Training method and device for intelligent recognition model of transformer fault based on text

Through the transformer fault intelligent identification model based on coding structure and cross-attention network, combined with text label domain screening method and generative adversarial network, the problem of diversity and uncertainty recognition in transformer fault text is solved, and the accuracy of fault type judgment is improved Sex and efficiency.

CN117493886BActive Publication Date: 2024-11-12HENAN KECHAI ELECTRIC POWER EQUIPMENT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311545744.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-11-16
Publication Date
2024-11-12
Estimated Expiration
2043-11-16

AI Technical Summary

Technical Problem

The prior art is difficult to effectively identify the diversity, confusion and uncertainty in the transformer fault text, resulting in low accuracy of fault type discrimination, and traditional methods are time-consuming and low accuracy.

Method used

Using text semantic model and cross-attention network based on coding structure, combining text label domain screening method and generative adversarial network, the model is trained through multi-label and tag imbalance adaptive loss function, and intelligent transformer fault identification model is generated, and the label chain vector is used. and mask enhance text features and improve classification accuracy.

Benefits of technology

It improves the classification accuracy of the transformer fault text, optimizes the model's performance in the case of label imbalance, enhances the ability to identify data of fewer types of labels, and reduces time cost.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117493886B_ABST
    Figure CN117493886B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of deep learning, and in particular to a training method and device for a text-based transformer fault intelligent recognition model, the method comprising inputting a multi-label transformer fault text into a text semantic model based on a coding structure to obtain a transformer fault text semantic coding vector, and inputting the fault category text semantic vector obtained through a text semantic model embedding layer into a cross-attention network for interaction to obtain a transformer fault category result. During training, a multi-label and label imbalance adaptive loss function MSLoss is used to optimize the model effect. The present invention proposes a text label domain screening method TLDF and a generative adversarial network LM-GAN based on a label domain mask to optimize the training effect of the model in the case of label imbalance. The present invention improves the model performance and robustness by optimizing the loss function and enhancing the few-sample data using the label domain feature information in the text, thereby achieving high-accuracy intelligent recognition of transformer faults.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of natural language processing and deep learning, and in particular to a training method and device for a text-based transformer fault intelligent recognition model. Background Art

[0002] Transformer failure mainly refers to abnormalities or failures in the operation of the transformer, which may cause damage to the power equipment or stop normal operation. Common transformer failures include winding short circuit, insulation aging, grounding faults, etc. These failures may be caused by equipment aging, external environmental factors or manufacturing defects. Timely detection and repair of transformer failures is essential to ensure the stable operation of the power system and can also extend the service life of the equipment.

[0003] After a transformer fails, relevant staff will record the fault phenomenon caused by the fault in text. However, due to the differences in knowledge reserves, language habits, observation angles and observation integrity of the recorders, the transformer fault text is diverse and confusing, and the fault text category is uncertain. This increases the time cost of manual review of transformer fault texts and reduces the accuracy of fault type identification. Therefore, it is necessary to use an intelligent transformer fault identification method based on deep learning and natural language processing technology.

[0004] The traditional intelligent recognition method is to have relevant experts write templates and rules to extract features from fault texts, and then use traditional machine learning algorithms, such as random forests and support vector machines, to classify faults. This method is labor-intensive and time-consuming and has a low accuracy rate. The current fault recognition method using natural language processing technology based on deep learning cannot cope well with complex situations such as unfamiliar semantics and unbalanced labels in transformer fault texts. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a training method and device for a text-based transformer fault intelligent recognition model to solve the problems that the traditional fault text recognition model has low accuracy and cannot solve the diversity, confusion and uncertainty problems unique to transformer fault text.

