A method and system for classifying power label knowledge based on large model

Abstract

The present invention discloses a method and system for classifying electric power label knowledge based on a large model, and relates to the field of data processing. The method includes: collecting electric power source data and determining effective source data; determining a hierarchical label map and supervising the training of an open source large model; introducing an unlabeled standard semantic library to determine the initialization classification model; reading downstream task features, performing supervised adjustments, and determining a large classification model; receiving a pre-classification task, making a classification decision, and determining a classification result; if the classification result of the pre-classification task is vacant, receiving a newly added defect type and performing class incremental learning. This method solves the technical problem in the prior art that electric power knowledge labels are highly specialized, resulting in scarce label category data and poor interpretability, and realizes the introduction of domain knowledge to enhance the interpretability of the model, achieves label classification of standard knowledge, and is used to predict and infer new sample data.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and specifically to a large-model-based power label knowledge classification method and system. Background Art

[0002] The classification method of power label knowledge based on large models has shown great potential and advantages in the application of the power industry. This method mainly relies on the deep learning and processing capabilities of large models on massive data to achieve effective classification of power label knowledge. In order to quickly obtain the required information from massive information, it is necessary to classify the information and label the corresponding classification labels. It is usually represented by key features that are highly relevant to the information and easy to identify, so that users can retrieve and filter. Although the power label knowledge classification in the prior art has the ability to handle large amounts of data and complex relationships, there are still some potential defects. Label knowledge in the power field is often highly professional and specific, resulting in very scarce data for certain labels or categories, especially power standard knowledge. Large models may encounter difficulties in processing such sparse data because they require a large number of samples to learn effective representations. Large models, especially deep learning models, usually have complex internal structures and parameters, which makes their decision-making process difficult to explain. This is particularly important in the power field because it is necessary to understand how the model makes classification decisions in order to trust and adopt these decisions in practical applications.

[0003] In summary, the existing technology has technical problems that power knowledge labels are highly professional, resulting in scarce label category data and poor interpretability. Summary of the invention

[0004] Based on this, it is necessary to provide a large-model-based power label knowledge classification method and system that can introduce domain knowledge to enhance the interpretability of the model, realize label classification of standard knowledge, and be used to predict and infer new sample data in response to the above technical problems.

[0005] In the first aspect, a method for classifying power label knowledge based on a large model is provided, the method comprising: collecting power source data and preprocessing it to determine effective source data, wherein the preprocessing mode is defect type screening and class parallel processing; classifying and determining a hierarchical label map based on the effective source data, and supervising the training of an open source large model based on the hierarchical label map and the effective source data, wherein the training standard comprises a classification dimension and a thinking chain deduction dimension, wherein the thinking chain deduction dimension comprises a thinking chain mode, and the thinking chain mode is an integrated visualization mode of the thinking chain; introducing an unlabeled standard semantic library, performing standardized training on the open source large model, and determining an initialization classification model; reading downstream task features, performing supervised adjustment on the initialization classification model, and determining a classification large model; receiving a pre-classification task, making a classification decision in combination with the classification large model, and determining a classification result, wherein the classification result comprises a belonging class and a thinking chain.

[0006] Determine data quality standards, identify the power source data, extract data based on defect types, determine multiple data classes, and there is data overlap in the multiple data classes; perform parallel processing on the multiple data classes to determine multiple preprocessing results, wherein data parallel processing is performed based on a preprocessing module, and the preprocessing module includes multiple parallel processing branches trained based on a difference preprocessing algorithm; perform homologous fitting and regularization on the multiple preprocessing results to determine the effective source data.

[0007] Identify the valid source data, perform data clustering based on the power label class, and determine N data sets; traverse the N data sets, extract the first data set and perform class label feature extraction and hierarchical association integration to determine a type of knowledge graph; complete the feature graph construction of the Nth data set and determine N types of knowledge graphs; perform hierarchical association integration on the one type of knowledge graph up to the N types of knowledge graphs to determine the hierarchical label graph.

[0008] Based on class relevance, the nodes of the one category of knowledge graphs up to the N categories of knowledge graphs are associated between the graphs to determine an aggregate knowledge graph, wherein the aggregate knowledge graph includes at least one; based on likelihood probability, a learning rate label is performed on the aggregate knowledge graph, wherein the likelihood probability is negatively correlated with the learning rate; based on the degree of feature personalization, invalid graph nodes are removed from the aggregate knowledge graph; and the aggregate knowledge graph after the learning rate label and the invalid graph nodes are removed is fitted to determine the hierarchical label graph.

