A Transformer Condition Comprehensive Analysis System and Method

Abstract

This invention discloses a comprehensive transformer condition analysis system, comprising: a data processing module that preprocesses transformer multimodal data to form training samples and fine-tuning samples; a model building module that performs unsupervised training on an improved Transformer model to achieve the first transformer health state classification, generating a Transformer pre-trained model and embedding a transformer fault identification model based on k-means and SVM algorithms; a human feedback learning technology module that scores the Transformer pre-trained model based on the fine-tuning samples and updates the learning strategy parameters according to the scoring mechanism to derive a reward model; a system that optimizes the model using a proximal strategy based on the fine-tuning samples to generate answers and assigns scores, and performs reinforcement learning on the reward model to update the strategy parameters of the strategy optimization model; and a second transformer health state classification based on the reward model and the strategy optimization model to derive the transformer condition. This invention improves the accuracy and efficiency of fault identification, making fault diagnosis more efficient and accurate.

Description

Technical Field

[0001] This invention relates to the field of power grid equipment condition analysis technology, specifically to a transformer condition comprehensive analysis system and method. Background Technology

[0002] The operating status of power equipment is directly related to the safety of the power system. Among them, transformers are the most critical equipment in the power system. Once a failure occurs, it may affect the safety of society and public utilities and the stability of the economy.

[0003] In order to identify transformer faults, experts have studied various condition analysis and fault identification technologies. Traditional methods include temperature detection, sound detection, and oil sample analysis. Modern methods widely use dissolved gas analysis (DGA) technology. Most condition analysis methods often require a lot of time and manpower, and due to information asymmetry, they may lead to poor fault identification results.

[0004] Currently, there are already cases of combining power equipment with AI in fields such as fault diagnosis, image recognition, and big data analysis. However, the GPT4 model can complete various tasks in different application scenarios, possessing a higher level of intelligence, a wider range of adaptability, and stronger generative capabilities than traditional AI. This effectively improves the efficiency of information management, condition assessment, and design and manufacturing for a vast number of diverse power equipment. Therefore, inventing a comprehensive transformer condition analysis system that utilizes the GPT4 model to improve the efficiency of information management, condition assessment, and design and manufacturing for a vast number of diverse power equipment, while simultaneously improving fault identification accuracy and ensuring the safe and stable operation of transformers, is an urgent technical problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide a transformer condition comprehensive analysis system and method. This invention can improve the efficiency of information management, condition assessment and design and manufacturing of a large number of various types of power equipment, while improving the accuracy of fault identification and ensuring the safe and stable operation of transformers.

[0006] To achieve this objective, the present invention provides a transformer condition comprehensive analysis system, which includes: The data processing module performs data preprocessing based on the collected transformer multimodal data to form training samples and fine-tuning samples; The model building module performs unsupervised training on the training samples based on the improved Transformer model to achieve the first classification of transformer health status, thereby generating a Transformer pre-trained model. The transformer fault identification model based on the k-means algorithm and SVM algorithm is embedded in the Transformer pre-trained model to obtain a Transformer pre-trained model with a transformer fault identification model. The human feedback learning technology module is used to assign scores to the Transformer pre-trained model with transformer fault identification model based on fine-tuned sample training. It generates corresponding answer 1 for each transformer health state classification, sorts the answer 1 according to the closeness of each answer 1 to the understanding of the transformer health state to obtain the scoring mechanism of the reward model, and updates the reinforcement learning policy parameters in the Transformer pre-trained model with transformer fault identification model according to the scoring mechanism of the reward model to obtain the trained reward model. The human feedback learning technology module is also used to randomly select sub-fine-tuning samples from the fine-tuning samples of various health states of the transformer, use the sub-fine-tuning samples to generate answer two using the proximal policy optimization model, assign a score to answer two according to the scoring mechanism of the reward model, and perform reinforcement learning on the trained reward model according to the score of answer two and the set score gradient to update the policy parameters of the policy optimization model, thus obtaining the updated policy optimization model. The analysis module is used to perform a second classification of the health status of each transformer based on the trained reward model and the updated policy optimization model, and to derive the transformer status corresponding to different categories based on the results of the second classification.

[0007] Preferably, it also includes a model optimization module that optimizes the encoder, decoder, sharing, layer normalization, and position encoding parameters of the Transformer pre-trained model generated by unsupervised training using fine-tuning samples, to obtain an optimized Transformer pre-trained model.