[0006] In a first aspect of the present invention, the present invention provides a training method for a text-based transformer fault intelligent identification model, comprising the following steps:

[0007] Obtain transformer fault text data, wherein the transformer fault text data includes transformer fault information text and corresponding multiple transformer fault category texts, each fault category text indicating a fault category label;

[0008] Input the fault category text into the embedding layer of the text semantic model based on the encoding structure, and output the fault category text semantic encoding vector through the encoding layer of the text semantic model based on the encoding structure;

[0009] Processing the fault information text into a text to be enhanced by a text label domain screening method, and forming a text pair to be enhanced with the fault information text before processing and the text to be enhanced;

[0010] Input the text pair to be enhanced into the embedding layer of the text semantic model based on the encoding structure, and output an initial enhanced text vector pair;

[0011] The initial enhanced text vector is input into the generative adversarial network based on the label domain mask to train and output the intermediate enhanced text vector;

[0012] Pass the intermediate enhanced text vector through the linear layer to obtain the target enhanced text vector;

[0013] The target enhanced text vector is directly input into the encoding layer of the text semantic model based on the encoding structure, and the fault information text is simultaneously input into the embedding layer of the text semantic model based on the encoding structure, and the fault information text semantic encoding vector is output through the encoding layer of the text semantic model based on the encoding structure;

[0014] Input the fault information text semantic encoding vector and the fault category text semantic vector into the cross-attention network, process them through the classifier, and output the transformer fault category result vector; the classifier includes an output Dropout layer, a linear layer, and an activation sigmoid layer;

[0015] The transformer fault category result vector is combined with multiple transformer fault category texts to construct a multi-label and label imbalance adaptive loss function, and the transformer fault intelligent identification model is obtained through loss training.

[0016] In an embodiment of the present invention, the text semantic model based on the coding structure is composed of a Chinese natural language pre-training model with correction as Mask, and the Chinese natural language pre-training model includes at least an embedding layer and an encoding layer, the embedding layer is mainly used to extract the embedding vector of the text, and the encoding layer is mainly used to obtain the encoding vector of the text. The generative adversarial network based on label domain mask (Generative Adversarial Network based on Label domain Mask, LM-GAN for short) is composed of a generator composed of an Encoder block and a Decoder block in transformers, and a discriminator composed of a bidirectional long short-term memory network BiLSTM, a linear layer, and an activation function.

[0017] In a second aspect of the present invention, the present invention further provides a training device for a text-based transformer fault intelligent identification model, the device comprising:

[0018] An acquisition module, used to acquire transformer fault text data, wherein the transformer fault text data includes transformer fault information text and corresponding multiple transformer fault category texts, each fault category text indicating a fault category label;

[0019] The training module is used to input the fault category text into a text semantic model based on a coding structure, and output a fault category text semantic coding vector; process the fault information text into a to-be-enhanced text through a text label domain screening method, and form a to-be-enhanced text pair with the fault information text before processing and the to-be-enhanced text; input the to-be-enhanced text pair into a text semantic model based on a coding structure, and output an initial enhanced text vector pair; input the initial enhanced text vector pair into a generative adversarial network based on a label domain mask for training, and output an intermediate enhanced text vector; pass the intermediate enhanced text vector through a linear layer to obtain a target enhanced text vector; input the fault information text into a text semantic model based on a coding structure, and output an initial enhanced text vector pair; A text semantic model of a coding structure is constructed, and the target enhanced text vector is input into the coding layer of the text semantic model based on the coding structure at the same time, and a fault information text semantic coding vector is output; the fault information text semantic coding vector and the fault category text semantic vector are input into a cross-attention network, and processed by a classifier to output a transformer fault category result vector; the classifier includes an output Dropout layer, a linear layer and an activation sigmoid layer; the transformer fault category result vector is combined with multiple transformer fault category texts to construct a multi-label and label imbalance adaptive loss function, and a transformer fault intelligent recognition model is obtained through loss training.