[0009] Integrate batch training data and randomly screen test samples; input the test samples into the classification model for convergence verification, and determine the loss amount based on the sample verification results; if the loss amount is greater than the loss threshold, combine the cross entropy loss function to optimize the classification model.

[0010] Determine a decision map label, wherein there are multiple decision map labels; based on the learning rate, perform weight distribution and overall analysis on the decision map labels to determine the belonging category; based on the thinking chain model, screen effective derivation features, perform pattern integration on the derivation steps, and determine the thinking chain; use the belonging category and the thinking chain as the classification result.

[0011] If the classification result of the pre-classification task is blank, receive a newly added defect type, and execute class incremental learning of the classification model; wherein, executing class incremental learning of the classification model includes: if the pre-classification task is a newly added defect type, perform standardized prediction on the task information and determine the prediction sample; perform class incremental learning on the classification model in combination with the prediction sample to determine the pre-learning large model; perform balance rate and overfitting verification on the pre-learning large model to determine the verification result, wherein the balance rate is the balance ratio of the classification effect before and after class incremental learning; identify the verification result, and perform loss compensation on the pre-learning large model.

[0012] In the second aspect, a large-model-based power label knowledge classification system is provided, the system comprising: an effective source data determination module, the effective source data determination module is used to collect power source data and preprocess it to determine the effective source data, wherein the preprocessing mode is defect type screening and class parallel processing; an open source large model training module, the open source large model training module is used to classify and determine a hierarchical label map based on the effective source data, and supervise the training of the open source large model based on the hierarchical label map and the effective source data, wherein the training standard includes a classification dimension and a thought chain deduction dimension, and the thought chain deduction dimension includes A thinking chain mode, which is an integrated visualization mode of the thinking chain; a standardized learning module, which is used to introduce an unlabeled standard semantic library, perform standardized training on the open source large model, and determine the initialization classification model; a classification large model determination module, which is used to read downstream task features, perform supervised adjustment on the initialization classification model, and determine the classification large model; a classification result determination module, which is used to receive pre-classification tasks, make classification decisions in combination with the classification large model, and determine the classification results, wherein the classification results include belonging categories and thinking chains. The system also includes: a data extraction module, which is used to determine data quality standards, identify the power source data, extract data based on defect types, and determine multiple data classes, where there is data overlap in the multiple data classes; a parallel processing module, which is used to perform parallel processing on the multiple data classes and determine multiple preprocessing results, wherein the preprocessing module performs data parallel processing, and the preprocessing module includes multiple parallel processing branches trained based on a difference preprocessing algorithm; a homologous fitting and regularization module, which is used to perform homologous fitting and regularization on the multiple preprocessing results to determine the effective source data.

[0013] According to a third aspect, a computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps according to the first aspect when executing the computer program.

[0014] According to a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps described in the first aspect are implemented.

[0015] The above-mentioned large-model-based power label knowledge classification method and system builds a hierarchical label map through power source data to effectively classify, integrate and directly visualize label information. The hierarchical label map is used as the underlying structure, and the open source large model is supervised and trained in combination with the effective source data, including training based on classification dimensions and thinking chain deduction dimensions. Processing is performed based on the training model, which can effectively improve the objectivity of classification and the accuracy of classification results. The unlabeled standard semantic library is introduced to perform standardized training on the open source large model, and the downstream task features are read to perform supervised adjustments to determine the classification large model, ensure the adaptability and standardization of the model, and combine the classification large model to make classification decisions and determine the classification results including the attribution category and the thinking chain. The technical problem that the power knowledge labels in the prior art are highly professional, resulting in scarce label category data and poor interpretability is solved. The interpretability of the model is enhanced by introducing domain knowledge, and the label classification of standard knowledge is realized. The classification decision is made in combination with the classification large model to determine the classification results, which is used to predict and infer the accuracy of new sample data.

[0016] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of a flow chart of a method for classifying power label knowledge based on a large model in an embodiment;

[0018] Figure 2 A schematic diagram of a process of optimizing learning of a classification big model of a power label knowledge classification method based on a big model in one embodiment;

[0019] Figure 3 is a structural block diagram of a large model-based power label knowledge classification system in one embodiment;

[0020] Figure 4 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment.

