Transformer fault diagnosis method and system, and computer program product

Abstract

The invention belongs to a fault diagnosis method, and provides a transformer fault diagnosis method, a transformer fault diagnosis system and a computer program product aiming at the technical problems that an existing multi-modal data fusion method cannot deeply mine complex correlation among modals, is difficult to generalize in application and neglects potential complementary information among different data. The transformer fault diagnosis result can be quickly and accurately obtained by obtaining the power time sequence signal of the transformer and the archived text data of the transformer. The multi-modal large model comprises a modal feature extractor, a feature mapper and a large language model, effective fusion and feature extraction of multi-modal data are achieved, then the precision and robustness of transformer fault diagnosis are remarkably improved, the problems that in the prior art, data are single, information is isolated, and the intelligent level is insufficient are solved, and the fault diagnosis efficiency is improved. Therefore, more comprehensive and accurate fault identification and classification capability can be provided in a complex scene.

Description

Technical Field

[0001] The present application relates to a fault diagnosis method, and specifically to a transformer fault diagnosis method, system, and computer program product. Background Art

[0002] With the development of intelligent power systems, the diversity and quantity of data are increasing. The data not only contains a large amount of time series signal data (such as voltage, current, frequency, etc.), but also includes rich text data (such as fault reports, maintenance logs, equipment status descriptions, etc.). However, how to effectively integrate data of different modes and fully explore the potential complementary information of these data has become an important challenge in current research.

[0003] However, most of the existing multimodal data fusion methods have the following shortcomings: (1) Traditional fusion methods usually simply concatenate or weighted average data from different modalities, and cannot deeply explore the complex relationships between modalities. (2) Many fusion methods rely on a large amount of labeled data, which is expensive to obtain and difficult to generalize in applications. (3) Single-modal data processing methods often ignore the potential complementary information between different data sources, resulting in poor performance of the model when facing complex scenarios. Summary of the invention

[0004] The present application aims to provide a transformer fault diagnosis method, system, and computer program product to address the technical problems that existing multimodal data fusion methods are unable to deeply explore the complex correlations between modes, are difficult to generalize in applications, and ignore the potential complementary information between different data.

[0005] In order to achieve the above objectives, this application adopts the following technical solutions: In a first aspect, the present application proposes a transformer fault diagnosis method, comprising: Obtain the transformer's power timing signal and transformer archived text data; Inputting the power time series signal and the text data into a multimodal large model to obtain a transformer fault diagnosis result; The multimodal large model includes a modal feature extractor, a feature mapper and a large language model; The modal feature extractor is used to process each modal signal in the power time series signal respectively to generate a corresponding feature vector; The feature mapper is used to map the feature vectors generated corresponding to each modal signal into the feature space of the text data to obtain mapping features; The large language model is used to extract text features from the text data, and fuse the text features with mapping features to obtain transformer fault diagnosis results.

[0006] Further, the modal feature extractor includes a position encoding module, a multi-layer Transformer encoder and a multi-layer Transformer decoder connected in sequence; The position encoder is used to obtain the position information of each modal signal in the power time series signal; The multi-layer Transformer encoder is used to gradually extract the multi-level features of each modal signal in the power time series signal through a self-attention mechanism and a feedforward neural network; The multi-layer Transformer decoder is used to step by step input the original data according to the position information and the extracted multi-level features to generate the feature vector corresponding to each modal signal.

[0007] Furthermore, the feature mapper adopts a KAN network.

[0008] Furthermore, the training method of the multimodal large model includes: The multimodal large model is trained using positive sample pairs and negative sample pairs; the positive sample pairs include power timing signals and corresponding fault description texts, and the negative sample pairs include power timing signals and irrelevant fault description texts; The loss function used in training is the InfoNCE loss function.

[0009] Furthermore, the InfoNCE loss function includes:

[0010] in, represents the InfoNCE loss function, Represents a positive sample pair The similarity of is the temperature parameter, represents the power timing signal in the positive sample pair, Represents the fault description text in the positive sample pair.

[0011] Furthermore, the text data archived by the transformer includes documents corresponding to the maintenance report, fault record and maintenance log of the transformer.