[0020] Beneficial effects of the present invention:

[0021] 1. The present invention uses a label chain vector composed of fault category texts to assist in decoding the encoding vector of the text semantic model based on the encoding structure, so that the classifier is separated from the monotonous label mapping numerical space, fully utilizes the label semantic representation information, and improves the classification accuracy;

[0022] 2. The present invention adopts the multi-label and label imbalance adaptive loss function MSLoss, which enables the model to adjust the weight of the loss according to the classification accuracy of the correct label when calculating the loss, and uses the cosine function to adaptively coordinate the overall loss scale during training and the sensitivity of the correct label and the wrong label to the confidence;

[0023] 3. The present invention adopts the Text Label Domain Filtering method (TLDF) and the generative adversarial network LM-GAN based on label domain mask for few-class label fault data enhancement. The label domain confidence difference value of each language block is characterized by calculating the Euclidean distance of the label confidence between the original text and the incomplete text of the language block, and the representation area of ​​the fault category in the original text is detected. After mask replacement, the label domain information in the original text is fitted at the embedded representation vector level, and the label features in the incomplete text of the language block are filled in from the high-dimensional vector level, so as to achieve the purpose of enhancing the few-class label fault data and optimize the prediction effect of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0025] Figure 1 It is an operation flow chart of a text-based transformer fault intelligent identification model according to an embodiment of the present invention;

[0026] Figure 2 A training flow chart of a text-based transformer fault intelligent identification model according to an embodiment of the present invention

[0027] Figure 3 This is an algorithm principle diagram of a label domain masked generative adversarial network LM-GAN of a text-based transformer fault intelligent identification model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0029] In some embodiments, the text-based transformer fault intelligent identification method of the present invention can be applied to a scenario where a computer device and a server coexist, wherein the computer device communicates with the server through a network. The computer device can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, and portable wearable devices. The server can be implemented as an independent server.

[0030] Figure 1is a schematic diagram of a training structure of a text-based transformer fault intelligent identification model according to an embodiment of the present invention, such as Figure 1 As shown, in an embodiment of the present invention, the transformer fault category text and the transformer fault information text are input into the transformer fault intelligent recognition model. After a series of processing, the fault category confidence can be output. The fault category confidence can reflect the fault type of the transformer fault text and accurately identify the transformer fault category. In this process, the transformer fault intelligent recognition model plays a key role. In order to obtain a higher recognition accuracy, the embodiment of the present invention will optimize and train it through a suitable training method so that the transformer fault intelligent recognition model can accurately identify the fault category confidence of the transformer fault text.

[0031] In an embodiment of the present invention, the transformer fault intelligent identification model includes a text semantic model based on a coding structure, a cross-layer attention network Cross-Attention, a Dropout layer, a Linear layer and a Sigmoid layer, and the text semantic model based on the coding structure is composed of a Chinese natural language pre-training model MacBERT with a corrected Mac as a Mask. The model reduces the gap between the pre-training and fine-tuning stages by using similar word masks. The Chinese natural language pre-training model used in this embodiment includes at least an embedding layer and an encoding layer. The embedding layer is mainly used to extract the embedding vector of the text, and the encoding layer is mainly used to obtain the encoding vector of the text.

[0032] In the embodiment of the present invention, in addition to using the transformer fault intelligent identification model to obtain the fault category confidence, Figure 1 As shown, the present invention also uses the TLDF sieving method to process the transformer fault text, and trains the generative adversarial network LM-GAN based on the label domain mask to obtain an enhanced text vector of the few-class label data, and uses the enhanced text vector to enhance the training of the transformer fault intelligent recognition model to obtain a transformer fault intelligent recognition model with a higher recognition rate; wherein the generative adversarial network LM-GAN based on the label domain mask is composed of a generator composed of an Encoderblock and a Decoder block in transformers, and a discriminator composed of a bidirectional long short-term memory network BiLSTM, a linear layer, and an activation function.