[0021] Explanation of the accompanying drawings: effective source data determination module 11, open source large model training module 12, standardization learning module 13, classification large model determination module 14, classification result determination module 15. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0023] like Figure 1 As shown, the present application provides a method for classifying power label knowledge based on a large model, the method comprising:

[0024] Collect and pre-process the power source data to determine the effective source data, where the pre-processing mode is defect type screening and class parallel processing;

[0025] The power label knowledge classification method based on a large model is a method for classifying power label knowledge using a machine learning model with large-scale parameters and complex computing structures, wherein a large model refers to a machine learning model with large-scale parameters and complex computing structures, and power label knowledge classification refers to the systematic classification and organization of various labels in the power field, so as to better organize, manage and utilize relevant power knowledge. The present application provides a power label knowledge classification method based on a large model, which combines the advantages of big data processing and deep learning, can achieve efficient and accurate classification of power labels, and can be used to predict and infer new sample data.

[0026] Collecting power source data refers to collecting data from various sources of the power system. The power source refers to the source of electric energy, that is, various types of power generation facilities. The power source data usually refers to the original data directly obtained from these power generation facilities and their related equipment, including equipment operating status, power generation, energy conversion efficiency, equipment fault records, etc., which are of great significance for the monitoring, optimization and fault prediction of the power system. Preprocessing refers to the pre-use processing of the power source data, including defect type screening and class parallel processing; wherein the defect type screening includes preliminarily screening out data related to a specific defect type according to the type, operating environment and historical fault records of the power equipment, extracting features related to the defect type from the screened data, and labeling the extracted features, that is, assigning one or more defect type labels to each feature; class parallel processing includes dividing the screened and labeled data into blocks according to the defect type, each data block contains data of the same or similar defect type, and processing multiple data blocks at the same time, which greatly speeds up the speed of data preprocessing, and merges the processing results of each data block to form a complete preprocessing data set. The power source data is collected and effectively preprocessed to determine the effective source data, which includes valuable and accurate parts for subsequent analysis. Through the above method, the effective part of the power source data can be effectively determined, providing high-quality data support for subsequent power label knowledge classification tasks.

[0027] Determining data quality standards, identifying the power source end data, extracting data based on defect types, and determining multiple data classes, wherein the multiple data classes have data overlap;

[0028] Performing parallel processing on the multiple data classes to determine multiple preprocessing results, wherein the data is processed in parallel based on a preprocessing module, the preprocessing module includes multiple parallel processing branches trained based on a difference preprocessing algorithm;

[0029] Homologous fitting and regularization are performed on the multiple preprocessing results to determine the effective source data.

[0030] Data quality standards are the basis for ensuring data accuracy and reliability, including integrity, whether the data is complete, without missing or missing; accuracy, whether the data truly reflects the actual situation, without errors or deviations; consistency, whether the data are logically consistent, without contradictions or conflicts; timeliness, whether the data is up-to-date and can reflect the current system status. Identifying the power source data refers to judging whether the power source data meets the standards according to the data quality standards, and extracting data based on defect types refers to extracting power source data based on known defect types. For example, if you are concerned about the fault type of a wind turbine, you need to extract operating parameters and fault records related to the wind turbine from the data. In the process of extracting data, it may be found that there is overlap between multiple data classes. For example, some data may be related to multiple defect types at the same time. For example, if a part of the data has multiple defects, it is necessary to perform multi-end parallel processing and fitting. The preprocessing module includes multiple parallel processing branches trained based on the difference preprocessing algorithm. These branches can process multiple data classes at the same time to improve processing efficiency. Each branch can adopt different preprocessing strategies, such as data cleaning, standardization, feature selection, etc., to adapt to the characteristics of different data classes. Since multiple parallel processing branches adopt different preprocessing strategies according to different defects, it is necessary to fit and regularize the processed results, including data alignment, standardization or feature conversion, to ensure the homology and consistency between them, and determine the final valid source data according to the quality, relevance and validity of the processing results. The valid remote data will be used as input for subsequent power label knowledge classification. Through the above method, it is ensured that the determined valid source data can meet the needs of practical applications.