[0012] In a second aspect, the present application proposes a transformer fault diagnosis system, comprising: An acquisition module, used to acquire the power timing signal of the transformer and the text data archived by the transformer; A diagnosis module, used for inputting the power time sequence signal and the text data into a multi-modal large model to obtain a transformer fault diagnosis result; The multimodal large model includes a modal feature extractor, a feature mapper and a large language model; The modal feature extractor is used to process each modal signal in the power time series signal respectively to generate a corresponding feature vector; The feature mapper is used to map the feature vectors generated corresponding to each modal signal into the feature space of the text data to obtain mapping features; The large language model is used to extract text features from the text data, and fuse the text features with mapping features to obtain transformer fault diagnosis results.

[0013] Further, the modal feature extractor includes a position encoding module, a multi-layer Transformer encoder and a multi-layer Transformer decoder connected in sequence; The position encoder is used to obtain the position information of each modal signal in the power time series signal; The multi-layer Transformer encoder is used to gradually extract the multi-level features of each modal signal in the power time series signal through a self-attention mechanism and a feedforward neural network; The multi-layer Transformer decoder is used to step by step input the original input according to the position information and the extracted multi-level features to generate the feature vector corresponding to each modal signal; The feature mapper adopts a KAN network.

[0014] Furthermore, the training method of the multimodal large model includes: The multimodal large model is trained using positive sample pairs and negative sample pairs; the positive sample pairs include power timing signals and corresponding fault description texts, and the negative sample pairs include power timing signals and irrelevant fault description texts; The loss function used in training is the InfoNCE loss function.

[0015] In a third aspect, the present application relates to a computer program product, which includes instructions, and when the instructions are executed by a processor, the above-mentioned transformer fault diagnosis method is implemented.

[0016] Compared with the prior art, this application has the following beneficial effects: This application proposes a transformer fault diagnosis method, which can quickly and accurately obtain transformer fault diagnosis results by acquiring the transformer's power timing signal and the transformer's archived text data with the help of a multimodal large model. The multimodal large model includes a modal feature extractor, a feature mapper, and a large language model, which realizes the effective fusion and feature extraction of multimodal data, thereby significantly improving the accuracy and robustness of transformer fault diagnosis, overcoming the problems of single data, isolated information, and insufficient intelligence level in the prior art, thereby providing more comprehensive and accurate fault identification and classification capabilities in complex scenarios.

[0017] The present application also proposes a transformer fault diagnosis system and a computer program product, which have all the advantages of the above transformer fault diagnosis method. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 A first flow chart of a fault diagnosis method for this application; Figure 2 A second flow chart of a fault diagnosis method of the present application; Figure 3 This is a schematic diagram of a fault diagnosis system of the present application. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for which protection is sought, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0022] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0023] In the description of the embodiments of the present application, it should be noted that if the terms "upper", "lower", "horizontal", "inner", etc. appear, the orientation or position relationship indicated is based on the orientation or position relationship shown in the drawings, or the orientation or position relationship in which the invented product is usually placed when used. It is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0024] In addition, if the term "horizontal" appears, it does not mean that the component must be absolutely horizontal, but can be slightly tilted. For example, "horizontal" only means that its direction is more horizontal than "vertical", which does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0025] In the description of the embodiments of the present application, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", and "connect" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal connection of two components. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0026] As the power system transforms towards intelligence, the complexity of data processing and analysis also increases. The data generated in this process includes not only continuously changing time series signals (such as voltage, current and frequency, which fluctuate continuously over time and record the real-time status of the system), but also covers a large amount of text information (such as fault reports that describe the type, cause and impact of faults in detail, maintenance logs that record the time, content and results of maintenance activities, and equipment status descriptions that provide the current status and historical performance of equipment). Effectively integrating and utilizing these multimodal data is crucial to improving the operating efficiency of power systems, preventing faults, and responding to problems in a timely manner.

[0027] At present, traditional multimodal data fusion still has the following problems: (1) Simple splicing or weighted averaging: Traditional multimodal data fusion methods often adopt the most direct approach, which is to directly splice data from different modalities together, or obtain a comprehensive indicator through a weighted averaging algorithm. This method ignores the complex interactions and deep associations that may exist between data from different modalities. For example, the changes in time series signals may be intrinsically related to specific maintenance activities or equipment status descriptions, but simple splicing or averaging cannot capture this relationship.