[0033] Figure 2 is a flow chart of a training method for a text-based transformer fault intelligent identification model according to an embodiment of the present invention, such as Figure 2 As shown, the following steps may be specifically included:

[0034] 101. Obtain transformer fault text data, wherein the transformer fault text data includes transformer fault information text and a plurality of corresponding transformer fault category texts, each fault category text indicating a fault category label;

[0035] In an embodiment of the present invention, the transformer fault text data may be an original transformer fault analysis report, and this information is displayed in text form. Taking the transformer fault analysis report of a 380KV substation as an example, the transformer fault analysis report contains a large number of transformer fault information sample texts and corresponding multiple transformer fault category sample texts. These transformer fault information sample texts may include basic file information such as the manufacturer, model, capacity, substation name, equipment name, manufacturer, transformer equipment model, etc. of the equipment, fault phenomenon when the equipment fails, fault description information, test conditions performed during the fault diagnosis process, fault-related test detection description feature information, fault causes and other information.

[0036] In an embodiment of the present invention, the transformer fault category sample text reflects the fault type label, and each transformer fault information sample text may correspond to multiple fault labels. Transformer fault labels are mainly divided into internal faults and external faults, among which internal faults refer to various faults in the transformer oil tank, which mainly rely on gas and differential protection to cut off the transformer; external faults refer to various faults occurring in the insulating bushing outside the oil tank and its lead wires. At present, common transformer faults are summarized into the following four categories: winding faults, tap changer faults, core faults and gas protection failure, among which winding faults mainly include turn-to-turn short circuits, winding grounding, phase-to-phase short circuits, wire breaks and joint open welding. Common tap changer faults include surface melting, phase-to-phase contact discharge or combined discharge.

[0037] In the transformer fault analysis report, due to the differences in the knowledge reserves, language habits, observation angles and observation integrity of the recorders, the transformer fault text presents diversity and confusion, and the fault text category is uncertain. The same transformer may have multiple fault category labels, and these fault category labels may have label types with the same meaning. For example, phase-to-phase short circuit and coil short circuit may both refer to direct short circuits between different phases or different coils, which leads to repeated label types for transformer faults. In addition, for example, a transformer simultaneously has turn-to-turn short circuit, surface melting and phase-to-phase contact discharge faults. At the same time, among these faults, only a few transformers may have joint discharge faults, which will lead to imbalance in the transformer fault text data. Therefore, the embodiment of the present invention uses a data enhancement training method and adopts the text label domain screening method TLDF to enhance the fault information data of a few categories to optimize the model's training effect in the case of label imbalance.

[0038] 102. Input the fault category text into an embedding layer of a text semantic model based on a coding structure, and output a semantic coding vector of the fault category text via the coding layer of the text semantic model based on the coding structure;

[0039] In an embodiment of the present invention, the step of obtaining a semantic encoding vector of a fault category text includes:

[0040] Step 1: Pass all the category label texts of transformer fault through the word segmenter to obtain the corresponding token sequence;

[0041] Step 2: Concatenate the token sequences of different fault categories with the corresponding classification heads to obtain the concatenated token sequences of transformer fault sample category texts; concatenate the token sequences of each category of label texts according to the following formula:

[0042]

[0043] Where, T L The token sequence representing the concatenated transformer fault category text, The token sequence representing the label text of the k-th transformer fault, [CLS] k The classification header token representing the k-th transformer fault category.

[0044] Step 3: Determine the position of each classification head in the token sequence of the concatenated transformer fault category text; record the position of each [CLS] classification token as P CLS ;

[0045] Step 4: T L Input the embedding layer of the text semantic model based on the encoding structure to obtain the transformer fault category text semantic vector V L .

