A method and system for processing power operation and maintenance data based on small sample alignment

Abstract

The present invention discloses a method and system for processing electric power operation and maintenance data based on small sample alignment, the method comprising: integrating and analyzing different categories of operation and maintenance small sample data collected from a target electric power operation and maintenance system to obtain an electric power operation and maintenance data set to be labeled; in the process of using the labeled electric power operation and maintenance data set to perform correction training on a voting annotation model, introducing an annotation correction factor constructed by a first preset large model into a graphic and text double-layer classifier, the annotation correction factor being designed to achieve dynamic correction of itself according to an importance difference factor output by the first preset large model; inputting the electric power operation and maintenance data set to be labeled into the voting annotation model after correction training, and annotating the electric power operation and maintenance data set to be labeled according to the output voting results; aligning the obtained annotation results to obtain electric power operation and maintenance data; the present invention can generate high-quality expert model fine-tuning enhanced data and output reliable operation and maintenance results.

Description

Technical Field

[0001] The present invention relates to the field of electric power operation and maintenance technology, and in particular to an electric power operation and maintenance data processing method and system based on small sample alignment. Background Art

[0002] With the rapid development of the artificial intelligence industry, large language models (LLMs) have made significant progress. The demand for vertical expert large models (Expert LLMs) in various fields has gradually become the primary goal of technological breakthroughs. The fine-tuning data attributes in the field of power operation and maintenance are very complex. The data is sparse and heterogeneous, and is limited by the diversity of different devices, systems, scenarios, and time. In addition, many power operation and maintenance data cannot be involved in public channels because they involve power grid security and operation secrets, which limits the acquisition and use of data and increases the difficulty of data preparation.

[0003] When constructing a complex power system operation and maintenance knowledge data set, the existing technology does not properly handle data sparsity and has difficulty processing heterogeneous data during data processing, resulting in insufficient clarity in instruction construction. As a result, the fine-tuning process of the expert model often results in output confusion, knowledge hallucinations and other problems, making it impossible to formulate accurate and effective power operation and maintenance strategies.

[0004] It can be seen that how to optimize the processing process of power operation and maintenance data and improve the stability of power system operation and maintenance has become a technical problem that technical personnel in this field need to solve urgently. Summary of the invention

[0005] The present invention provides a method and system for processing power operation and maintenance data based on small sample alignment, so as to effectively process the power operation and maintenance data used for a large model, improve the subsequent processing effect of using the large model to cope with the power system operation and maintenance knowledge, and improve the reliability of the generated maintenance strategy.

[0006] In order to solve the above technical problems, an embodiment of the present invention provides a method for processing power operation and maintenance data based on small sample alignment, comprising: Integrate and analyze small sample data of different categories collected from the target power operation and maintenance system to obtain the power operation and maintenance data set to be labeled; In the process of using the labeled power operation and maintenance data set to perform correction training on the voting annotation model, the annotation correction factor constructed based on the first preset large model is introduced into the image-text double-layer classifier of the voting annotation model, and the annotation correction factor is designed to realize dynamic correction control of itself according to the importance difference factor output by the first preset large model; Inputting the electric power operation and maintenance data set to be annotated into the voting annotation model after the deviation correction training, and annotating the electric power operation and maintenance data set to be annotated according to the output voting results; The obtained annotation results are aligned to obtain the power operation and maintenance data.

[0007] Furthermore, the step of introducing the annotation correction factor constructed based on the first preset large model into the image-text double-layer classifier of the voting annotation model includes: Using the labeled power operation and maintenance data set to train a graphic-text double-layer Bayesian classifier, and outputting a prediction result to be corrected; Performing quantitative analysis on the difference between the prediction result to be corrected and the standard prediction result outputted based on the first preset large model to obtain the marked correction factor; The prediction result to be corrected is dynamically corrected by using the marked correction factor.

[0008] Furthermore, the marking correction factor is: in, is the correction factor, It is the prediction output of the large model for the sample; The Bayesian prediction result obtained from the current round of training; is the image feature matrix; is the text feature matrix.

[0009] Furthermore, the annotation correction factor is designed to realize dynamic correction control of itself according to the importance difference factor output by the first preset large model, including: The image set part and the text set part in the labeled power operation and maintenance data set are respectively input into the first preset large model for feature analysis to obtain the corresponding importance difference factor, which is expressed by the following formula: In the formula, It is the function of the large model to process image data and text data; is the importance difference factor corresponding to the image data, is the importance difference factor corresponding to the text data; The importance difference factor is normalized using a corrected linear unit to dynamically control the correction degree of the annotation correction factor.

[0010] Furthermore, the process of the correction training includes: Sort the graphic and text sample data input into the first preset large model according to the corresponding importance difference factor, and divide the graphic and text sample data into high-order power operation and maintenance samples and low-order power operation and maintenance samples under preset conditions; Different correction training methods are configured for the high-level power operation and maintenance samples and the low-level power operation and maintenance samples; Under different deviation correction training modes, high-order features are obtained by performing high-order feature interaction based on the high-order power operation and maintenance samples, a high-order sub-model is trained according to the high-order features, and a low-order sub-model is trained based on the low-order power operation and maintenance samples; Regularization technology is introduced to adjust the weight parameters of the low-order sub-model, and combined with the high-order sub-model to collaboratively optimize the dynamic correction control.