[0028] (2) Reliance on large amounts of labeled data: Many advanced data fusion techniques, especially those based on deep learning, require large amounts of labeled data to train models. However, in practical application scenarios such as power systems, obtaining high-quality, large-scale labeled data is very difficult and costly. In addition, even if sufficient labeled data can be obtained, the generalization ability of the model in new and unseen scenarios may be limited because the actual operating conditions are far more complex and changeable than the training data.

[0029] (3) Ignoring complementary information: Single-modal data processing methods often focus only on extracting information within that modality, while ignoring the possible complementarity between different modalities. For example, time series signal data can provide immediate feedback on the state of the system, while text data may contain the historical background and potential causes of such state changes. If these two types of information cannot be effectively integrated, the model may not be able to fully understand the true state of the system, resulting in deviations in predictions or decisions in complex scenarios.

[0030] In response to these problems, in recent years, the introduction of Transformer autoencoders and large language models has provided new ideas for multimodal data processing. These models have performed well in fields such as natural language processing and computer vision, and have powerful feature extraction and data fusion capabilities. However, in power system fault diagnosis, how to effectively apply these advanced models and combine the characteristics of multimodal data still needs in-depth exploration.

[0031] Based on the above situation, the present application proposes a transformer fault diagnosis method, system, and computer program product. The present application is described in detail below in conjunction with embodiments and drawings.

[0032] like Figure 1 As shown, it is a first flow chart of a transformer fault diagnosis method of the present application, which may include: S101, obtaining the power timing signal of the transformer and the text data archived by the transformer.

[0033] It should be noted that power time series signals usually include signals that change over time, such as voltage, current, and frequency, which can directly reflect the operating status of the transformer. These power time series signals can be acquired in real time through sensors installed on the transformer and recorded in the form of time series. In practical applications, the analysis of power time series signals helps to discover abnormal fluctuations or trends in the transformer and can provide important clues for fault diagnosis.

[0034] The text data archived by transformers usually includes text information such as fault reports, maintenance logs, and equipment status descriptions. These text information records the transformer's historical fault conditions, maintenance activities, and the current and past status of the equipment. Text data can provide rich background information to help understand the possible causes and context of transformer failures.

[0035] S102, inputting the power time series signal and the text data into a multimodal large model to obtain a transformer fault diagnosis result.

[0036] Wherein, the multimodal large model includes a modal feature extractor, a feature mapper and a large language model; The modal feature extractor is used to process each modal signal in the power time series signal respectively to generate a corresponding feature vector; The feature mapper is used to map the feature vectors generated corresponding to each modal signal into the feature space of the text data to obtain mapping features; The large language model is used to extract text features from the text data, and fuse the text features with mapping features to obtain transformer fault diagnosis results.

[0037] It should be noted that the modal feature extractor is responsible for processing each modal signal (such as voltage, current, etc.) in the power time series signal separately. It can use specific algorithms or models (such as convolutional neural networks, recurrent neural networks, etc.) to extract the features of each modal signal and generate corresponding feature vectors. These feature vectors are an abstract representation of the original signal and contain the key information of the signal. The task of the feature mapper is to map the feature vectors generated corresponding to each modal signal to the feature space of the text data. It can convert data of different modes (such as time series signals and text data) into the same feature space, making it possible to compare and merge them. The mapped features retain the key information of the original feature vectors and have a certain similarity with the features of the text data. The large language model is the core part of the multimodal large model and is responsible for extracting text features from text data. The large language model is usually based on deep learning technology and can capture semantic information and contextual relationships in the text. After extracting the text features, the large language model will fuse the text features with the mapping features. The fused features contain the comprehensive information of the time series signal and the text data, providing a more comprehensive basis for transformer fault diagnosis. Finally, the large language model will use the fused features for reasoning and judgment to obtain the fault diagnosis results of the transformer.

[0038] like Figure 2 As shown, it is a second flow chart of a transformer fault diagnosis method of the present application, which may include: S201, multimodal data acquisition and preprocessing.

