair sample set and generate a device fault map. The dynamic knowledge extractor is composed of a distillation mechanism and a graph neural network.

[0071] First, the generated initial question-and-answer pairs sample sets need to undergo preprocessing operations such as data cleaning. After the sequences are converted into text vectors, they are embedded into a pre-set large model for training. In this embodiment, the preferred large model is the Qwen-7B-Chat open-source large model.

[0072] Within the large model, a dynamic knowledge extractor, composed of a distillation mechanism and a graph neural network, is first introduced to optimize the knowledge transfer efficiency during the training process. Specifically, this embodiment introduces a distillation mechanism to dynamically adjust the weights of the transformer equipment operation indicators reflected in the vector, in order to generate fault knowledge vectors.

[0073] For example, three types of teacher models trained using this distillation mechanism are processed in parallel: an electrical characteristic model, a mechanical state model, and a chemical analysis model. In this example, the electrical characteristic model outputs the transformer discharge / overheating probability distribution, the mechanical state model outputs the transformer component degradation score, and the chemical analysis model outputs the correlation degree of oil cracking products. Then, a gating network dynamically assigns weights to the teacher models based on text keywords. For example, when "acetylene" appears, the weight of the chemical teacher model is increased to 0.82. Finally, the weighted models are fused to form a fault knowledge vector.

[0074] It should be understood that in the process of adjusting the weights, this embodiment also employs a cosine annealing strategy to adjust the loss weights generated by the distillation mechanism, as shown in the following formula:

[0075]

[0076] In the formula, where The distillation loss weights are t, the current training step, and T, the total number of steps. This adjustment strategy allows the model to gradually transition from relying on supervision in the early stages to autonomous optimization.

[0077] Next, the driving graph neural network extracts text entities from the fault knowledge vectors to construct node and edge features, generating a device fault map. For example, from "fan stoppage causes oil temperature to rise," two entities, "cooling fan" and "oil temperature," can be extracted. The relationship between them is "causes rise," forming a knowledge graph structure centered on "component—fault type—causal relationship." The graph neural network uses these entities and relationships to construct a node-edge graph structure. The node representations obtain contextual features through multi-layer aggregation and are finally embedded into vector form, which is then input into the attention mechanism module of the larger model.

[0078] S3. After transforming the equipment fault map vector, input it into the attention layer reconstructed by rotational position encoding to capture the equipment coupling relationship.

[0079] This step aims to introduce a rotational position encoding mechanism, called the Trans-RoPE mechanism, into the multi-head self-attention layer of a large model. This mechanism captures the multi-dimensional coupling relationships between various indicators such as transformer oil temperature, load, and operating time, generating the final attention output. Specifically, the vector representation of the device fault map is input into the attention layer, and feature normalization is performed. The normalization result is then mapped to a rotation angle using a learnable frequency parameter.

[0080] For example, the equipment fault map is vectorized (dimension 16×1024×4096), and this vector is accompanied by M dynamic indicators / features, such as oil temperature t. i , load l i Running time h iNormalize these three metrics. If the learnable frequency parameter defined for each dimension of the vector is... The process of mapping the normalized index to a rotation angle using the learnable frequency parameter can be represented as follows:

[0081]

[0082] In the formula, The rotation angle reflects the dynamic characteristics of index d at position i; This is the normalized result of the indicator.

[0083] In this embodiment, the rotation angle is designed to dynamically adjust the rotation angle weight using a gating network under a preset index threshold. For example, a learnable frequency parameter can be used. =0.35 The oil temperature rotation angle θ = 1.7 × 0.35 ≈ 0.6 radians. When the gated network detects that the oil temperature change rate exceeds the threshold of 8℃ / h (the preset oil temperature threshold is 2℃ / h), the oil temperature weight α is increased from 0.2 to 0.6, thereby enhancing the large model's attention to this oil temperature index.

[0084] Furthermore, Trans-RoPE can dynamically assign weights to the location information of each indicator at different points in time. For example: "In the lead-up to a sudden failure, the weight of oil temperature increases and the weight of load decreases."