[0031] Classify and determine a hierarchical label map based on the effective source data, and supervise and train an open source big model based on the hierarchical label map and the effective source data, wherein the training standard includes a classification dimension and a thinking chain derivation dimension, the thinking chain derivation dimension includes a thinking chain mode, and the thinking chain mode is an integrated visualization mode of the thinking chain;

[0032] After preprocessing and screening of effective source data, high-quality data used for model training can accurately reflect the characteristics and distribution of actual problems. Hierarchical label map is a structure used to organize and represent the relationship between labels at different levels. In the power sector, this can include multiple levels from high-level general labels to low-level specific labels, conduct in-depth exploration of effective source data, understand the inherent structure and characteristics of the data, and define different levels of labels based on data characteristics and business needs. For example, in the power sector, high-level labels may be power generation types, while low-level labels may be specific power generation methods, such as thermal power generation, hydropower generation, etc. Determine the hierarchical relationship between labels and build a hierarchical label structure. Supervised training is to use data with known labels to train the model so that it can automatically classify or predict unknown data, perform necessary preprocessing on effective source data, such as encoding, normalization, etc., to meet the input requirements of the model; select a suitable open source large model as the basis. Considering the complexity of the power sector and the diversity of data, it is necessary to select a model with strong representation learning capabilities. Map the hierarchical label map to the output layer of the model to ensure that the model can output the prediction results corresponding to the label map, configure the training parameters such as learning rate, batch size, etc. in conjunction with the effective source data, and use the effective source data with labels to train the model. During the training process, the model will learn to extract useful features from the input data and make predictions based on the labels. During the training process, the training standards include classification dimensions and thought chain deduction dimensions, and there is a thought chain mode; the classification dimension includes specified standards and some classification label samples, which prompts the model to generate classified data that meets specific standards; the thought chain deduction dimension requires that when a large model generates these data, it must be accompanied by the reasoning process of the thought chain in order to understand the inference process and logic of the model when generating samples. Through the above method, the hierarchical label map can be determined based on the effective source data, and the open source large model can be supervised and trained in conjunction with these data to achieve accurate classification and prediction of power data.

[0033] Identify the valid source data, perform data clustering based on power tag classes, and determine N data sets;

[0034] Traversing the N data sets, extracting the first data set and performing class label feature extraction and hierarchical association integration to determine a class of knowledge graphs;

[0035] Complete the construction of the feature graph of the Nth data set and determine N types of knowledge graphs;

[0036] The one type of knowledge graph up to the N types of knowledge graphs are hierarchically associated and integrated to determine the hierarchical label graph.

[0037] According to the data quality standard determined previously, valid data that meets the standard is screened out from the power source data, and the valid data will serve as the basis for subsequent clustering and analysis. According to business needs and data characteristics, define power-related label classes, such as power generation type, equipment status, fault type, etc., extract features related to the power label class from the valid source data, and cluster the data according to the power label class. When clustering, it is necessary to select appropriate clustering algorithms and parameters to ensure the accuracy and effectiveness of the clustering results. After clustering, N different data sets will be obtained, each of which represents a specific power label class. Traverse the N data sets, extract the first data set, and perform class label feature extraction and hierarchical association integration to determine a class of knowledge graphs. Starting from the first data set, traverse the N data sets one by one. For each data set, extract features related to the class label. Based on the extracted features, perform hierarchical association integration to construct a graph of this class of knowledge, that is, a class of knowledge graphs. When constructing a knowledge graph, it is necessary to make full use of professional knowledge and experience in the power field to ensure the accuracy and practicality of the graph, where hierarchical association integration includes determining the relationship between features, hierarchical structure, etc. Repeat the above steps until the construction of the feature graph of the Nth data set is completed, thereby obtaining N types of knowledge graphs. Analyze the relationship between these N types of knowledge graphs, determine the hierarchical structure and association paths between them, and based on the analysis results, hierarchically integrate these N types of knowledge graphs. When performing hierarchical association integration, it is necessary to comprehensively consider the correlation and consistency between different graphs to ensure that the final hierarchical label graph has integrity and consistency, forming a complete hierarchical label graph. The hierarchical label graph not only shows the relationship between different power label classes, but also reveals the hierarchical structure and dependencies. Through the above steps, a power-related hierarchical label graph can be constructed based on effective source data, providing strong support for subsequent power label knowledge classification and application.