[0039] Multimodal data refers to data information from different sources or in different forms. In this application, it mainly includes power timing signal data and text data. Among them, the power timing signal data is composed of sensor signals such as current, voltage, temperature and oil chromatography collected during the operation of the transformer. These signals reflect the real-time operating status of the transformer. The text data comes from documents such as the transformer's maintenance reports, fault records and maintenance logs. These documents contain a lot of expert experience and descriptions of historical faults. During the data collection process, the power timing signal can be obtained in real time through special sensors and data acquisition equipment, while the text data can be extracted from the maintenance management system or collected by manual entry.

[0040] The following methods can be used to preprocess the power time series signal data: (1) Clean and filter the collected raw data to remove noise and outliers.

[0041] (2) Normalize the cleaned and filtered power time series signal data to eliminate the dimensional differences between different physical quantities. For example, normalize the signal data such as current, voltage, temperature and oil chromatography separately to make their numerical ranges consistent for subsequent feature extraction.

[0042] (3) In order to further enhance the model’s ability to capture temporal features, the time series signal data can be differentiated, smoothed, or windowed as needed to divide the long-span data into several small time segments, so that the model can effectively learn the changing trends and patterns in the time series.

[0043] Natural Language Processing (NLP) technology can be used to preprocess text data. First, the text data is subjected to basic processing such as word segmentation, part-of-speech tagging, and stop word removal. Then, the processed text data is converted into a vector representation. This step usually uses word embedding technology. In this embodiment, Word2Vec can be used to map text information into a high-dimensional vector space to facilitate fusion with time series signal data.

[0044] Specifically, word segmentation is to recombine a continuous sequence of characters into a sequence of words according to certain specifications. For Chinese text, word segmentation is the first step of preprocessing because Chinese characters themselves do not directly represent words and need to be segmented into meaningful word units through word segmentation. Then, part-of-speech tagging can be used to assign a part of speech to each word in the text, such as noun, verb, adjective, etc. Part-of-speech tagging helps subsequent natural language processing tasks, such as semantic analysis and anaphora resolution. Part-of-speech tagging can be completed manually or by specific algorithms, such as Hidden Markov Model, Conditional Random Field, etc. Stop words refer to words that are very common in natural language text but have no practical help or contribution to text analysis, such as "de", "le", "shi", etc. These stop words will interfere with the training effect of the model. Therefore, they need to be removed in text preprocessing. Removing stop words usually depends on a predefined list of stop words, and these lists can be customized according to different languages and application scenarios. Word embedding is a technique that represents words as real-valued vectors. By mapping words into a low-dimensional vector space, words with similar semantics are closer in the vector space. Word embedding can not only capture the semantic relationships between words but also help reduce computational complexity. After steps such as word segmentation, part-of-speech tagging, removing stop words, and word embedding, the text data is transformed into vector representations. These vectors can be used as inputs for machine learning models and can be used for tasks such as text classification, analysis, and named entity recognition. The dimension and nature of the vector representation depend on the choice and training of the word embedding model.

[0045] S202, design the network structure of the multimodal large model.

[0046] In this embodiment, the multimodal large model includes a modal feature extractor, a feature mapper, Among them, the modal feature extractor is responsible for extracting feature vectors that can represent the device state from data of different modalities. Specifically, the modal feature extractor processes power time series signals (such as current, voltage, temperature, oil chromatography, etc.) and extracts key features in each modality through advanced deep learning models, such as the Transformer autoencoder.

[0047] Power time series signals usually contain rich dynamic information, such as the fluctuations of current and voltage, the temperature changes of equipment, and the changes in gas components in oil chromatography analysis, etc. These signals have the characteristics of time series, that is, the values of the signals change continuously over time and are often affected by multiple factors, such as load changes, equipment aging, environmental conditions, etc. Modal feature extraction refers to extracting feature quantities that reflect different dynamic characteristics of the system from the original signals. For power time series signals, through modal feature extraction, the operating state of the power system can be better understood, the health status of the equipment can be monitored, and potential faults can be predicted. These feature quantities are crucial for the safe and stable operation of the power system.