[0085] Next, a rotation matrix is ​​generated by weighting the rotation angle. During attention calculation, this rotation matrix is ​​used to encode and reconstruct the query matrix and key-value matrix in the attention matrix, generating a reconstructed attention matrix to capture device coupling relationships. For example, it can be observed that in a scenario of sudden oil temperature changes, the value of the 128th dimension (oil temperature-load coupling coefficient) in the output feature jumps from 0.45 to 0.91. For more details, please refer to... Figure 2 As shown in the figure, the example illustrates a transformer experiencing a heavy gas fault under 750kV conditions. The reconstruction process of the attention mechanism in this embodiment involves performing a Hadamard product operation between each dimension of the query matrix and the key matrix and the rotation matrix. This operation enables the large model to accurately capture coupling effects such as "when the load rate is 90%, every 1°C increase in oil temperature corresponds to a 0.8% acceleration in insulation aging," providing explicit physical sensing features for subsequent diagnosis.

[0086] S4. Fine-tune and optimize the output of the attention layer and iterate the training process to obtain a well-trained fault diagnosis model.

[0087] In this step, the deep features output by the attention layer will enter the hybrid fine-tuning decoding stage. Specifically, the MoE-DoRA hybrid fine-tuning framework, i.e., the expert-fine-tuning hybrid framework, is used in the decoding layer of the large model.

[0088] In the MoE (Expert Routing Network) part, the probability of each feature vector in the reconstructed attention matrix of the attention output is predicted by deploying an expert routing network. Feature vectors with probability values ​​exceeding a preset condition are selected as activation experts. Specifically, the routing network M generates expert selection probabilities based on the input feature vector x and performs weighted aggregation through a gating function.

[0089] In this process, for example, if the probability prediction result for a discharge fault feature is 1.1, exceeding the preset threshold of 1.0, then the routing network M will activate the discharge fault expert E3 with an activation weight of 0.73. In general, if the large model's probability prediction results for various fault features are: "Busset type fault: 0.83, Cooling system abnormality: 0.12, Protection device fault: 0.05", and the predicted probability of a certain type of fault exceeds the threshold (e.g., 0.70), a fault determination is triggered, and the corresponding maintenance strategy is generated by the currently activated expert sub-network. For details, please refer to... Figure 3 As shown on the far right, it can be seen that the routing component FFNN selects a specific activation expert FFNN1 by outputting probabilities.

[0090] Meanwhile, in this decoding layer, a DoRA fine-tuning mechanism is introduced to work in parallel with the expert routing network. In this embodiment, each expert network corresponds to an independent DoRA low-rank adapter, which is responsible for decomposing the original weights of the activated experts into magnitude components and direction components. The DoRA fine-tuning mechanism performs low-rank updates on the direction components and merges the updated direction components with the magnitude components of the original weights to optimize training.

[0091] For example, in combination Figure 3 The middle block diagram and Figure 4 As shown, the DoRA low-rank adapter decomposes the weight matrix W of the input activation expert to obtain the magnitude components of the weights: ||W||c∈R 1×n And the directional components of the weights: V∈R m×n .

[0092] During the low-rank update process, the directional component V is adjusted using low-rank matrices A and B to generate the updated ΔV = A × B. The adapted direction is V + ΔV. The adapted weights are then merged with the magnitude component to obtain the new merged weight W', as shown below:

[0093]

[0094] During collaborative parallel training, the routing network and the adapter jointly optimize the parameter loss, where the loss function L includes the task loss L. task With expert load balancing loss L balance , means as follows:

[0095]

[0096] The second term in the formula is L. balance p i Let λ be the historical call frequency of expert routing network i, λ be the balancing coefficient, and N be the number of expert routing networks.

[0097] After fully implementing the entity extraction, attention reconstruction, and parameter fine-tuning processes described above, iteratively train the large model until the error meets the requirements. Combined with... Figure 3 As shown on the far left, in this embodiment, the output layer of the large model is set as a combination of a normalization layer and a fully connected layer. During iterative training, a greedy algorithm is preferably integrated into the output layer of the large model to select the result corresponding to the highest prediction probability from the output probability distribution as the target output result. Specifically, the large model converts the output result into a probability distribution through a function, and then uses a greedy algorithm to select the word with the highest probability from the probability distribution as the final prediction result.