[0011] Furthermore, after aligning the obtained annotation results to obtain the power operation and maintenance data, the method further includes: Inputting the power operation and maintenance data into a second preset large model to generate a power operation and maintenance knowledge graph to obtain large model enhanced data; The second preset large model is fine-tuned and trained using the large model enhancement data, and the operation and maintenance expert large model formed after the fine-tuning and training is used to perform power operation and maintenance analysis to obtain and execute the power operation and maintenance strategy.

[0012] Furthermore, the generation of the power operation and maintenance knowledge graph includes: Inputting the power operation and maintenance data into the second preset large model in the form of instances, and generating data enhancement samples corresponding to each instance according to the output results; Preprocessing the instance and the data augmentation sample respectively to generate corresponding feature vectors; The kernel function is used to perform similarity analysis on the feature vectors to determine target enhanced sample data so as to construct the power operation and maintenance knowledge graph.

[0013] Furthermore, the step of inputting the to-be-annotated electric power operation and maintenance data set into the voting annotation model after the deviation correction training is completed, and annotating the to-be-annotated electric power operation and maintenance data set according to the output voting result includes: Inputting the to-be-annotated power operation and maintenance data set into the trained voting annotation model and decision annotation model for voting annotation respectively, and obtaining a first voting result and a second voting result; the decision annotation model is designed to be obtained by training a third preset large model with a few-sample learning technology; Performing a voting decision on the power operation and maintenance data set to be labeled according to the first voting result and the second voting result; The first preset large model is used to correct the voting decision result to label the electric power operation and maintenance data set to be labeled.

[0014] Furthermore, after obtaining the power operation and maintenance data, the method further includes: Constructing a simulation operation model of the target power operation and maintenance system, inputting the power operation and maintenance data into the simulation operation model, and obtaining a first simulation operation result; Applying the created power operation and maintenance simulation event to the simulation operation model to obtain a second simulation operation result; The first simulation operation result and the second simulation operation result are compared and analyzed to determine a verification result of the power operation and maintenance data.

[0015] Another embodiment of the present invention provides a power operation and maintenance data processing system based on small sample alignment, comprising: The data acquisition module is used to integrate and analyze small sample data of different categories of operation and maintenance collected from the target power operation and maintenance system to obtain the power operation and maintenance data set to be labeled; A correction control module, used to introduce a label correction factor constructed based on a first preset large model into a picture-text double-layer classifier of the voting labeling model during the process of performing correction training on the voting labeling model using the labeled power operation and maintenance data set, wherein the label correction factor is designed to realize dynamic correction control of itself according to the importance difference factor output by the first preset large model; A data annotation module, used for inputting the electric power operation and maintenance data set to be annotated into the voting annotation model after the deviation correction training is completed, and annotating the electric power operation and maintenance data set to be annotated according to the output voting result; The data enhancement module is used to align the obtained annotation results to obtain power operation and maintenance data.

[0016] Compared with the prior art, the embodiments of the present invention have the following advantages: (1) The embodiment of the present invention adopts small sample alignment technology and power operation and maintenance knowledge graph to hierarchically integrate expert-level power system operation and maintenance knowledge data, optimize it step by step, and effectively construct the power operation and maintenance knowledge graph with small samples as the core to construct high-quality large model training data, providing a solid foundation for subsequent vertical domain expert fine-tuning.

[0017] (2) Combined with the double-layer Bayesian classification voting annotator based on power space multimodality and large model correction, it can automatically annotate data in small sample and heterogeneous data environments and perform efficient structured alignment, which greatly improves the efficiency of data processing and provides reliable support for further analysis of power operation and maintenance data and construction of knowledge graphs. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flowchart of a method for processing power operation and maintenance data based on small sample alignment in one embodiment of the present invention; Figure 2It is a schematic diagram of a structured flow of power operation and maintenance data based on small sample alignment in one embodiment of the present invention; Figure 3 It is a schematic diagram of the construction process of the power operation and maintenance knowledge graph in one embodiment of the present invention; Figure 4 It is a specific implementation example diagram of constructing a data set based on the Align-PKG technology in one embodiment of the present invention; Figure 5 This is an example diagram of a power operation and maintenance knowledge graph in one embodiment of the present invention; Figure 6 It is a structural schematic diagram of an electric power operation and maintenance data processing system based on small sample alignment in one embodiment of the present invention. DETAILED DESCRIPTION

[0019] 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. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] In the description of this application, the terms "first", "second", "third", etc. are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first", "second", "third", etc. may explicitly or implicitly include one or more of the feature. In the description of this application, unless otherwise specified, "plurality" means two or more.