[0046] 103. Processing the fault information text into a to-be-enhanced text by a text label domain screening method, and forming a to-be-enhanced text pair with the fault information text before processing and the to-be-enhanced text;

[0047] In an embodiment of the present invention, the text label domain screening method includes:

[0048] Step 1: Filter out the minority fault information text C whose data volume is lower than the average data volume from the fault information text, and divide the minority fault information text C into a sequence S consisting of multiple language blocks according to punctuation marks. C ;

[0049] Step 2: Input the few-class fault information text C as the original text C into the pre-trained transformer fault intelligent recognition model to obtain the confidence R of the original text C for each fault category;

[0050] Step 3: Translate each chunk of the original text C The content is removed to obtain semantically incomplete text Will Input the pre-trained transformer fault intelligent recognition model to obtain incomplete text Confidence level for each fault category

[0051] Step 4: Compare the confidence R of the original text C for each fault category with the incomplete text Confidence level for each fault category Processing to obtain confidence The Euclidean distance with confidence R, and get each chunk The label domain confidence difference value of ;

[0052] Step 5: Repeat step 3 to obtain the label domain confidence difference values ​​of all language blocks. After sorting, take the language block with the largest label domain confidence difference value. Corresponding incomplete text is the text to be enhanced, and forms a text pair to be enhanced with the original text C

[0053] It can be understood that the confidence level described in the present invention is the classification prediction probability of each fault category. For example, if the confidence level of the original text C about the winding fault is 0.7, it means that the probability of the winding fault is 0.7.

[0054] In this embodiment, in order to obtain the target enhanced text vector, a series of processing needs to be performed on the to-be-enhanced text pair. The processing will be described in detail below in conjunction with steps 104 to 106:

[0055] 104. Input the to-be-enhanced text pair into an embedding layer of a text semantic model based on an encoding structure, and output an initial enhanced text vector pair;

[0056] 105. The initial enhanced text vector is input into a generative adversarial network based on a label domain mask to train, and an intermediate enhanced text vector is output;

[0057] 106. Pass the intermediate enhanced text vector through a linear layer to obtain a target enhanced text vector;

[0058] In the embodiment of the present invention, more specifically, Figure 3 As shown, the process of obtaining the target enhanced text vector includes:

[0059] Step 1: Treat the enhanced text pair After each text in is segmented, the text to be enhanced is compared with the minority fault information, that is, the original text C. Each chunk Position P CLS Fill the sequence with the mask token sequence

[0060] Step 2: Sequence After passing through the embedding layer of the text semantic model based on the encoding structure, it is input into the generator of the generative adversarial network to obtain the disguised text semantic vector

[0061] Step 3: After the original text C passes through the embedding layer of the text semantic model based on the encoding structure, the original text semantic vector V is obtained;

[0062] Step 4: Disguise text semantic vector The discriminator of the LM-GAN network is input with the original text semantic vector V to judge the authenticity of the text vector. is judged as “true”, it is used as the enhanced text vector of the original text C. If the disguised text semantic vector If it is judged as "fake", the original text semantic vector V and the disguised text semantic vector Calculate the loss training generator and use the discriminator to compare the original text semantic vector V with the disguised text semantic vector The identification results and the true labels are used to calculate the loss for training the discriminator.

[0063] 107. Directly input the target enhanced text vector into the encoding layer of the text semantic model based on the encoding structure, and simultaneously input the fault information text into the embedding layer of the text semantic model based on the encoding structure, and output the fault information text semantic encoding vector through the encoding layer of the text semantic model based on the encoding structure;

[0064] In an embodiment of the present invention, the process of obtaining the semantic encoding vector of the fault information text may include: Step 1: Inputting the fault information text into the embedding layer of the text semantic model based on the encoding structure to obtain the fault information text embedding vector V ori ;

[0065] Step 2: Pass the target enhanced text vector through the Linear layer to obtain the target enhanced text embedding vector V LM-GAN ;

[0066] Step 3: Embed the target enhanced text into vector V LM-GAN and the fault information text embedding vector V ori Merge into the total text embedding vector V all ;

[0067] Step 4: Embed the total text into vector V all Input the encoding layer of the text semantic model based on the encoding structure to obtain the fault category text semantic vector V L .