[0098] For example, in this embodiment, the preferred normalization layer is RMSnorm, while the fully connected layer introduces a gating mechanism by adding an activation function SwiGLU, allowing the large model to select different activation modes in different input regions. Finally, a linear layer extracts features and outputs the prediction result. It is worth noting that after each round of training iteration, validation and evaluation are required. The results are fed back to the CGAN rare fault generation stage and the parameter fine-tuning stage to correct and optimize the large model parameters, thereby improving the model's output accuracy.

[0099] Specifically, this invention will provide a concrete example of large-scale model training to elaborate on the above steps:

[0100] 1. In this embodiment, the number of characters acquired is 1.4 × The transformer operation procedure data yielded 17,526 question-and-answer pairs; transformer test specification data with 9.9 × 10⁶ characters was selected to construct 13,148 question-and-answer pairs; transformer fault cases and defect records with 1.3 × 10⁷ characters were selected to construct 15,974 question-and-answer pairs. These question-and-answer pairs were randomly split into three sets: 70% for training, 15% for validation, and 15% for testing. The training set was then cleaned and preprocessed using a generative adversarial network to convert it into an input sequence index (QA question-and-answer pair samples) with a maximum sampling size of 100,000 and a training dataset length truncated to 1024.

[0101] 2. The input sequence index undergoes data cleaning and preprocessing. A dynamic knowledge distillation mechanism based on cosine annealing is employed to adjust the distillation loss weights. A domain knowledge extractor based on a graph neural network is constructed to automatically extract the fault ontology from QA question-answering pairs. The output sequence is processed into integer identifiers of different lengths through an embedding layer. The integer identifiers are then transformed into a vector sequence using an embedding matrix as the initial input to the large model. In this example, the Qwen-7B-Chat open-source large model is selected for improvement and training.

[0102] 3. The vector sequence constructed in step 2 is trained using multi-level optimized Trans-RoPE rotation encoding. Rotation transformation is introduced into the query and key vectors to capture the position information during attention calculation. The parameters are jointly optimized through backpropagation, enabling the large model to adaptively adjust the importance of different dimensions and capture complex relationships such as the coupling effect of oil temperature and load and the cumulative effect of running time. The relative positional relationship between elements in the sequence is injected into the vector input large model decoding layer.

[0103] 4. The decoding layer adds a low-rank adapter layer with a DoRA rank of 8 between the fully connected layer and the attention layer, freezes all parameters of the pre-trained model, and sets the learning rate to 5e. -5 The computational model is fp16, and the parameters of the adapter layer are added during training. Fault information is extracted from unstructured text data and continuously updated. The MoE-DoRA framework achieves fine-grained adaptation of fault scenarios through dynamic routing mechanism to form knowledge-driven learning. The gradient accumulation is 16.

[0104] 5. Repeat steps 1-4. After completing one round of training, evaluate the model's performance on the validation set to prevent overfitting and determine whether to stop training. After 10 rounds of training, continuously update the model's weights and biases.

[0105] 6. The trained large model was tested on the test set. Specifically, for ease of evaluation, three evaluation metrics—precision, recall, and F1 score—were introduced to evaluate the model's results in diagnosing transformer equipment fault types and fault state levels. BLEU-4 (a machine translation evaluation metric) and ROUGE-1, ROUGE-2, and ROUGE-L (text summarization evaluation metrics based on recall) were also introduced to evaluate the generated results of transformer equipment fault repair strategies.

[0106] To verify whether the trained large model, after MoE-DoRA fine-tuning, forms more knowledge links and effectively improves vertical domain-specific question answering, this example selects 70 different transformer equipment fault scenarios and uses a scoring system to evaluate the transformer fault repair strategy generation capability of the trained large model. In this example, the improved and trained large model is named TransQwen.

[0107] See test results Figure 5 As shown, the TransQwen model can generate precise maintenance strategies tailored to transformer equipment faults, reducing ambiguity in operation. Furthermore, the generated maintenance strategies are comprehensive, detailed, safe, and systematic. Figure 5 As can be seen, the BLEU-4, ROUGE-1, ROUGE-2, and ROUGE-L indices evaluated by the TransQwen model are 79.86, 86.18, 81.92, and 84.58, respectively; the total score (SUM) and average score rate (AVG) of the assigned scores are 89.94 and 88.65, respectively. It is evident that the large-scale prediction model trained in this example—TransQwen—significantly improves performance compared to other models (the traditional Qwen-7B-Chat and a pre-trained general Qwen model), and can more accurately diagnose transformer equipment faults.