[0008] Preferably, after dividing the training samples into equivalence classes, validation data is selected to form validation samples, and the accuracy of the optimized Transformer pre-trained model with transformer fault identification model is tested based on the validation samples.

[0009] Preferably, the collected transformer multimodal data is obtained from offline historical data of the transformer and includes structural, text, image, and video data of the transformer.

[0010] Preferably, the data preprocessing methods include: for structured data, missing values ​​and outliers are first processed, categorical variables are converted into standardized numerical variables, and the data is normalized; for text data, text recognition, segmentation, processing, and reconstruction are performed based on GPT4 natural language processing technology to form corpus samples; for image and video data, parsing is performed based on Transformer model encoding technology.

[0011] Preferably, the fine-tuning samples include high-quality labeled information such as transformer status analysis reports and operation and maintenance inspection reports annotated by power professionals, which are used to fine-tune the reinforcement learning model based on human feedback; the training samples are unlabeled raw text data of the equipment, including equipment ledgers, maintenance records, fault records, operation and maintenance records, technical standards, and technical specifications.

[0012] The beneficial effects of this invention are: This invention provides a comprehensive transformer condition analysis system. Based on the GPT4 multimodal algorithm, it preprocesses collected transformer multimodal data. The GPT4 algorithm can process and integrate different types of data, forming standardized training samples and fine-tuning samples suitable for subsequent model training, improving the effectiveness and accuracy of subsequent model training. Unsupervised training of the training samples using an improved Transformer model allows the system to learn the inherent structure and features of the data, improving the model's generalization ability. Embedding a transformer fault identification model based on k-means and SVM algorithms into the Transformer pre-trained model further enhances the accuracy and efficiency of fault identification. Reinforcement learning techniques based on human feedback optimize the Transformer pre-trained model using fine-tuning samples, enabling the model to adjust parameters based on human evaluators' scores, improving performance and updating model scores. This guides the model to learn how to more accurately identify faults, enhancing its ability to interact with humans and improving fault identification accuracy. Through the combined effect of these improvements, this system enhances the accuracy and efficiency of fault identification, making fault diagnosis more efficient and accurate. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the structure of the present invention; Figure 2 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 a part of the embodiments of the present invention, not all of them. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0015] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1 A transformer condition comprehensive analysis system, such as Figure 1 As shown, it includes: The data processing module performs data preprocessing based on the collected transformer multimodal data to form training samples and fine-tuning samples; The model building module performs unsupervised training on the training samples based on the improved Transformer model to achieve the first classification of transformer health status, thereby generating a Transformer pre-trained model. The transformer fault identification model based on the k-means algorithm and SVM algorithm is embedded in the Transformer pre-trained model to obtain a Transformer pre-trained model with a transformer fault identification model. The human feedback learning technology module is used to assign scores to the Transformer pre-trained model with transformer fault identification model based on fine-tuned sample training. It generates corresponding answer 1 for each transformer health state classification, sorts the answer 1 according to the closeness of each answer 1 to the understanding of the transformer health state to obtain the scoring mechanism of the reward model, and updates the reinforcement learning policy parameters in the Transformer pre-trained model with transformer fault identification model according to the scoring mechanism of the reward model to obtain the trained reward model. The human feedback learning technology module is also used to randomly select sub-fine-tuning samples from the fine-tuning samples of various health states of the transformer, use the sub-fine-tuning samples to generate answer two using the proximal policy optimization model, assign a score to answer two according to the scoring mechanism of the reward model, and perform reinforcement learning on the trained reward model according to the score of answer two and the set score gradient to update the policy parameters of the policy optimization model, thus obtaining the updated policy optimization model. The analysis module is used to perform a second classification of the health status of each transformer based on the trained reward model and the updated policy optimization model, and to derive the transformer status corresponding to different categories based on the results of the second classification.

[0016] In the above technical solution, external knowledge and evaluation standards are introduced through human feedback to improve the decision-making quality of the model. The scoring mechanism that rewards the model can guide the direction of model optimization, making it more in line with the needs of practical applications.

[0017] In the above technical solution, unsupervised pre-training is used to master the expression patterns, contextual logic, and knowledge reserves of training samples, enabling them to understand multimodal information of power equipment, perform state reasoning, generate equipment information, and learn from historical data.

[0018] In the above technical solutions, the embedded transformer fault identification model can employ methods such as dissolved gas analysis (DGA) and the three-ratio method.