[0068] 108. Input the fault information text semantic encoding vector and the fault category text semantic vector into a cross-attention network, process them through a classifier, and output a transformer fault category result vector; the classifier includes an output Dropout layer, a linear layer, and an activation sigmoid layer;

[0069] In an embodiment of the present invention, the step of obtaining a transformer fault classification result includes:

[0070] Step 1: Transform the fault category text semantic vector V L With the weight matrix W q Multiply them together to get the Q matrix, and convert the fault text semantic encoding vector V C Respectively with the weight matrix W k , W v After multiplication, we get K and V matrices;

[0071] Step 2: Input the Q, K, and V matrices into the cross-attention network to obtain the interaction result vector O CA ;

[0072] Step 3: Transform the interaction result vector O CA Middle position P CLS The vectors at are extracted and concatenated according to the sequence dimension to obtain the vector

[0073] Step 4: Convert the vector The hidden layer dimension is scaled to 2 through the Dropout layer and the Linear layer in sequence, and the hidden layer is normalized by the Sigmoid function to obtain the transformer fault category result vector O.

[0074] The present invention uses a label chain vector spliced ​​from fault category texts to assist in decoding the encoding vector of a text semantic model based on an encoding structure, so that the classifier breaks away from the monotonous label mapping numerical space, fully utilizes the label semantic representation information, and improves classification accuracy.

[0075] 109. The transformer fault category result vector is combined with multiple transformer fault category texts to construct a multi-label and label imbalance adaptive loss function, and the transformer fault intelligent recognition model is obtained through loss training.

[0076] In this embodiment of the present invention, the multi-label and label imbalance adaptive loss function is as follows:

[0077]

[0078] In the formula, m represents the number of samples of fault information text when calculating the loss, k represents the number of labels in the fault text, represents the jth true label of the i-th fault information text, represents the j-th predicted label of the i-th fault information text, ln(·) represents the natural logarithm function, and cos(·) represents the cosine function.

[0079] The present invention adopts the multi-label and label imbalance adaptive loss function MSLoss, which enables the model to adjust the weight of the loss according to the classification accuracy of the correct label when calculating the loss, and uses the cosine function to adaptively coordinate the overall loss scale during training and the sensitivity of the correct label and the wrong label to the confidence, thereby improving the accuracy of the transformer fault intelligent identification model.

[0080] Based on the same concept of the present invention, the present invention also provides a training device for a text-based transformer fault intelligent identification model, the device comprising:

[0081] An acquisition module, used to acquire transformer fault text data, wherein the transformer fault text data includes transformer fault information text and corresponding multiple transformer fault category texts, each fault category text indicating a fault category label;

[0082] The training module is used to input the fault category text into a text semantic model based on a coding structure, and output a fault category text semantic coding vector; process the fault information text into a to-be-enhanced text through a text label domain screening method, and form a to-be-enhanced text pair with the fault information text before processing and the to-be-enhanced text; input the to-be-enhanced text pair into a text semantic model based on a coding structure, and output an initial enhanced text vector pair; input the initial enhanced text vector pair into a generative adversarial network based on a label domain mask for training, and output an intermediate enhanced text vector; pass the intermediate enhanced text vector through a linear layer to obtain a target enhanced text vector; input the fault information text into a text semantic model based on a coding structure, and output an initial enhanced text vector pair; A text semantic model of a coding structure is constructed, and the target enhanced text vector is input into the coding layer of the text semantic model based on the coding structure at the same time, and a fault information text semantic coding vector is output; the fault information text semantic coding vector and the fault category text semantic vector are input into a cross-attention network, and processed by a classifier to output a transformer fault category result vector; the classifier includes an output Dropout layer, a linear layer and an activation sigmoid layer; the transformer fault category result vector is combined with multiple transformer fault category texts to construct a multi-label and label imbalance adaptive loss function, and a transformer fault intelligent recognition model is obtained through loss training.

[0083] The transformer fault intelligent recognition model processed by the training method and device of the present invention can be applied to the processing process of unknown transformer fault information text. When the unknown transformer fault information text is input into the transformer fault intelligent recognition model of the present invention, the corresponding fault type information can be output, thereby realizing the intelligent recognition of the transformer fault information text.