[0038] Based on the class relevance, associate nodes between the one class of knowledge graphs up to the N classes of knowledge graphs to determine an aggregate knowledge graph, wherein the aggregate knowledge graph includes at least one;

[0039] Based on the likelihood probability, the learning rate of the overall knowledge graph is marked, wherein the likelihood probability is negatively correlated with the learning rate;

[0040] Based on the degree of feature personalization, invalid graph nodes are removed from the overall knowledge graph;

[0041] The learning rate identifier and the overall knowledge graph after the invalid graph nodes are removed are fitted to determine the hierarchical label graph.

[0042] Based on class relevance, the nodes between the spectra are associated, which includes class relevance analysis of one class of knowledge graphs to N classes of knowledge graphs. The class relevance includes the degree of association between different classes and how different classes interact with each other, and the association points between different categories of knowledge are determined. According to the analysis results, the nodes are associated between the knowledge graphs to ensure that the relevant information can be circulated and integrated between the graphs. After the node association, at least one total knowledge graph is formed, and the total knowledge graph integrates the association information of multiple categories of knowledge, including at least one independent graph of an unrelated power label class. For each node or sub-graph in the total knowledge graph, its likelihood probability is calculated according to its actual frequency of occurrence and relevance in the power field. The likelihood probability is a core concept in statistics. When the result is known, the possible environmental parameters that produce this result are inferred. For example, in the context of coin tossing, it is used to infer the nature of the coin or other influencing factors that may exist in the tossing process after the tossing result is known. In this application, there is a negative correlation between the likelihood probability and the learning rate, and the learning rate is identified for each node or sub-graph. Nodes or sub-graphs with higher likelihood probabilities will be assigned lower learning rates. Based on the degree of feature personalization, invalid graph nodes are removed from the total knowledge graph. For each node or sub-graph in the total knowledge graph, its degree of feature personalization is evaluated. The degree of feature personalization refers to its uniqueness and contribution to the overall graph. Based on the evaluation results, invalid graph nodes are removed, that is, those nodes or sub-graphs with low degree of feature personalization and small contribution to the overall graph are removed to reduce redundancy and noise. The learning rate identifier is fitted with the total knowledge graph after the invalid graph nodes are removed, that is, the learning rate identifier is placed in the total knowledge graph after the removal, ensuring that the learning rate can reasonably reflect the importance of different nodes or sub-graphs in the graph. The total knowledge graph after fitting is integrated with the previous hierarchical association integration results to form a more complete and accurate hierarchical label graph. According to the above method, a more comprehensive, accurate and practical hierarchical label graph can be constructed to provide strong support for the classification and application of label knowledge in the power field.

[0043] Determine a decision map label, wherein the decision map label is multiple;

[0044] Based on the learning rate, weight distribution and overall analysis are performed on the decision graph labels to determine the belonging category;

[0045] Based on the thinking chain model, effective derivation features are screened, the derivation steps are integrated in a pattern, and the thinking chain is determined;

[0046] The belonging category and the thought chain are taken as the classification result.

[0047] Determine the decision map label, the decision map label refers to a highly refined feature identification obtained by applying abstraction, induction, reasoning and other algorithms to the target object based on the needs of the business scenario, which is used for differentiated management and decision-making. In the decision-making process, the decision map label can help better understand and analyze the data, discover the associations and laws between the data, and make more accurate decisions. In this application, the decision map label should be able to fully cover the different aspects and levels of the power field, so as to facilitate subsequent classification and decision-making. When determining the decision map label, it is necessary to ensure the comprehensiveness and accuracy of the label to avoid omissions or errors. Map the defined decision map label with the previous hierarchical label map to ensure that the correspondence between the label and the data is accurate. According to the learning rate previously identified for each node or sub-map, calculate the learning rate corresponding to each decision map label, and based on the learning rate, weight the decision map label. Labels with higher learning rates will obtain larger weights, and vice versa. When performing weight distribution, it is necessary to reasonably set the weight of the learning rate to ensure the rationality and effectiveness of the weight distribution. Overall analysis refers to comprehensively considering the relationship between weight distribution and labels to determine the degree of influence of each label on the classification results. Based on the thinking chain model, effective derivation features are screened, and features related to decision map labels are extracted from the data. Based on the thinking chain model, features that can effectively derive classification results are screened. The features should be able to form a coherent derivation path to support the rationality of classification decisions. The screened effective derivation features are patterned and integrated to form a thinking chain. When screening effective derivation features, it is necessary to focus on the representativeness and interpretability of the features to avoid introducing irrelevant or redundant features. This helps to understand the logic and process of classification decisions. According to the results of weight distribution and overall analysis, the belonging class of each decision map label is determined. The belonging class represents the classification result of the data under a specific label. The determined belonging class is combined with the thinking chain to form a complete classification result. The classification result not only includes the belonging class of the data, but also shows the thinking chain process of deriving to the belonging class. Through the above method, the decision map label can be determined, the weight distribution and overall analysis can be performed, the effective derivation features can be screened, and the classification results can be finally determined. This helps to provide strong support for the classification and application of label knowledge in the power field.