[0021] In the description of the present application, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" 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 a connection between the two elements. The terms "vertical", "horizontal", "left", "right", "upper", "lower" and similar expressions used herein are only for illustrative purposes, and do 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 invention. The term "and / or" used herein includes any and all combinations of one or more related listed items. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0022] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in the present invention have the same meaning as those commonly understood by those skilled in the art. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood by specific circumstances.

[0023] Faced with the complex attributes of data in the field of power operation and maintenance, which are sparse and highly heterogeneous, and are limited by the diverse characteristics of different devices, systems, scenarios and time, the present invention proposes a method called Align-PKG (Alignment-Poweroperating Knowledge Graph), which uses small sample alignment technology and power operation and maintenance knowledge graph to hierarchically integrate expert-level power system operation and maintenance knowledge data, optimize step by step, and construct high-quality large model training data with small samples as the core, providing a solid foundation for subsequent vertical domain expert fine-tuning. Based on this, an embodiment of the present invention provides a method for processing power operation and maintenance data based on small sample alignment. For details, please refer to Figure 1 , Figure 1 The figure shows a flow chart of a method for processing power operation and maintenance data based on small sample alignment in one embodiment of the present invention, which includes the following steps: S1. Integrate and analyze small sample data of different categories collected from the target power operation and maintenance system to obtain the power operation and maintenance data set to be labeled.

[0024] It is understandable that the field of power information operation and maintenance needs to face various complex scenarios and redundant content, involving different types of data such as equipment fault diagnosis process, maintenance standard requirements, resource management methods, and scheduling measures requirements. Based on this, the embodiment of the present invention sets standard small sample data from three types of representative data, including "equipment monitoring records", "fault reports" and "scheduling documents". For details, please refer to Figure 2 , Figure 2 Shown is a schematic diagram of the structuring process of power operation and maintenance data based on small sample alignment in one embodiment of the present invention.

[0025] As an example, this embodiment represents the device monitoring record as , the fault report is expressed as , the scheduling document is represented as .

[0026] For each type of data, a standard sample is selected as the basic example. It is worth noting that the samples of equipment monitoring record type need to cover the operating status and historical data of power equipment; the samples of fault report type need to record detailed information of equipment faults and the fault handling process; the samples of scheduling document type need to include power scheduling arrangements and execution status. After integrating these data, we get the power operation and maintenance data set to be labeled. These data contain a picture set of relevant professional images and texts.

[0027] It is worth noting that in order to realize model training, it is necessary to build a labeled power operation and maintenance data set in advance.

[0028] S2. In the process of correcting the voting annotation model using the labeled power operation and maintenance data set, the annotation correction factor constructed based on the first preset large model is introduced into the image-text double-layer classifier of the voting annotation model. The annotation correction factor is designed to achieve dynamic correction control of itself according to the importance difference factor output by the first preset large model.

[0029] The embodiment of the present invention considers integrated learning technology and adopts voting to automatically compare and annotate the power operation and maintenance data set to be annotated.

[0030] In some embodiments of the present invention, two different types of voting taggers are introduced to perform voting determination, including a Bayesian classification voting tagger and a visual language large model voting tagger.

[0031] Among them, the embodiment of the present invention trains a two-layer Bayesian classifier of images and texts by using the image set and text set in the labeled power operation and maintenance data set to construct a voting labeling model, and gives classification results under the three labels of "equipment monitoring records", "fault reports" and "scheduling documents". Through the large model correction factor control item, adaptive adjustment of the Bayesian classification voting labeler is achieved.

[0032] Specifically, in some embodiments of the present invention, the labeled power operation and maintenance data set is divided into an image set part and a text set part. The image data is first loaded and preprocessed. The ordinary image is converted into a grayscale image, and then the median filter method is used to remove the noise in the image. Then the text data is loaded and vectorized using a large language model to obtain the processed image data. and text data As an example, the bge-large-zh-v1.5 large model can be used to vectorize text data.

[0033] Getting image data and text data Afterwards, this embodiment trains the image-text double-layer Bayesian classifier through these data, and outputs the prediction results to be corrected. Specifically, The embodiment of the present invention uses the following formula to train the image layer Bayesian classifier: in, is the image feature In category The conditional probability under ; is the category prior probability; is the marginal probability of the image feature.

[0034] The conditional probability It is expressed by the following formula: in, is the first feature vector of the image Features is the total dimension of the image feature vector.

[0035] Furthermore, the text layer Bayesian classifier is trained using the following formula: Text features In category The conditional probability under ; is the class prior probability; is the marginal probability of the text feature.

[0036] For the conditional probability The calculations are: in, is the first Features is the total dimension of the text feature vector.

[0037] According to the above formula, after completing one round of training, the auxiliary correction of the large model is introduced according to the results obtained from the training set. In the embodiment of the present invention, the prediction result output by the Bayesian classifier, i.e., the prediction result to be corrected, is quantitatively analyzed for differences with the standard prediction result output by the first preset large model, so as to obtain the labeled correction factor. The labeled correction factor is expressed as follows: in, To mark the correction factor, The prediction output of the first preset large model for the sample; The Bayesian prediction result obtained from the current round of training; is the image feature matrix; is the text feature matrix.