[0019] In the above technical solution, the sample data is offline data. By accessing the data interface of the data platform, the relevant historical data of the transformer can be exported for research, including data filtering, discretization and dimensionality reduction.

[0020] The above technical solution also includes a model optimization module that optimizes the encoder, decoder, shared layer normalization, and positional encoding parameters of the Transformer pre-trained model generated by unsupervised training using fine-tuning samples, thereby obtaining an optimized Transformer pre-trained model.

[0021] In the above technical solution, after dividing the training samples into equivalence classes, validation data is selected to form validation samples. The accuracy of the optimized Transformer pre-trained model with transformer fault identification model is tested based on the validation samples.

[0022] In the above technical solution, GPT4 natural language processing technology is used to perform text recognition, segmentation, processing and reconstruction on transformer multimodal data to form corpus samples.

[0023] In the above technical solution, the collected transformer multimodal data is obtained based on the transformer's offline historical data and includes transformer structural data, text data, image data, and video data.

[0024] In the above technical solutions, structural data includes model data, operational data, online monitoring data, maintenance records, environmental factors, etc.; text data includes experimental reports, fault cases, technical standards, technical specifications, etc.; image data includes fault waveforms, inspection pictures, etc.; and video data includes remote monitoring videos, etc.

[0025] In the above technical solutions, for structured data, missing values ​​and outliers are first processed, categorical variables are converted into standardized numerical variables, and the data is normalized to eliminate the influence of different dimensions between features; for text data, text recognition, segmentation, processing, and reconstruction are performed based on GPT4 natural language processing technology to form corpus samples; for image and video data, parsing is performed based on the encoding technology of the Transformer model.

[0026] In the above technical solution, the data preprocessing methods are as follows: for structured data, missing values ​​and outliers are first processed, categorical variables are converted into standardized numerical variables, and the data is normalized; for text data, text recognition, segmentation, processing, and reconstruction are performed based on GPT4 natural language processing technology to form corpus samples; for image and video data, parsing is performed based on Transformer model encoding technology.

[0027] In the above technical solution, by preprocessing the multimodal data of the transformer, the quality and consistency of the input data are ensured, laying a solid foundation for subsequent model training and helping to improve the training efficiency and accuracy of the model.

[0028] In the above technical solution, the improved Transformer model mainly optimizes the mask self-attention mechanism. The working principle of the self-attention mechanism involves labeling text segments and converting these labeled text segments into vectors representing the importance of the label in the input sequence. The specific steps are as follows: For each token in the input sequence, create a query, key, and value vector. Calculate the similarity between the query vector of the token and the key vectors of other tokens by taking the dot product of the two vectors. Input the calculation result into a softmax function with a mask to calculate the normalized weight. Multiply the calculated weight by the value vector of each token to produce a final vector that represents the importance of the tokens in the sequence. The specific formula for calculating the similarity between the query vector of a tag and the key vector of other tags is as follows: Where q is the query vector, It is the transpose of the key vector, and v is the value vector. It is the dimension of the vector. It's a mask.

[0029] In the above technical solution, the query, key, and value vectors are trained by the model to obtain vector functions similar to querying, responding, and storing values. In the attention mechanism, the query is used to evaluate the relevance between other words; the key vector is a label or description of all words in the text, similar to what we use to match when searching for related words. In the attention mechanism, we use the relationship between the query and the key to determine the relevance between different words; the value vector is the actual word representation, usually learned through a neural network. When we use the query and key to evaluate the relevance between different words, we use these value vectors to calculate the final representation of the current word. The value vectors are weighted and combined to represent the meaning or importance of the current word.

[0030] In the above technical solution, the scaled dot product result The input is fed into the softmax function, which calculates the attention weight of each key relative to the query. The softmax function converts the original score into a probability distribution, so that the sum of the attention weights of all keys is 1.

[0031] In the above technical solution, in the power text classification task, some inputs need to be ignored or specially processed because the input lengths may be inconsistent. In order to maintain the uniformity of the inputs, padding operations are required to make all inputs the same length. By importing the masked_softmax function and using the mask tensor, the weights of these invalid or special positions can be set to negative infinity, so that the output of these positions is 0 when performing the softmax operation.

[0032] In the above technical solution, an improved Transformer model is used for unsupervised training to achieve a preliminary classification of transformer health status. Unsupervised learning can discover the inherent structure of data without explicit labels. By embedding k-means and SVM algorithms and combining the advantages of multiple algorithms, the classification accuracy and robustness of the model are improved, and the accuracy of fault identification can be further enhanced.