[0048] Introduce an unlabeled standard semantic library, perform standardized training on the open source large model, and determine the initialization classification model;

[0049] The unlabeled standard semantic library is a collection of a large amount of unlabeled text data. These data are carefully screened and processed to reflect the semantic information of a specific field or topic. Since these data do not have clear labels, unsupervised learning methods are required to extract the semantic features and structures therein. In this application, the field of the unlabeled standard semantic library is the power field; the open source large model usually has huge parameters and complex structures, and can adapt to a variety of tasks. In order to optimize for a specific task, it is necessary to perform standardized learning on it. This includes pre-training the model using the unlabeled standard semantic library and fine-tuning the model according to task requirements. Standardized learning is to optimize for the power field. After standardized learning, an initialization classification model for the classification of power label knowledge can be obtained. The initialization classification model already has certain semantic understanding and classification capabilities, which can be used as the basis for subsequent training and optimization. By introducing the unlabeled standard semantic library to perform standardized learning on the open source large model, an initialization classification model for a specific task can be obtained. The initialization classification model provides a solid foundation for subsequent classification tasks, and the performance can be improved through further training and optimization.

[0050] Read downstream task features, perform supervised adjustment on the initialized classification model, and determine a large classification model;

[0051] Downstream task features refer to data features related to specific classification tasks. These features usually reflect the specific needs and characteristics of the tasks. For example, in each device in the power system, obtaining downstream task features refers to obtaining operating status and power information. Supervised adjustment refers to using labeled downstream task data to train and adjust the initialization classification model so that the model can better adapt to the needs of specific tasks. The initialization classification model is supervised and adjusted to determine the classification model. The classification model fully learns the characteristics and laws of the downstream tasks and can make accurate classification predictions on new data.

[0052] like Figure 2 As shown, batch training data is integrated and test samples are randomly selected;

[0053] Inputting the test sample into the classification model for convergence verification, and determining the loss amount based on the sample verification result;

[0054] If the loss amount is greater than the loss threshold, the classification model is optimized and learned in combination with the cross entropy loss function.

[0055] Integrate batch training data, randomly screen test samples, and organize all available training data into batches in a fixed manner, such as randomly or sequentially, which can effectively utilize memory and allow the model to gradually learn the distribution of data during training. From the integrated batch training data, a part is randomly selected as a test sample. These test samples will be used to verify the performance of the classification model and adjust the classification model based on the prediction results of these samples. By inputting the test sample into the classification model, observe the difference between the prediction result of the classification model and the true label. This difference is usually measured by calculating a loss function, such as cross entropy loss. The loss amount is a value calculated according to the loss function, which indicates the prediction performance of the classification model on the test sample. The smaller the loss value, the more accurate the model's prediction. If the loss amount is greater than the loss threshold, it means that the performance of the current classification model has not reached the expected standard and needs further training and optimization. The loss threshold is set by the staff. The cross entropy loss function is a commonly used loss function in classification tasks, which is used to measure the difference between the probability distribution predicted by the model and the true label. Based on the calculation results of the cross entropy loss function, the parameters of the model are updated through the back propagation algorithm, and the optimization algorithm is used to accelerate and stabilize the learning process to reduce the loss value and improve the performance of the model. For example, the gradient descent method can be used to iteratively calculate the gradient of the objective function and update the parameters in the opposite direction of the gradient to minimize the loss function or maximize the utility function. By continuously adjusting the parameters, the prediction results of the model gradually approach the true value. The above method is a typical machine learning training cycle. Through continuous iteration and optimization, the model gradually learns to extract features from the input data and perform accurate classification.