[0048] Transformer autoencoder is a deep learning architecture based on the Transformer model, designed to process sequence data, especially for time series signals with obvious long-term and short-term dependencies. Because the Transformer model has a self-attention mechanism, it can effectively capture the dependencies between distant time points in the sequence without relying on traditional recurrent neural networks. The autoencoder structure consists of two parts: the encoder and the decoder. The encoder is responsible for mapping the input time series signal into a feature vector in a low-dimensional feature space, and the decoder attempts to reconstruct the original input from the feature vector, thereby ensuring that the extracted features retain the main information of the input signal during the training process.

[0049] In the modal feature extractor, the power time series signal is first processed by the Transformer encoder to generate the corresponding feature vector. The specific process is: First, the input power time series signal passes through the PositionalEncoding module to obtain the position information in the sequence. Then, the input power time series signal passes through several layers of Transformer encoders. In each layer of encoder, the multi-level features in the power time series signal are gradually extracted through the self-attention mechanism and the feed-forward neural network. The self-attention mechanism enables the model to focus on the key time points in the signal, thereby capturing complex timing dependencies. After multi-layer processing of the Transformer encoder, the final generated feature vector contains a high-dimensional representation of the input signal, which can fully describe the dynamic changes and potential patterns of the signal.

[0050] For power time series signals of different modes (current, voltage, temperature, oil chromatography, etc.), each mode of power time series signal will generate a corresponding feature vector after being processed by an independent Transformer autoencoder. The modal feature extractor ensures that different types of time series signals can be represented by a unified high-dimensional feature space, thereby achieving effective fusion and collaborative processing of different modal data. In addition, the extracted feature vectors also have powerful expressive capabilities, which can identify subtle changes in the operating status of equipment in complex fault scenarios, significantly improving the accuracy and robustness of fault diagnosis.

[0051] The feature mapper is responsible for mapping feature vectors of different modalities into a unified feature space, thereby achieving effective fusion of multimodal data.

[0052] The core task of the feature mapper is to project feature vectors from different modes (such as current, voltage, temperature, oil chromatography and other time series signals) into the feature space of text data. This process aims to solve the heterogeneity problem between multimodal data, so that the features of different modes can be expressed in the same coordinate system, which is convenient for subsequent model processing and fusion analysis. In multimodal data fusion, data of different modes often have different physical properties and statistical characteristics, so an effective mapping mechanism is needed to unify these differentiated features into a common representation space.

[0053] In this embodiment, the feature mapper can use a KAN network. The structure of the KAN network is based on the Kolmogorov-Arnold representation theorem, which states that any continuous function of multiple variables can be represented as a combination of a finite number of single variable functions. In the KAN network, this representation is used to design the layers and connections of the network. The KAN network structure in this embodiment is as follows, including an input layer, a hidden layer, and an output layer: (1) Input layer: The KAN network receives multi-dimensional input vectors, each of which is passed to multiple univariate functions for processing.

[0054] (2) Hidden layer: The neurons in each layer are actually constructed through multiple learnable univariate functions (such as B-spline functions or Gaussian radial basis functions), which are used to perform nonlinear transformations on the input vectors.

[0055] (3) Output layer: The features processed by multiple hidden layers are combined into an output, usually achieved through addition or other simple operations.

[0056] In traditional neural networks, the activation function is fixed, such as ReLU or Sigmoid. But in KAN networks, the activation function is learnable, usually using B-spline or RBF functions. The shapes and parameters of these activation functions are automatically adjusted during training to best fit the data.

[0057] A key feature of the KAN network is that it does not use traditional linear weights. Instead, the "weights" in the KAN network are expressed by nonlinear functions of a single variable, which are applied to the edges of each connection. This makes each connection in the KAN network not just a linear transformation, but a complex nonlinear mapping. In summary, through the feature mapper, feature vectors of different modalities are mapped into the text feature vector space.

[0058] S203, multimodal large model pre-training based on cross-modal contrastive learning.

[0059] Methods for pre-training multimodal large models may include: (1) Prepare positive and negative sample pairs.

[0060] First, extract a pair of related data from the multimodal dataset. For example, a power time series signal and its corresponding fault description text can be used as a positive sample pair. These positive sample pairs represent the same fault information under different modes. Then, randomly select unrelated modal data pairs as negative sample pairs. For example, a power time series signal and an irrelevant text description. The goal of these negative sample pairs is to help the model learn the difference between modes.