[0084] It can be understood that the training method and device of the text-based transformer fault intelligent identification model of the present invention belong to the same concept of the present invention, and their corresponding features can be referenced to each other, so the present invention will not repeat them.

[0085] A person skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium, which can include: ROM, RAM, disk or CD, etc.

[0086] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A training method for a text-based transformer fault intelligent identification model, characterized in that: The following steps are involved: Obtain transformer fault text data, wherein the transformer fault text data includes transformer fault information text and corresponding multiple transformer fault category texts, each fault category text indicating a fault category label; The steps of obtaining the semantic encoding vector of the fault category text include: Step 11: Pass all transformer fault category texts through the word segmenter to obtain the corresponding token sequences of different fault categories; Step 12: Concatenate the token sequences of different fault categories with the corresponding classification heads to obtain the concatenated token sequence of the transformer fault sample category text; Step 13: Determine the position of each classification head in the token sequence of the concatenated transformer fault category text; Step 14: Input the concatenated token sequence of the transformer fault sample category text into the embedding layer of the text semantic model based on the encoding structure to obtain the transformer fault sample category text semantic vector; Input the fault category text into the embedding layer of the text semantic model based on the encoding structure, and output the fault category text semantic encoding vector through the encoding layer of the text semantic model based on the encoding structure; The fault information text is processed into a text to be enhanced by a text label domain screening method, and the fault information text before processing and the text to be enhanced form a text pair to be enhanced; the text label domain screening method includes: Step 21: Filter out the minority fault information texts whose data volume is lower than the average data volume from the fault information texts, and divide the minority fault information texts into a sequence consisting of multiple language blocks according to punctuation marks; Step 22: Input the few-category fault information text as the original text into the pre-trained transformer fault intelligent recognition model to obtain the confidence of the original text for each fault category; Step 23: remove the content of each language block in the original text to obtain semantically incomplete text, input the semantically incomplete text into the pre-trained transformer fault intelligent recognition model, and obtain the confidence of the incomplete text for each fault category; Step 24: Process the confidence of the original text on each fault category and the confidence of the semantically incomplete text on each fault category to obtain the Euclidean distance between the two confidences, and obtain the label domain confidence difference value of each language block; Step 25: Repeat steps 23 to 24 to obtain the label domain confidence difference values ​​of all language blocks. After sorting the label domain confidence difference values ​​of all language blocks, take the incomplete text corresponding to the language block with the largest label domain confidence difference value as the text to be enhanced, and form a text pair to be enhanced with the original text; Input the text pair to be enhanced into the embedding layer of the text semantic model based on the encoding structure, and output an initial enhanced text vector pair; The initial enhanced text vector is input into the generative adversarial network based on the label domain mask to train and output the intermediate enhanced text vector; wherein, the structure of the generative adversarial network LM-GAN based on the label domain mask includes a generator and a discriminator, the generator includes an encoder encoder and a decoder decoder, the encoder encoder includes an encoder block of three layers of transformers, the decoder decoder includes a decoder block of three layers of transformers, and the discriminator includes a BiLSTM bidirectional long short-term memory network, a linear layer, and a Sigmoid activation function; The intermediate enhanced text vector is passed through the linear layer to obtain the target enhanced text vector. The process of obtaining the target enhanced text vector includes: Step 31: After segmenting each text in the to-be-enhanced text pair, the text is compared with the minority fault information, i.e., the original text, and the position of each language block in the to-be-enhanced text is filled with the mask token sequence to obtain an updated sequence; Step 32: After the update sequence passes through the embedding layer of the text semantic model based on the encoding structure, it is input into the generator of the generative adversarial network to obtain the disguised text semantic vector; Step 33: After the original text passes through the embedding layer of the text semantic model based on the encoding structure, the original text semantic vector is obtained; Step 34: Input the disguised text semantic vector and the original text semantic vector into the discriminator of the LM-GAN network to judge the authenticity of the text vector. If the disguised text semantic vector is judged to be "true", it is used as the enhanced text vector of the original text. If the disguised text semantic vector is judged to be "false", the original text semantic vector and the disguised text semantic vector are used to calculate the loss training generator, and the discriminator is used to calculate the loss training of the discriminator based on the identification results of the original text semantic vector and the disguised text semantic vector and the real label; the target enhanced text vector is directly input into the encoding layer of the text semantic model based on the encoding structure, and the fault information text is simultaneously input into the embedding layer of the text semantic model based on the encoding structure, and the fault information text semantic encoding vector is output through the encoding layer of the text semantic model based on the encoding structure; Input the fault information text semantic encoding vector and the fault category text semantic vector into the cross-attention network, process them through the classifier, and output the transformer fault category result vector; the classifier includes an output Dropout layer, a linear layer, and an activation sigmoid layer; The transformer fault category result vector is combined with multiple transformer fault category texts to construct a multi-label and label imbalance adaptive loss function, and a transformer fault intelligent recognition model is obtained through loss training. The multi-label and label imbalance adaptive loss function is as follows: In the formula, m represents the number of samples of fault information text when calculating the loss, k represents the number of labels in the fault text, represents the jth true label of the i-th fault information text, represents the j-th predicted label of the i-th fault information text, ln(·) represents the natural logarithm function, and cos(·) represents the cosine function.