[0038] Understandable, marking the correction factor Indicates the difference between the big model prediction and the Bayesian prediction, and is used to correct the Bayesian result. This embodiment introduces a big model correction after each round of training, dynamically adjusts the weight and output of the Bayesian classifier, and can gradually reduce the training error of the classifier and get closer to the actual sample distribution. It is worth noting that the first preset big model selected in this embodiment is a multimodal big model. Preferably, the MiniCPM-V 2.6 end-side multimodal big model is selected.

[0039] In order to achieve dynamic control of the annotation correction factor and improve the accuracy and robustness of the double-layer Bayesian model, the embodiment of the present invention constructs an importance difference factor for controlling the data enhancement ratio and adjusting the weights in model training by using a large model. Specifically: First, the preprocessed image set part and the text set part in the labeled power operation and maintenance data set are respectively input into the first preset large model for feature analysis to obtain the corresponding importance difference factors.

[0040] Traverse the image data separately and text data The individual data content (image and text data are calculated and sorted separately) is in the interval In which It means it is not important. Indicates importance.

[0041] Assume the image data is , the text data is ; The two are respectively passed through the feature extraction module of the multimodal large model to output their respective importance difference values, namely the importance difference factors.

[0042] Specifically, the importance difference factor is expressed by the following formula: In the formula, It is the function of the large model to process image data and text data, that is, after using Prompt to declare the "above scoring rules", the multimodal large model is used to directly output and ; is the importance difference factor corresponding to the image data, is the importance difference factor corresponding to the text data.

[0043] Next, the importance difference factor is normalized using a rectified linear unit to dynamically control the degree of correction of the labeled correction factor. By normalizing the importance difference factor using a rectified linear unit (ReLU), a dynamic adjustment parameter can be obtained, which is used to control the degree of influence of the correction factor on the next round of prediction results.

[0044] Specifically, the dynamic adjustment parameters are expressed as: in, To dynamically adjust the parameters, is the importance difference value corresponding to the sample.

[0045] Understandably, the next round of prediction results after correction It is expressed by the following formula: The training and correction process is repeated based on the above formula until the model converges. It is worth noting that in the process of controlling the annotation correction factor using the importance difference factor, the embodiment of the present invention considers the different characteristics of each sample data used to train the classifier and considers configuring different training strategies for the sample data.

[0046] Specifically, the graphic sample data input into the first preset large model are sorted according to the corresponding importance difference factors, that is, the samples are grouped, and the graphic sample data are divided into high-order power operation and maintenance samples and low-order power operation and maintenance samples under preset conditions. It is worth noting that high order and low order reflect the value of the samples. The purpose of dividing the samples in this way is to help the model pay more attention to high-value samples during the training process. The samples obtained after grouping are expressed as follows: As an example, in this embodiment, the first 33% of samples are high-value power operation and maintenance data. , the last 67% of the samples are low-value power operation and maintenance data .

[0047] After determining the value of the sample data, a scenario-based special group training configuration is set, that is, different correction training methods are configured for the high-order power operation and maintenance samples and the low-order power operation and maintenance samples. As an example, the high-order power operation and maintenance samples can be set to perform E rounds of training, and the configuration is: For a single round of training of low-level power operation and maintenance samples, the configuration is: In the above-mentioned different correction training modes, high-order features are obtained by performing high-order feature interaction based on the high-order power operation and maintenance samples, and high-order sub-models are trained according to the high-order features.

[0048] In some embodiments of the present invention, the most critical and representative subsets can be screened from the high-value samples. In-depth feature engineering is performed on these samples to extract high-order features to capture complex patterns and correlations in the data. These high-order features are used to train a dedicated sub-model for the correction task of handling high-value samples.

[0049] A low-order sub-model is trained based on the low-order power operation and maintenance samples, and regularization technology is introduced to adjust the weight parameters of the low-order sub-model. Finally, the high-order sub-model is combined to collaboratively optimize the process of dynamic correction control of the marked correction factor.

[0050] It can be understood that high-order feature interactions can capture the complex correlations and dependencies in high-order samples. In some embodiments of the present invention, deep learning techniques, such as deep neural networks or graph neural networks, are used to model high-order feature interactions. For low-order samples, a relatively simple model can be used for training to reduce computational complexity and overfitting risks. The embodiments of the present invention combine high-order and low-order deviation correction control methods to make full use of key information in high-value samples and diversity in low-value samples.

[0051] It is worth noting that after training the image layer Bayesian classifier and the text layer Bayesian classifier, the two classifiers need to be integrated to obtain a dual-layer Bayesian classifier for making comprehensive decisions on unlabeled power operation and maintenance data. Specifically, the output results of the image Bayesian classifier and the text Bayesian classifier are combined to calculate the final classification probability, which is expressed as follows: in, is the output of the image Bayesian classifier; is the output of the text Bayesian classifier; , is a dynamic weight, satisfying .