[0108] S5. Real-time acquisition of transformer equipment operation data and input into the fault diagnosis model to generate fault diagnosis results and maintenance strategies.

[0109] After the trained large model is deployed, multi-source data of transformer equipment is collected in real time, such as structured monitoring data: oil chromatography gas concentration, infrared thermogram, load curve, etc., as well as unstructured text records: inspection description, abnormal sound, etc. These data are standardized to generate a unified vector representation, which is then input into the trained fault diagnosis model for diagnosis. After various layers of operation, the model finally outputs the fault diagnosis results and the corresponding maintenance strategy.

[0110] For example, the fault diagnosis includes: Fault type: At the high-voltage side bushing of phase A of the No. 1 main transformer in a 110kV substation, insulating oil flows downwards around the porcelain bushing along the upper edge. Component attributes: The No. 1 main transformer is an oil-immersed self-cooled transformer; the 110kV side bushing is a high-voltage oil-paper capacitor type bushing with a capacitance of 308pF, a creepage distance of 3906mm, and a weight of approximately 80kg.

[0111] The corresponding maintenance strategies can be: 1. Power outage: Immediately notify relevant departments and obtain a power outage permit to ensure the transformer is completely de-energized and the safety of maintenance personnel is guaranteed. 2. On-site isolation: Identify the leak location and isolate the faulty area promptly. 3. Oil draining: Drain the leaked oil and replace the insulating oil. 4. Bushing removal: Check for loose connections between the A-phase bushing oil conservator and the porcelain bushing flange, and remove the bushing to confirm whether the connection between the cable-carrying aluminum pipe and the lower conical porcelain bushing is broken. 5. High-voltage bushing replacement: Ensure all components are tightly connected and use extended high-voltage side cable-carrying installation. 6. Equipment testing: Conduct infrared thermography retesting to ensure all components are operating normally without overheating or leakage. Perform comprehensive infrared thermography checks on transformers of the same model, and simultaneously test the dielectric strength, moisture content, and other indicators of the insulating oil to ensure the oil quality is qualified. Conduct an energization test. 7. Regular inspection and maintenance: Conduct infrared testing according to the "Condition-Based Maintenance and Testing Procedures for Transmission and Transformation Equipment," especially increasing the frequency of testing during high-temperature and high-load operation, and paying attention to checking the bushing oil level. Strengthen the monitoring and tracking of oil chromatography and high-voltage tests on equipment with recurring defects, and promptly overhaul or replace equipment with serious problems. 8. Training and Improvement: Enhance the professional training of maintenance personnel, update maintenance manuals and operating procedures, and ensure the standardization and effectiveness of maintenance work. Simultaneously, strengthen the whole-process management of transformer selection, ordering, acceptance, and commissioning to ensure the use of reliable products.

[0112] It should be understood that, in practical applications, the execution status of maintenance strategies can be further fed back to the monitoring terminal for staff to understand, so as to update the equipment fault graph based on the feedback results and iteratively optimize the fault diagnosis model. For example, if the diagnosed fault type is cooling oil circuit blockage, the operation performed is to start the backup oil pump and clean it. If the feedback result is that the flow rate recovers to 95% after cleaning, the knowledge graph can be updated accordingly to improve the handling speed when a similar fault occurs again.

[0113] In summary, this invention constructs a fault diagnosis model based on a large language model. It integrates historical transformer fault cases and uses CGAN adversarial generation technology to expand rare samples and construct an initial QA sample set. A dynamic knowledge distillation mechanism combined with a graph neural network is used to extract equipment fault ontology from unstructured text and generate a knowledge graph. A Trans-RoPE rotational position encoding layer is introduced after the model embedding layer to accurately capture the coupling relationship of equipment states. A MoE-DoRA hybrid fine-tuning framework is used to achieve efficient parameter updates. Finally, after normalization and a fully connected layer, fault diagnosis results and customized maintenance strategies are output, forming a complete intelligent power equipment operation and maintenance system from diagnosis to strategy, improving the efficiency and accuracy of power equipment fault detection.