[0033] In the above technical solution, the specific method of embedding the transformer fault identification model based on the k-means algorithm and the SVM algorithm into the Transformer pre-trained model is as follows: The k-means algorithm is used to divide the preprocessed online monitoring and operational data into k distinct clusters based on the transformer's state. Data points are then iteratively allocated to each cluster, ensuring that data points within the same cluster have identical coordinates, while data points in different clusters have different coordinates. The goal of the k-means algorithm is to minimize the total distance between data points within a cluster and the cluster center point. The specific calculation formula is as follows: in, It is the j-th cluster. It is a cluster Data points in It is a cluster The center point; A classification model is constructed using the SVM algorithm to associate transformer gas analysis results with known fault types. A kernel function is used to map the processed training sample data to a high-dimensional feature space. The specific formula for constructing an optimal hyperplane is as follows: in, For sample data, Let be the normal vector of the hyperplane. For a high-dimensional feature space, It is the displacement; The core idea of ​​SVM is to find a hyperplane that maximizes the margin between different classes. Then, by integrating multiple constraints into a single optimization problem, the complexity of the problem is simplified. The maximum margin classifier and constraints for nonlinear data are as follows: in, As a penalty factor, Less than The training error, For insensitive loss functions, greater than The training error, Let n be the number of sample data and i be the i-th training data set. For input sample points, There are two categories; Introducing Lagrange multipliers, the formula for calculating the optimal hyperplane function is: Where SV represents the support vectors. For kernel function, Lagrange coefficient ( ), It is the displacement; A radial basis function kernel is used as the kernel function to process transformer fault information. By transforming this data into a high-dimensional space, different data points can be distinguished to facilitate more effective fault analysis and prediction. The specific formula for calculating the constraint equation of the optimal classification hyperplane is as follows: in, These are kernel function parameters.

[0034] In the above technical solution, the operating data and online monitoring data are: voltage (U), current (I), hydrogen (H2), carbon monoxide (CO), methane (CH4), ethylene (C2H4), acetylene (C2H2), ethane (C2H6), and the present invention also includes data such as water (H2O) and carbon dioxide (CO2).

[0035] In the above technical solution, the output parameters are transformer status type information, which are divided into ten groups, namely: partial discharge, low-energy discharge (spark discharge), high-energy discharge (arc discharge), thermal fault T>300℃, thermal fault 700℃>T>300℃, thermal fault T>700℃, insulation aging, insulation deterioration caused by other factors, oil contamination, water leakage into oil, and normal.

[0036] In the above technical solution, the proximal policy algorithm is a gradient-based algorithm for policy optimization, aiming to learn a policy that maximizes the cumulative reward based on experience during training, while limiting the number of policy updates. Its total loss function consists of three terms: a CLIP term, a value function term, and an entropy reward term, as shown in the following formula: in, For strategy parameters, The clip() function limits the variation of the probability ratio. This is the expected estimate. This is the estimated value of the dominance function. It is a positive number, such as 0.1 or 0.2.

[0037] In the above technical solutions, the performance of the model is further improved through strategy optimization, random selection of subsamples can increase the generalization ability of the model, and reinforcement learning updates ensure that the model can be continuously improved.

[0038] In the above technical solution, the fine-tuning samples include high-quality labeled information such as transformer status analysis reports and operation and maintenance inspection reports annotated by power professionals, which are used to fine-tune the reinforcement learning model based on human feedback.

[0039] In the above technical solution, for transformer fault identification, different clusters represent transformers with similar conditions, such as similar voltage, current, and dissolved gas content in oil.

[0040] In the above technical solution, the Transformer pre-trained model is suitable for processing text, image, and video data. For numerical data such as runtime data and online monitoring, it is necessary to embed mature fault identification algorithms. Therefore, the k-means algorithm and SVM algorithm are embedded.

[0041] In the above technical solution, k-means clustering is used to divide data points into different categories, and SVM is used to build a classification model to associate different categories with specific fault types, thereby achieving accurate diagnosis of power transformer faults.

[0042] In the above technical solution, the transformer fault identification model imports the model and function package into Python and calls the embedded Transformer pre-trained model.

[0043] In the above technical solution, the fine-tuning samples include high-quality labeled information processed from transformer status analysis reports and operation and maintenance inspection reports annotated by power professionals, which are used to fine-tune and train the reinforcement learning model based on human feedback; the training samples are unlabeled raw text data of equipment, including equipment ledgers, maintenance records, fault records, operation and maintenance records, technical standards and technical specifications.