[0052] It is understandable that if a piece of data only has text content, or only has image content, the other dynamic weight is reset to 0, and the above integrated classification probability is used to make a labeling decision for the power operation and maintenance data set to be labeled.

[0053] In this embodiment, the dynamic weight The adjustment process is designed to be based on image characteristics and text features The information entropy contribution is dynamically adjusted. Specifically, the weight dynamic adjustment process can be expressed by the following formula: in, is the total information entropy of image features, is the total information entropy of text features.

[0054] The dynamic weight mechanism designed in this embodiment enables image or text features to contribute more to the final decision when the amount of information is high, thereby achieving better multimodal fusion.

[0055] S3: input the electric power operation and maintenance data set to be labeled into the voting and labeling model after the deviation correction training, and label the electric power operation and maintenance data set to be labeled according to the output voting result.

[0056] In this embodiment, the power operation and maintenance data set to be labeled is input into the trained voting labeling model and decision labeling model for voting labeling, respectively, to obtain a first voting result and a second voting result, and a voting decision is made on the power operation and maintenance data set to be labeled according to the first voting result and the second voting result. It is worth noting that the decision labeling model is designed to be obtained by training a third preset large model with a few-sample learning technology.

[0057] It can be understood that the embodiment of the present invention introduces a visual language large model voting annotator, that is, a trained decision annotation model is combined with a voting annotation model to achieve voting decision. Preferably, the Qwen-VL-Chat visual language large model is selected as the third preset large model, and a small number of samples (Few-shot) technology is used for voting.

[0058] Specifically, based on the small sample data of different categories of operation and maintenance ( , , ), extract the voting scores of each class of samples from the first voting result and the second voting result to calculate the scores of the corresponding class of samples, and determine the target voting result according to the scores. The calculation process of the sample score is expressed by the following formula: For each category, the classifier with the highest average score is selected as the final judgment category, which is expressed as follows: According to the above formula, output the voting results And the corresponding score results: .

[0059] In the voting decision process, the embodiment of the present invention uses the first preset large model to correct the voting decision result to achieve the labeling of the power operation and maintenance data set to be labeled, which is specifically expressed by the following formula: in, A preliminary decision for the omnibus vote; For large models , The correction result; It is an adjustment parameter to control the influence of large model correction.

[0060] Here are some examples of the output annotation results: if The "Equipment Monitoring Record" label will be scored 3 points, and the others will be scored 0 points; if The "Fault Report" label will be scored 3 points, and the others will be scored 0 points; if The "Scheduling Document" label will be scored 3 points, and the others will be scored 0 points.

[0061] S4. Align the obtained annotation results to obtain power operation and maintenance data.

[0062] like Figure 2 As shown, it can be seen that in order to further overcome the challenges brought by the sparsity and heterogeneity of operation and maintenance data, this embodiment performs structured alignment on the labeled power operation and maintenance data to obtain the initial power operation and maintenance data for expert fine-tuning.

[0063] It is worth noting that in order to ensure the reliability of the enhanced data, in some embodiments of the present invention, before the power operation and maintenance data is input into the second preset large model, the accuracy and validity of the power operation and maintenance data are pre-verified by simulation technology. Specifically, the following steps are included: First, based on advanced simulation technology, such as power system simulation software or digital twin technology, a simulation operation model of the target power operation and maintenance system is constructed, and the power operation and maintenance data is input into the simulation operation model to obtain the first simulation operation result. This step can simulate the actual operation status of the power operation and maintenance system and reveal the behavioral characteristics of the system under normal and abnormal conditions.

[0064] Then, the created power operation and maintenance simulation event is applied to the simulation operation model to observe the response and change of the system to obtain the second simulation operation result. The simulation event may include operation and maintenance scenarios such as equipment failure, load mutation, and grid structure change.

[0065] Finally, the first simulation operation result and the second simulation operation result are compared and analyzed to determine the verification result of the power operation and maintenance data, the verification result is fed back, and the data is corrected as needed.

[0066] Furthermore, in order to construct high-quality fine-tuning data for the expert large model, the embodiment of the present invention inputs the power operation and maintenance data into the second preset large model to generate a power operation and maintenance knowledge graph. Figure 3 , Figure 3 Shown is a schematic diagram of the construction process of the power operation and maintenance knowledge graph in one embodiment of the present invention.

[0067] In some embodiments of the present invention, the Qwen2.5-14B-Instruct large model is selected as the second preset large model, an electric power operation and maintenance knowledge graph with "aligned sample characteristics" is generated, and the kernel density estimation method is used for data processing to obtain large model enhanced data.

[0068] Specifically, this embodiment inputs the power operation and maintenance data into the second preset large model in the form of instances, generates data enhancement samples corresponding to each instance according to the output results, preprocesses the instances and the data enhancement samples respectively, and generates corresponding feature vectors.