[0114] One embodiment of the present invention provides a transformer equipment fault diagnosis system. For details, please refer to [link / reference]. Figure 6 , Figure 6 The diagram shown illustrates the structure of a transformer equipment fault diagnosis system according to one embodiment of the present invention, including:

[0115] The training sample construction module M1 is used to integrate and analyze various operational fault cases of transformer equipment in the target power system and construct an initial question-answer pair sample set.

[0116] The text entity extraction module M2 is used to input the initial question-and-answer pair sample set into a preset large model for training. During the training process, a preset dynamic knowledge extractor is introduced to extract text entities from the initial question-and-answer pair sample set and generate a device fault map. The dynamic knowledge extractor is composed of a distillation mechanism and a graph neural network.

[0117] The attention reconstruction module M3 is used to transform the device fault map vector and input it into the attention layer reconstructed by rotation position encoding to capture device coupling relationships;

[0118] The iterative training module M4 is used to fine-tune and optimize the output of the attention layer and iteratively train the model to obtain a well-trained fault diagnosis model.

[0119] The fault diagnosis module M5 is used to collect the operating data of the transformer equipment in real time and input it into the fault diagnosis model to generate fault diagnosis results and maintenance strategies.

[0120] like Figure 7 As shown, this embodiment of the invention also provides a computer device. Figure 7 This is a structural block diagram of a preferred embodiment of a computer device provided by the present invention. The computer device includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the method described above.

[0121] Preferably, the computer program can be divided into one or more modules / units (such as computer program 1, computer program 2, ...), and the one or more modules / units are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program in the computer device.

[0122] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or the processor can be any conventional processor. The processor is the control center of the terminal device, connecting various parts of the terminal device through various interfaces and lines.

[0123] The memory mainly includes a program storage area and a data storage area. The program storage area can store the operating system, applications required for at least one function, etc., while the data storage area can store related data, etc. Furthermore, the memory can be a high-speed random access memory, or a non-volatile memory, such as a plug-in hard drive, a SmartMedia Card (SMC), a Secure Digital (SD) card, and a Flash Card, or other volatile solid-state storage devices.

[0124] It should be noted that the aforementioned terminal devices may include, but are not limited to, processors and memory, as will be understood by those skilled in the art. Figure 7 The structural block diagram is merely an example of a terminal device and does not constitute a limitation on the terminal device. It may include more or fewer components than shown, or combine certain components, or use different components. Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium may be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0125] Accordingly, embodiments of the present invention provide a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the steps in the method of the above embodiments, for example... Figure 1 Steps S1 to S5 as described above.

[0126] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for diagnosing transformer equipment faults, characterized in that, include: Integrate and analyze various operational failure cases of transformer equipment in the target power system to construct an initial question-and-answer pair sample set; The initial question-and-answer pair sample set is input into a preset large model for training. During the training process, a preset dynamic knowledge extractor is introduced to extract text entities from the initial question-and-answer pair sample set to generate a device fault map. The dynamic knowledge extractor is composed of a distillation mechanism and a graph neural network. The device fault map vector is transformed and then input into the attention layer reconstructed by rotational position encoding to capture device coupling relationships; The output of the attention layer is fine-tuned and optimized by parameters, and the training process is iterated to obtain a well-trained fault diagnosis model. Real-time acquisition of transformer equipment operation data and input into the fault diagnosis model to generate fault diagnosis results and maintenance strategies; The process of generating the equipment fault map includes: After the initial question and answer pairs are sequence-transformed, they are input into a preset large model. A distillation mechanism is introduced to dynamically optimize the weights of transformer equipment operation indicators and generate fault knowledge vectors. The driving graph neural network extracts text entities from the fault knowledge vector to construct node features and edge features, thereby generating the equipment fault map; The introduced distillation mechanism dynamically optimizes the weights of transformer equipment operation indicators and generates a fault knowledge vector, including: The teacher model trained using this distillation mechanism is processed in parallel, and then the weights of the teacher model are dynamically allocated according to the text keywords through a gating network. Finally, the weighted fusion is used to form a fault knowledge vector.