[0044] In the above technical solution, the labeled power-related data samples can further improve the pre-trained model, making the model more suitable for the needs of power users.

[0045] When applied in the field of power equipment, the above technical solution can combine the model trained with a large amount of power equipment data with the power cloud computing platform to achieve efficient collaborative processing of computing power and data.

[0046] The above technical solution reduces the professional knowledge requirements for data analysts and simplifies the operation complexity of processing multi-source heterogeneous data by realizing functions such as multi-scale condition assessment of power transformers, intelligent fault diagnosis, equipment operation reports, and operation and maintenance plan generation.

[0047] Example 2 A comprehensive transformer condition analysis method, such as Figure 2 As shown, the collected transformer multimodal data is preprocessed using the GPT4 multimodal algorithm to form training samples and fine-tuning samples. Unsupervised training is performed on the training samples based on the improved Transformer model, and a transformer fault identification model based on k-means and SVM algorithms is embedded in the Transformer pre-trained model. The improved Transformer model is then strengthened and optimized using reinforcement learning techniques based on human feedback, and the scoring policy is updated to obtain a trained reward model and an updated policy optimization model. The accuracy of the test results is verified, and a comprehensive analysis method is formed.

[0048] The specific method for comprehensive condition analysis of oil-immersed transformers based on the improved GPT4 includes the following steps: Data preprocessing is performed on the collected transformer multimodal data to form training samples and fine-tuning samples; The improved Transformer model is used to perform unsupervised training on the training samples to achieve the first classification of transformer health status, thereby generating a Transformer pre-trained model. A transformer fault identification model based on k-means algorithm and SVM algorithm is then embedded into the Transformer pre-trained model to obtain a Transformer pre-trained model with a transformer fault identification model. The Transformer pre-trained model with transformer fault identification model is scored based on the fine-tuned sample training. The health status of each transformer is classified to generate a corresponding answer 1. The answer 1 is sorted according to the closeness of the understanding of the transformer health status of each answer 1 to obtain the scoring mechanism of the reward model. The reinforcement learning policy parameters in the Transformer pre-trained model with transformer fault identification model are updated according to the scoring mechanism of the reward model to obtain the trained reward model. Sub-fine-tuning samples are randomly selected from the fine-tuning samples of various health states of the transformer. The sub-fine-tuning samples are used to generate answer two using the proximal policy optimization model. Answer two is scored according to the scoring mechanism of the reward model. Based on the score of answer two, the trained reward model is subjected to reinforcement learning according to the set score gradient to update the policy parameters of the policy optimization model, and the updated policy optimization model is obtained. Based on the trained reward model and the updated policy optimization model, the health status of each transformer is classified a second time, and the transformer status corresponding to different categories is obtained based on the results of the second classification.

[0049] Example 3 A computer program product includes a computer program that, when executed by a processor, implements the steps of the method described in Embodiment 2.

[0050] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0051] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0052] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0053] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0054] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A transformer condition comprehensive analysis system, characterized in that, It includes: The data processing module performs data preprocessing based on the collected transformer multimodal data to form training samples and fine-tuning samples; The data preprocessing methods include: for structured data, missing values ​​and outliers are first processed, categorical variables are converted into standardized numerical variables, and the data is normalized; for text data, text recognition, segmentation, processing, and reconstruction are performed based on GPT4 natural language processing technology to form corpus samples; for image and video data, parsing is performed based on Transformer model encoding technology. The model building module performs unsupervised training on the training samples based on the improved Transformer model to achieve the first classification of transformer health status, thereby generating a Transformer pre-trained model. The transformer fault identification model based on the k-means algorithm and SVM algorithm is embedded in the Transformer pre-trained model to obtain a Transformer pre-trained model with a transformer fault identification model. The human feedback learning technology module is used to assign scores to the Transformer pre-trained model with transformer fault identification model based on fine-tuned sample training. It generates corresponding answer one for each transformer health state classification, sorts answer one according to the closeness of each answer one's understanding of the transformer health state to obtain the scoring mechanism of the reward model, and updates the reinforcement learning policy parameters in the Transformer pre-trained model with transformer fault identification model according to the scoring mechanism of the reward model. The human feedback learning technology module is also used to randomly select sub-fine-tuning samples from the fine-tuning samples of various health states of the transformer, use the sub-fine-tuning samples to optimize the model using the proximal policy to generate answer two, assign a score to answer two according to the scoring mechanism of the reward model, and perform reinforcement learning on the trained reward model again according to the score gradient based on the score of answer two, and update the reinforcement learning policy parameters in the Transformer pre-trained model with transformer fault identification model again to obtain the updated Transformer pre-trained model with transformer fault identification model; The analysis module is used to perform a second classification of the health status of each transformer based on the updated Transformer pre-trained model with a transformer fault identification model, and to derive the transformer status corresponding to different categories based on the results of the second classification.