[0069] In some embodiments of the present invention, each data in the aligned data sample can be traversed and selected as (instance), and perform embedding operation on the instance, using feature representation based on TF-IDF (a text feature extraction method), setting a text The eigenvector of: in, For Text Middle The TF-IDF value of the word, is the dictionary size.

[0070] Since high-dimensional feature kernel density estimation (KDE) can estimate the probability density function for data points in high-dimensional space, it is suitable for text similarity measurement in complex power operation and maintenance scenarios. Based on this, the embodiment of the present invention uses the kernel function to perform similarity analysis on the feature vectors, determines the target enhanced sample data, and constructs the power operation and maintenance knowledge graph. The construction process of the power operation and maintenance knowledge graph is described in detail below: First, let the reference text be And the enhanced result is The eigenvector of and , processed using kernel density estimation, is expressed as follows: in, It's on point The estimated probability density function value at ; is the number of samples, i.e. the dimension of the feature vector; is the bandwidth parameter, controlling the smoothness; is the kernel function Text and The eigenvector of and Substituting into the kernel density estimation formula, we get their probability density estimates, which are expressed by the following formula: Among them, the integral represents the similarity in all feature spaces, that is, and The density estimation overlap is calculated to measure the degree of overlap of texts in the high-dimensional feature space.

[0071] Next, based on the above formula, the most similar and least similar enhanced samples to the original text instance are selected. The similarity calculation process is expressed as follows: It is understandable that the most similar samples highlight the authenticity of the large model simulation data, while the least similar samples highlight the sample flexibility and robustness. Based on this, this embodiment preferably selects the samples with "highest similarity" and "lowest similarity" (a total of 2) as the output results.

[0072] Finally, the second preset large model is fine-tuned and trained using the large model enhancement data, and the power operation and maintenance analysis is performed through the large operation and maintenance expert model formed after fine-tuning and training, and the corresponding power operation and maintenance strategy is obtained and executed. Specifically, the present invention provides an example of a fixed alignment special sample, as shown below: { "Sample ID":"001", "Category":"Equipment Monitoring Records", / / Sample category - equipment monitoring records, fault reports, scheduling documents "Original text":"The equipment is running normally and no fault warning is found.", "Annotation result": { "Status": "Normal", / / The running status of the marked device "Fault Type": null, / / Fault type, if there is no fault, it is null "Risk Level": "Low", / / Risk level, low / medium / high "Operation suggestion": "Continue to monitor the operation of the device" / / Operation suggestion based on the annotation "Image reference": [image link] }, "Feature vector": [0.23, 0.45, 0.67, 0.12, 0.89], / / Feature vector of text "Voting Result": [R1, R2], "Alignment Information": { "Standard sample": "Equipment monitoring record sample 001", / / Standard sample ID aligned with this sample "Similarity": 0.87, / / Similarity between the sample and the standard sample }, "Knowledge Graph": { "Entity": [ { "Entity ID":"Device001", "Type":"Device", "Relationship":"Running Status", "Value": "Normal" } ], "relation": [ { "Source Entity ID":"Device001", "Target Entity ID":"Running Status", "Relationship Type": "Status" } ] } } The following is an explanation of some of the characteristic words in the above content: Sample ID: unique identifier for each sample; Category: The category to which the sample data belongs, including "equipment monitoring record", "fault report" or "dispatch document"; Labeling results: including labeling information such as sample classification, fault type, risk level, and operation suggestions; Feature vector: the feature vector converted from the sample text, used for model processing; Alignment information: the ID of the standard sample aligned with the sample, the similarity between the sample and the standard sample, the data source, and other information.

[0073] It is worth noting that after fine-tuning and training, the expert language model can generate corresponding strategy recommendations based on the input operation and maintenance scenarios and problems. These recommendations can include fault prediction, maintenance guidance, equipment management and other aspects.

[0074] In order to give full play to the advantages of the above contents of the present invention, the above embodiments of the present invention will be analyzed in detail below. Specifically, please refer to Figure 4 , Figure 4 A specific implementation example diagram of constructing a data set based on the Align-PKG technology in one of the embodiments of the present invention is shown. It can be seen that this example sets up a data processing scenario for the power operation and maintenance expert model to achieve data structuring and effective extraction of knowledge.

[0075] The specific steps are as follows: S1: Data collection: In this stage, Align-PKG first collects multiple data types from the power operation and maintenance system, collects the operating status parameters of the equipment in real time; records the historical fault conditions of the equipment, including the time of occurrence of the fault, the type of fault, and the treatment measures; and extracts log information such as equipment maintenance history, maintenance personnel, and maintenance content. Data quality is ensured through data preprocessing, and word segmentation, stop word removal, and normalization processing techniques are used to assist in data cleaning to ensure data accuracy and consistency and generate structured feature vectors. Representative samples are extracted from different data sources to ensure coverage of different equipment and fault types. Specifically, please refer to the data content shown in the table below.