[0061] After modal feature extraction and feature mapping, the positive sample pair is converted into a pair of text feature vectors, and the negative sample pair is also converted into a pair of text feature vectors.

[0062] (2) Loss function design.

[0063] Use dot product or cosine similarity to calculate the similarity between positive sample pairs, and the similarity between negative sample pairs.

[0064] In this embodiment, the InfoNCE loss function is used: the commonly used loss function for contrastive learning is the InfoNCE loss. For each positive sample pair, its similarity in the joint feature space is maximized, while the similarity of the negative sample pair is minimized. The specific formula is as follows:

[0065] in, Represents a positive sample pair The similarity of is the temperature parameter, which is used to adjust the distribution range of similarity. Represents the InfoNCE loss function. The positive sample pairs and negative sample pairs are input into the multimodal large model together to calculate their feature representations. The model is then trained and the model parameters are updated using the aforementioned contrastive learning loss function. In order to enhance the generalization ability of the multimodal large model, multiple positive and negative sample pairs are usually introduced into a batch and contrastive learning is performed within the batch. This strategy can help the model better learn the complex relationship between different modalities.

[0066] S204, inputting the power time series signal of the transformer and the text data archived by the transformer into the trained multi-modal large model to obtain the transformer fault diagnosis result.

[0067] The synergy of the large language model combined with the power time series proposed in this application realizes the effective fusion and feature extraction of multimodal data. Compared with traditional methods, this application not only improves the accuracy and robustness of fault diagnosis, but also overcomes the problems of single data, isolated information and insufficient intelligence level, thereby providing more comprehensive and accurate fault identification and classification capabilities in complex scenarios. In addition, in order to solve the problem of scarce labeled data, this application introduces cross-modal comparative learning, which effectively integrates power time series signals with text data through self-supervised learning and shared feature space. The InfoNCE loss function is used to maximize the similarity of positive sample pairs while minimizing the similarity of negative sample pairs, thereby enhancing the modal alignment and discrimination capabilities of the multimodal large model. Not only does it achieve efficient feature fusion between different modalities, but it also significantly improves the performance of the multimodal large model in multimodal tasks, and has strong generalization and robustness, providing more comprehensive and accurate support for fault identification and classification in complex power scenarios.

[0068] like Figure 3 As shown, it is a schematic diagram of a transformer fault diagnosis system proposed in this application, which may include: An acquisition module, used to acquire the power timing signal of the transformer and the text data archived by the transformer; A diagnosis module, used for inputting the power time sequence signal and the text data into a multi-modal large model to obtain a transformer fault diagnosis result; The multimodal large model includes a modal feature extractor, a feature mapper and a large language model; The modal feature extractor is used to process each modal signal in the power time series signal respectively to generate a corresponding feature vector; The feature mapper is used to map the feature vectors generated corresponding to each modal signal into the feature space of the text data to obtain mapping features; The large language model is used to extract text features from the text data, and fuse the text features with mapping features to obtain transformer fault diagnosis results.

[0069] It should be noted that in the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the system embodiments described above are merely schematic. For example, the division of each module is only a logical function division. There may be other division methods in actual implementation. For example, multiple modules can be combined or integrated into another device, or some features can be ignored or not executed. The module described as a separate component may or may not be physically separated. The component displayed as a module may be a physical unit or multiple physical units, that is, it may be located in one place, or it may be distributed in multiple different places. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0070] In addition, each module in each embodiment of the present invention may be integrated into a processing unit, each module may exist physically separately, or two or more modules may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of a software functional unit.

[0071] The present application also proposes a computer program product, which includes instructions, and when the instructions are executed by a processor, the steps in the above transformer fault diagnosis method embodiment are implemented. Alternatively, when the processor executes the computer program, the functions of each module / unit in the above system embodiment are implemented.

[0072] The computer program may be divided into one or more modules / units, and the one or more modules / units are stored in a memory and executed by the processor to accomplish the present invention.

[0073] The carrier for implementing the above-mentioned computer program product may be a computer device, or the computer program product may be stored in a computer storage medium.

[0074] The computer device may be a desktop computer, a notebook computer, a palm computer, a cloud server, etc. The computer device may include, but is not limited to, a processor and a memory.