[0056] Receiving a pre-classification task, making a classification decision in combination with the classification model, and determining a classification result, wherein the classification result includes a belonging category and a thinking chain;

[0057] Receive a pre-classification task, which includes the task of identifying the power knowledge of the target, combining the trained classification model to make a classification decision, and determine the classification result. The classification model will process and analyze the input data based on the knowledge and patterns it has learned, so as to output the classification result. The classification result usually includes two aspects, including the belonging category, that is, the specific category to which the data is classified and the thinking chain, or the reasoning path, that is, the logic or feature sequence based on which the model obtains the classification result. The belonging category means that the model classifies the input data into a pre-defined category based on its characteristics. For example, in an image classification task, the model will classify a picture into one of the categories such as cats, dogs or other animals. The thinking chain or reasoning path focuses more on explaining how the model makes this classification decision. It usually involves how the model processes and transforms the input data internally, and which features or patterns have an impact on the final classification result, which can be obtained by viewing the output, activation state or feature importance of each layer of the model. Through the above method, the classification model can be used to efficiently and accurately classify the pre-classification task and obtain valuable information about the model decision-making process. This is of great significance for improving the reliability, interpretability and optimization performance of the model.

[0058] If the pre-classification task is a newly added defect type, perform standardized prediction on the task information to determine the prediction sample;

[0059] In combination with the prediction samples, performing class incremental learning on the classification large model to determine a pre-learning large model;

[0060] Performing a balance rate and overfitting check on the pre-learning large model to determine a check result, wherein the balance rate is a balance ratio of the classification effect before and after the incremental learning;

[0061] The verification result is identified, and loss compensation is performed on the pre-learned large model.

[0062] When the classification result of the pre-classification task is empty, that is, the model cannot classify the input data into any existing category, it means that the input data belongs to a completely new category that the model has not yet learned. In this case, receiving the newly added defect type, that is, this new category and performing class incremental learning of the classification model is an effective solution. Class incremental learning is suitable for scenarios where the number of categories is increasing. In class incremental learning, the model needs to continuously learn new categories while maintaining the ability to recognize the categories that have been learned. By performing class incremental learning, the classification model can continuously adapt to new classification tasks and maintain the ability to recognize existing categories. When the pre-classification task is a new defect type, it is first necessary to perform standardized prediction on the task information. Standardized prediction usually involves preprocessing the task data, such as data cleaning, feature extraction and conversion, so as to convert it into a format that can be processed by the classification model. Based on the learning effect of the unlabeled standard semantic library, sample prediction is performed. Since the task data is a single group or too few, sample prediction is required to determine the prediction samples. These prediction samples will be used as representative data of the newly added category for subsequent class incremental learning. Combined with the prediction samples, class incremental learning is performed on the classification model. The purpose of incremental learning is to enable the classification model to recognize new defect types while maintaining the ability to recognize the original categories. During the learning process, the classification model will adjust its parameters and internal structure to adapt to the new classification task. It includes adding new nodes to the output layer of the classification model to represent the newly added categories, and using the predicted samples to train the nodes to obtain the pre-learning model. After completing the incremental learning, the pre-learning model needs to be checked for balance rate and overfitting. The balance rate refers to the balance ratio of the classification effect of the model in each category before and after incremental learning, which is used to evaluate whether the performance of the model in different categories is balanced. Overfitting check is to check whether the model is overfitting the training data, resulting in poor performance on the test data. With incremental learning, the classification model has a certain degree of forgetting, that is, the network is more inclined to new categories; and due to the small number of incremental learning samples, there is a certain degree of overfitting problem. These checks help identify problems with the model and provide a basis for subsequent loss compensation. According to the verification results, loss compensation is performed on the pre-learned large model to correct the performance degradation or deviation that may occur in the model during the incremental learning process. Specific compensation methods include adjusting model parameters, adding regularization terms, introducing new training strategies, etc. Through loss compensation, the classification performance of the model on new and old categories can be further improved, making it more stable and reliable.