[0076] Power operation and maintenance data type collection table S2: Constructing a knowledge graph In this phase, Align-PKG constructs a knowledge graph for power operation and maintenance, extracts key entities such as equipment A, equipment B, fault type C, etc., and forms basic knowledge triples; identifies the relationship between different entities, such as "equipment A and equipment B both belong to power transformers", and records the logical relationship. Use Neo4j to store the above triples as a knowledge graph for subsequent query and reasoning. For details, please refer to Figure 5 , Figure 5 Shown is an example diagram of a power operation and maintenance knowledge graph in one embodiment of the present invention.

[0077] S3: Knowledge Representation Enhancement of Expert Model Fine-tuning Data In this phase, Align-PKG combines small sample alignment technology to match structured data in the knowledge graph with expert knowledge to ensure the consistency and logic of knowledge; for each sample, define feature vectors and establish an alignment relationship matrix between samples. Define reasoning rules, for example: "If device A has fault type B, maintenance record C may be required"; collect feedback information from the model in actual application and analyze whether the model's decision is in line with expectations. Adjust reasoning rules and knowledge graphs based on feedback to continuously optimize the model's learning ability and knowledge representation ability.

[0078] S4: Self-optimization During this phase, Align-PKG will continue to optimize. Based on newly collected data and user feedback, the knowledge graph will be regularly updated to ensure the timeliness and accuracy of the model knowledge. By comparing and analyzing the performance of data in different fine-tuning scenarios, the strategic recommendations for the fine-tuning power operation and maintenance expert LLM will be adjusted to ensure that it can adapt to various complex situations in power operation and maintenance. A monitoring mechanism will be implemented to track the performance of data in actual fine-tuning applications to ensure its continuous improvement in quality.

[0079] In summary, the embodiment of the present invention can better extract detailed features in power operation and maintenance data by constructing a double-layer Bayesian classification voting annotator based on multimodal large model correction, and control the correction process by designing importance difference factors. This dynamic adjustment system can improve the accuracy of the classifier; and combined with the Qwen-VL-Chat strong annotator, an effective data annotation and alignment mechanism is formed, which can automatically annotate data in small sample and heterogeneous data environments, and perform efficient structured alignment, greatly improving the efficiency of data processing; by aligning structured and large models to build a power operation and maintenance knowledge graph, it can provide high-quality expert fine-tuning enhanced data, so that the expert model can output reliable and effective operation and maintenance results, thereby improving the stability of power operation and maintenance.

[0080] An embodiment of the present invention provides a power operation and maintenance data processing system based on small sample alignment. For details, see Figure 6 , Figure 6 The figure shows a flow chart of a power operation and maintenance data processing system based on small sample alignment in one embodiment of the present invention, including: The data acquisition module M1 is used to integrate and analyze small sample data of different categories of operation and maintenance collected from the target power operation and maintenance system to obtain the power operation and maintenance data set to be labeled; A correction control module M2 is used to introduce a label correction factor constructed based on a first preset large model into a picture-text double-layer classifier of the voting labeling model during the process of performing correction training on the voting labeling model using the labeled power operation and maintenance data set, wherein the label correction factor is designed to realize dynamic correction control of itself according to the importance difference factor output by the first preset large model; A data annotation module M3 is used to input the electric power operation and maintenance data set to be annotated into the voting annotation model that has completed the deviation correction training, and annotate the electric power operation and maintenance data set to be annotated according to the output voting result; The data enhancement module M4 is used to align the obtained annotation results to obtain power operation and maintenance data.

[0081] The technical features and technical effects of the power operation and maintenance data processing system based on small sample alignment proposed in the embodiment of the present invention are the same as the technical features and technical effects of the power operation and maintenance data processing method based on small sample alignment proposed in the embodiment of the present invention, and will not be elaborated here.

[0082] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for processing power operation and maintenance data based on small sample alignment, characterized in that: include: Integrate and analyze small sample data of different categories collected from the target power operation and maintenance system to obtain the power operation and maintenance data set to be labeled; In the process of using the labeled power operation and maintenance data set to perform correction training on the voting annotation model, the annotation correction factor constructed based on the first preset large model is introduced into the image-text double-layer classifier of the voting annotation model, and the annotation correction factor is designed to realize dynamic correction control of itself according to the importance difference factor output by the first preset large model; Inputting the electric power operation and maintenance data set to be annotated into the voting annotation model after the deviation correction training, and annotating the electric power operation and maintenance data set to be annotated according to the output voting results; The obtained annotation results are aligned to obtain the power operation and maintenance data.

2. The power operation and maintenance data processing method based on small sample alignment according to claim 1, characterized in that: The method of introducing the annotation correction factor constructed based on the first preset large model into the image-text double-layer classifier of the voting annotation model includes: Using the labeled power operation and maintenance data set to train a graphic-text double-layer Bayesian classifier, and outputting a prediction result to be corrected; Performing quantitative analysis on the difference between the prediction result to be corrected and the standard prediction result outputted based on the first preset large model to obtain the marked correction factor; The prediction result to be corrected is dynamically corrected by using the marked correction factor.