2. The transformer equipment fault diagnosis method as described in claim 1, characterized in that, The integrated analysis of various operational fault cases of transformer equipment in the target power system constructs an initial question-and-answer pair sample set, including: Collect transformer operation procedures, transformer test specifications data and historical transformer fault data, and annotate them with question-and-answer pairs to integrate them into operational fault cases; Generative adversarial networks are used to augment the runtime failure cases to construct the initial question-answer pair sample set.

3. The transformer equipment fault diagnosis method as described in claim 1, characterized in that, The step of converting the device fault map vector and inputting it into the attention layer reconstructed by rotational position encoding to capture device coupling relationships includes: The vector representation of the device fault map is input into the attention layer and feature normalization is performed. The normalization result is mapped to a rotation angle using a learnable frequency parameter; wherein the rotation angle is designed to dynamically correct the rotation angle weight using a gating network under a preset index threshold condition. The rotation matrix is ​​generated by weighted analysis of the rotation angles. During attention calculation, the rotation matrix is ​​used to perform encoding reconstruction on the query matrix and key matrix to generate a reconstructed attention matrix to capture the device coupling relationship.

4. The transformer equipment fault diagnosis method as described in claim 1, characterized in that, The process of fine-tuning and optimizing the parameters includes: An expert routing network is deployed to predict the probability of each feature vector in the reconstructed attention matrix, and feature vectors with probability values ​​exceeding preset conditions are used as activated experts. The DoRA fine-tuning mechanism is introduced to decompose the original weights of the activated expert into magnitude and direction components; The direction component is updated in low rank, and the updated direction component is merged with the amplitude component to optimize the training.

5. The transformer equipment fault diagnosis method as described in claim 1, characterized in that, The process of fine-tuning and optimizing the output of the attention layer and iteratively training also includes: During iterative training, a greedy algorithm is integrated into the output layer of a pre-defined large model to select the result with the highest predicted probability from the output probability distribution as the target output result.

6. The transformer equipment fault diagnosis method as described in claim 1, characterized in that, After the real-time acquisition of transformer equipment operating data and input into the fault diagnosis model to generate fault diagnosis results and maintenance strategies, the process also includes: The execution status of the maintenance strategy is fed back to the monitoring terminal; The equipment fault map is updated based on the feedback results, and the fault diagnosis model is iteratively optimized.

7. A transformer equipment fault diagnosis system, characterized in that, include: The training sample construction module is used to integrate and analyze various operational failure cases of transformer equipment in the target power system and construct an initial question-answer pair sample set; The text entity extraction module is used to input the initial question-and-answer pair sample set into a preset large model for training. During the training process, a preset dynamic knowledge extractor is introduced to extract text entities from the initial question-and-answer pair sample set and generate a device fault map. The dynamic knowledge extractor is composed of a distillation mechanism and a graph neural network. The attention reconstruction module is used to transform the device fault map vector and input it into the attention layer reconstructed by rotation position encoding to capture device coupling relationships; The iterative training module is used to fine-tune and optimize the output of the attention layer and iteratively train the model to obtain a well-trained fault diagnosis model. The fault diagnosis module is used to collect the operating data of the transformer equipment in real time and input it into the fault diagnosis model to generate fault diagnosis results and maintenance strategies. The process of generating the equipment fault map includes: After the initial question and answer pairs are sequence-transformed, they are input into a preset large model. A distillation mechanism is introduced to dynamically optimize the weights of transformer equipment operation indicators and generate fault knowledge vectors. The driving graph neural network extracts text entities from the fault knowledge vector to construct node features and edge features, thereby generating the equipment fault map; The introduced distillation mechanism dynamically optimizes the weights of transformer equipment operation indicators and generates a fault knowledge vector, including: The teacher model trained using this distillation mechanism is processed in parallel, and then the weights of the teacher model are dynamically allocated according to the text keywords through a gating network. Finally, the weighted fusion is used to form a fault knowledge vector.

8. A computer device, characterized in that, The device includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the transformer equipment fault diagnosis method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the device containing the computer-readable storage medium executes the computer program, it implements the transformer equipment fault diagnosis method as described in any one of claims 1 to 6.

Images