2. The transformer condition comprehensive analysis system according to claim 1, characterized in that: It also includes a model optimization module that optimizes the encoder, decoder, shared layer normalization, and positional encoding parameters of the Transformer pre-trained model generated by unsupervised training using fine-tuning samples, resulting in an optimized Transformer pre-trained model.

3. The transformer condition comprehensive analysis system according to claim 1, characterized in that: After dividing the training samples into equivalence classes, validation data is selected to form validation samples. The accuracy of the optimized Transformer pre-trained model with transformer fault identification model is tested based on the validation samples.

4. The transformer condition comprehensive analysis system according to claim 1, characterized in that: The collected transformer multimodal data is obtained from offline historical transformer data and includes structural, text, image, and video data of the transformer.

5. The transformer condition comprehensive analysis system according to claim 1, characterized in that: The improved Transformer model primarily optimizes the mask self-attention mechanism. The working principle of this mechanism involves labeling text segments and converting them into vectors representing the importance of each label within the input sequence. The specific steps are as follows: For each token in the input sequence, create a query, key, and value vector. Calculate the similarity between the query vector of the token and the key vectors of other tokens by taking the dot product of the two vectors. Input the calculation result into a softmax function with a mask to calculate the normalized weight. Multiply the calculated weight by the value vector of each token to produce a final vector that represents the importance of the tokens in the sequence. The specific formula for calculating the similarity between the query vector of a tag and the key vector of other tags is as follows: Where q is the query vector, It is the transpose of the key vector, and v is the value vector. It is the dimension of the vector. It's a mask.

6. The transformer condition comprehensive analysis system according to claim 1, characterized in that: The fine-tuning samples include high-quality labeled information processed from transformer status analysis reports and operation and maintenance inspection reports annotated by power professionals, which are used to fine-tune the reinforcement learning model based on human feedback; the training samples are unlabeled raw text data of equipment, including equipment ledgers, maintenance records, fault records, operation and maintenance records, technical standards and specifications.

7. A comprehensive analysis method for transformer condition, characterized in that, It includes the following steps: Based on the collected transformer multimodal data, data preprocessing was performed using the GPT4 multimodal algorithm to form training samples and fine-tuning samples. The data preprocessing methods included: for structured data, missing values ​​and outliers were handled, categorical variables were converted into standardized numerical variables, and the data was normalized; for text data, text recognition, segmentation, processing, and reconstruction were performed using GPT4 natural language processing technology to form corpus samples; and for image and video data, parsing was performed using encoding techniques based on the Transformer model. The improved Transformer model is used to perform unsupervised training on the training samples to achieve the first classification of transformer health status, thereby generating a Transformer pre-trained model. A transformer fault identification model based on k-means algorithm and SVM algorithm is then embedded into the Transformer pre-trained model to obtain a Transformer pre-trained model with a transformer fault identification model. The Transformer pre-trained model with transformer fault identification model is scored based on the fine-tuned sample training. The health status of each transformer is classified to generate a corresponding answer 1. The answer 1 is sorted according to the closeness of the understanding of the transformer health status of each answer 1 to obtain the scoring mechanism of the reward model. The reinforcement learning policy parameters in the Transformer pre-trained model with transformer fault identification model are updated according to the scoring mechanism of the reward model. Sub-fine-tuning samples are randomly selected from the fine-tuning samples of various health states of the transformer. The sub-fine-tuning samples are used to optimize the model using a near-end strategy to generate answer two. Answer two is scored according to the scoring mechanism of the reward model. Based on the score of answer two, the trained reward model is reinforced again according to the set score gradient. The reinforcement learning policy parameters in the Transformer pre-trained model with transformer fault identification model are updated again to obtain the updated Transformer pre-trained model with transformer fault identification model. The health status of each transformer is classified a second time based on the updated Transformer pre-trained model with transformer fault identification model, and the transformer status corresponding to different categories is obtained based on the results of the second classification.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method described in claim 7.

Images