2. The training method of a text-based transformer fault intelligent identification model according to claim 1 is characterized in that: The steps to obtain transformer fault classification results include: Step 41: Transform the fault category text semantic vector V L With the weight matrix W q Multiply them together to get the Q matrix, and convert the fault text semantic encoding vector V C Respectively with the weight matrix W k , W v After multiplication, we get K and V matrices; Step 42: Input the Q, K, and V matrices into the cross-attention network to obtain the interaction result vector O CA ; Step 43: Transform the interaction result vector O CA Middle position P CLS The vectors at are extracted and concatenated according to the sequence dimension to obtain the vector Step 44: Vector The hidden layer dimension is scaled to 2 through the Dropout layer and the Linear layer in sequence, and the hidden layer is normalized by the Sigmoid function to obtain the transformer fault category result vector O.

3. A training device for a text-based transformer fault intelligent recognition model, which is used to implement the training method for a text-based transformer fault intelligent recognition model as claimed in claim 1 or 2, characterized in that: The device comprises: An acquisition module, used to acquire transformer fault text data, wherein the transformer fault text data includes transformer fault information text and corresponding multiple transformer fault category texts, each fault category text indicating a fault category label; The training module is used to input the fault category text into a text semantic model based on a coding structure, and output a fault category text semantic coding vector; process the fault information text into a to-be-enhanced text through a text label domain screening method, and form a to-be-enhanced text pair with the fault information text before processing and the to-be-enhanced text; input the to-be-enhanced text pair into a text semantic model based on a coding structure, and output an initial enhanced text vector pair; input the initial enhanced text vector pair into a generative adversarial network based on a label domain mask for training, and output an intermediate enhanced text vector; pass the intermediate enhanced text vector through a linear layer to obtain a target enhanced text vector; input the fault information text into a text semantic model based on a coding structure, and output an initial enhanced text vector pair; A text semantic model of a coding structure is constructed, and the target enhanced text vector is input into the coding layer of the text semantic model based on the coding structure at the same time, and a fault information text semantic coding vector is output; the fault information text semantic coding vector and the fault category text semantic vector are input into a cross-attention network, and processed by a classifier to output a transformer fault category result vector; the classifier includes an output Dropout layer, a linear layer and an activation sigmoid layer; the transformer fault category result vector is combined with multiple transformer fault category texts to construct a multi-label and label imbalance adaptive loss function, and a transformer fault intelligent recognition model is obtained through loss training.