[0075] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0076] The memory may be used to store the computer programs and / or modules, and the processor implements various functions of the computer device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory.

[0077] If the module / unit integrated in the computer device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0078] The above are only preferred embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A transformer fault diagnosis method, characterized in that: include: Obtain the transformer's power timing signal and transformer archived text data; Inputting the power time series signal and the text data into a multimodal large model to obtain a transformer fault diagnosis result; The multimodal large model includes a modal feature extractor, a feature mapper and a large language model; The modal feature extractor is used to process each modal signal in the power time series signal respectively to generate a corresponding feature vector; The feature mapper is used to map the feature vectors generated corresponding to each modal signal into the feature space of the text data to obtain mapping features; The large language model is used to extract text features from the text data, and fuse the text features with mapping features to obtain transformer fault diagnosis results.

2. A transformer fault diagnosis method according to claim 1, characterized in that: The modal feature extractor includes a position encoding module, a multi-layer Transformer encoder and a multi-layer Transformer decoder connected in sequence; The position encoder is used to obtain the position information of each modal signal in the power time series signal; The multi-layer Transformer encoder is used to gradually extract the multi-level features of each modal signal in the power time series signal through a self-attention mechanism and a feedforward neural network; The multi-layer Transformer decoder is used to step by step input the original data according to the position information and the extracted multi-level features to generate the feature vector corresponding to each modal signal.

3. A transformer fault diagnosis method according to claim 2, characterized in that: The feature mapper adopts a KAN network.

4. A transformer fault diagnosis method according to claim 3, characterized in that: The training method of the multimodal large model comprises: The multimodal large model is trained using positive sample pairs and negative sample pairs; the positive sample pairs include power timing signals and corresponding fault description texts, and the negative sample pairs include power timing signals and irrelevant fault description texts; The loss function used in training is the InfoNCE loss function.

5. A transformer fault diagnosis method according to claim 4, characterized in that: The InfoNCE loss function includes: in, represents the InfoNCE loss function, Represents a positive sample pair The similarity of is the temperature parameter, represents the power timing signal in the positive sample pair, Represents the fault description text in the positive sample pair.

6. A transformer fault diagnosis method according to claim 5, characterized in that: The text data archived in the transformer includes documents corresponding to the transformer's maintenance report, fault record and overhaul log.

7. A transformer fault diagnosis system, characterized in that: include: An acquisition module, used to acquire the power timing signal of the transformer and the text data archived by the transformer; A diagnosis module, used for inputting the power time sequence signal and the text data into a multi-modal large model to obtain a transformer fault diagnosis result; The multimodal large model includes a modal feature extractor, a feature mapper and a large language model; The modal feature extractor is used to process each modal signal in the power time series signal respectively to generate a corresponding feature vector; The feature mapper is used to map the feature vectors generated corresponding to each modal signal into the feature space of the text data to obtain mapping features; The large language model is used to extract text features from the text data, and fuse the text features with mapping features to obtain transformer fault diagnosis results.

8. A transformer fault diagnosis system according to claim 7, characterized in that: The modal feature extractor includes a position encoding module, a multi-layer Transformer encoder and a multi-layer Transformer decoder connected in sequence; The position encoder is used to obtain the position information of each modal signal in the power time series signal; The multi-layer Transformer encoder is used to gradually extract the multi-level features of each modal signal in the power time series signal through a self-attention mechanism and a feedforward neural network; The multi-layer Transformer decoder is used to step by step input the original input according to the position information and the extracted multi-level features to generate the feature vector corresponding to each modal signal; The feature mapper adopts a KAN network.

9. A transformer fault diagnosis system according to claim 8, characterized in that: The training method of the multimodal large model comprises: The multimodal large model is trained using positive sample pairs and negative sample pairs; the positive sample pairs include power timing signals and corresponding fault description texts, and the negative sample pairs include power timing signals and irrelevant fault description texts; The loss function used in training is the InfoNCE loss function.

10. A computer program product, characterized in that: The computer program product comprises instructions, and when the instructions are executed by a processor, a transformer fault diagnosis method according to any one of claims 1 to 7 is implemented.

Images