3. The power operation and maintenance data processing method based on small sample alignment according to claim 1, characterized in that: The marking correction factor is: in, is the correction factor, It is the prediction output of the large model for the sample; The Bayesian prediction result obtained from the current round of training; is the image feature matrix; is the text feature matrix.

4. The power operation and maintenance data processing method based on small sample alignment according to claim 1, characterized in that: The annotation correction factor is designed to realize dynamic correction control of itself according to the importance difference factor output by the first preset large model, including: The image set part and the text set part in the labeled power operation and maintenance data set are respectively input into the first preset large model for feature analysis to obtain the corresponding importance difference factor, which is expressed by the following formula: In the formula, It is the function of the large model to process image data and text data; is the importance difference factor corresponding to the image data, is the importance difference factor corresponding to the text data; The importance difference factor is normalized using a corrected linear unit to dynamically control the correction degree of the annotation correction factor.

5. The power operation and maintenance data processing method based on small sample alignment according to claim 1, characterized in that: The process of the correction training includes: Sort the graphic and text sample data input into the first preset large model according to the corresponding importance difference factor, and divide the graphic and text sample data into high-order power operation and maintenance samples and low-order power operation and maintenance samples under preset conditions; Different correction training methods are configured for the high-level power operation and maintenance samples and the low-level power operation and maintenance samples; Under different deviation correction training modes, high-order features are obtained by performing high-order feature interaction based on the high-order power operation and maintenance samples, a high-order sub-model is trained according to the high-order features, and a low-order sub-model is trained based on the low-order power operation and maintenance samples; Regularization technology is introduced to adjust the weight parameters of the low-order sub-model, and combined with the high-order sub-model to collaboratively optimize the dynamic correction control.

6. The power operation and maintenance data processing method based on small sample alignment according to claim 1, characterized in that: After aligning the obtained annotation results to obtain the power operation and maintenance data, the method further includes: Inputting the power operation and maintenance data into a second preset large model to generate a power operation and maintenance knowledge graph to obtain large model enhanced data; The second preset large model is fine-tuned and trained using the large model enhancement data, and the operation and maintenance expert large model formed after the fine-tuning and training is used to perform power operation and maintenance analysis to obtain and execute the power operation and maintenance strategy.

7. The power operation and maintenance data processing method based on small sample alignment according to claim 6, characterized in that: The generation of the power operation and maintenance knowledge graph includes: Inputting the power operation and maintenance data into the second preset large model in the form of instances, and generating data enhancement samples corresponding to each instance according to the output results; Preprocessing the instance and the data augmentation sample respectively to generate corresponding feature vectors; The kernel function is used to perform similarity analysis on the feature vectors to determine target enhanced sample data so as to construct the power operation and maintenance knowledge graph.

8. The power operation and maintenance data processing method based on small sample alignment according to claim 1, characterized in that: The step of inputting the to-be-annotated electric power operation and maintenance data set into the voting and annotation model after the deviation correction training is completed, and annotating the to-be-annotated electric power operation and maintenance data set according to the output voting result, includes: Inputting the to-be-annotated power operation and maintenance data set into the trained voting annotation model and decision annotation model for voting annotation respectively, and obtaining a first voting result and a second voting result; the decision annotation model is designed to be obtained by training a third preset large model with a few-sample learning technology; Performing a voting decision on the power operation and maintenance data set to be labeled according to the first voting result and the second voting result; The first preset large model is used to correct the voting decision result to label the electric power operation and maintenance data set to be labeled.

9. The power operation and maintenance data processing method based on small sample alignment according to claim 6, characterized in that: After obtaining the power operation and maintenance data, the method further includes: Constructing a simulation operation model of the target power operation and maintenance system, inputting the power operation and maintenance data into the simulation operation model, and obtaining a first simulation operation result; Applying the created power operation and maintenance simulation event to the simulation operation model to obtain a second simulation operation result; The first simulation operation result and the second simulation operation result are compared and analyzed to determine a verification result of the power operation and maintenance data.

10. A power operation and maintenance data processing system based on small sample alignment, characterized in that: include: The data acquisition module is used to integrate and analyze small sample data of different categories of operation and maintenance collected from the target power operation and maintenance system to obtain the power operation and maintenance data set to be labeled; A correction control module, used to introduce a label correction factor constructed based on a first preset large model into a picture-text double-layer classifier of the voting labeling model during the process of performing correction training on the voting labeling model using the labeled power operation and maintenance data set, wherein the label correction factor is designed to realize dynamic correction control of itself according to the importance difference factor output by the first preset large model; A data annotation module, used for inputting the electric power operation and maintenance data set to be annotated into the voting annotation model after the deviation correction training is completed, and annotating the electric power operation and maintenance data set to be annotated according to the output voting result; The data enhancement module is used to align the obtained annotation results to obtain power operation and maintenance data.

Images