[0063] In summary, the present application has the following beneficial effects:

[0064] The technical problem that the power knowledge labels in the existing technology are highly professional, resulting in scarce label category data and poor interpretability is solved. The technical effect of introducing domain knowledge to enhance the interpretability of the model, realize label classification of standard knowledge, and use it to predict and infer the accuracy of new sample data is achieved.

[0065] like Figure 3 As shown, the embodiment of the present application includes a large model-based power label knowledge classification system, the system comprising:

[0066] An effective source end data determination module 11, the effective source end data determination module 11 is used to collect and pre-process the power source end data to determine the effective source end data, wherein the pre-processing mode is defect type screening and class parallel processing;

[0067] An open source big model training module 12, the open source big model training module 12 is used to classify and determine a hierarchical label map based on the effective source end data, and supervise the training of the open source big model based on the hierarchical label map and the effective source end data, wherein the training standard includes a classification dimension and a thinking chain derivation dimension, the thinking chain derivation dimension includes a thinking chain mode, and the thinking chain mode is an integrated visualization mode of the thinking chain;

[0068] A standardization learning module 13, wherein the standardization learning module 13 is used to introduce an unlabeled standard semantic library, perform standardization training on the open source large model, and determine an initialization classification model;

[0069] A classification large model determination module 14, wherein the classification large model determination module 14 is used to read downstream task features, perform supervised adjustment on the initialized classification model, and determine the classification large model;

[0070] A classification result determination module 15, wherein the classification result determination module 15 is used to receive a pre-classification task, make a classification decision in combination with the classification large model, and determine a classification result, wherein the classification result includes a belonging category and a thinking chain;

[0071] Furthermore, the embodiment of the present application also includes:

[0072] A data extraction module, the data extraction module is used to determine a data quality standard, identify the power source end data, extract data based on defect types, and determine multiple data classes, wherein the multiple data classes have data overlap;

[0073] A parallel processing module, the parallel processing module is used to perform parallel processing on the multiple data classes and determine multiple preprocessing results, wherein the preprocessing module is based on performing data parallel processing, and the preprocessing module includes multiple parallel processing branches trained based on a difference preprocessing algorithm;

[0074] A homologous fitting and regularization module is used to perform homologous fitting and regularization on the multiple preprocessing results to determine the effective source end data.

[0075] Furthermore, the embodiment of the present application also includes:

[0076] A data clustering module, the data clustering module is used to identify the valid source data, perform data clustering based on power tag classes, and determine N data sets;

[0077] A knowledge graph determination module, the knowledge graph determination module is used to traverse the N data sets, extract the first data set and perform class label feature extraction and hierarchical association integration to determine a class of knowledge graphs;

[0078] A feature graph construction module, wherein the feature graph construction module is used to complete the feature graph construction of the Nth data set and determine N types of knowledge graphs;

[0079] A hierarchical association integration module is used to perform hierarchical association integration on the one type of knowledge graph up to the N types of knowledge graphs to determine the hierarchical label graph.

[0080] Furthermore, the embodiment of the present application also includes:

[0081] An assembly knowledge graph determination module, the assembly knowledge graph determination module is used to associate nodes between the one type of knowledge graphs to the N types of knowledge graphs based on class relevance to determine an assembly knowledge graph, the assembly knowledge graph including at least one;

[0082] A learning rate identification module, wherein the learning rate identification module is used to identify the learning rate of the overall knowledge graph based on the likelihood probability, wherein the likelihood probability is negatively correlated with the learning rate;

[0083] An invalid graph node elimination module, which is used to eliminate invalid graph nodes from the overall knowledge graph based on the degree of feature personalization;

[0084] A hierarchical label graph determination module is used to fit the overall knowledge graph after the learning rate identifier and invalid graph nodes are removed to determine the hierarchical label graph.

[0085] Furthermore, the embodiment of the present application also includes:

[0086] A test sample screening module, which is used to integrate batch training data and randomly screen test samples;

[0087] A convergence verification module, the convergence verification module is used to input the test sample into the classification model to perform convergence verification and determine the loss amount based on the sample verification result;

[0088] The optimization learning module is used to optimize the learning of the classification model in combination with the cross entropy loss function if the loss amount is greater than the loss threshold.

[0089] Furthermore, the embodiment of the present application also includes:

[0090] A decision graph label determination module, the decision graph label determination module is used to determine a decision graph label, wherein the decision graph label is multiple;

[0091] A weight distribution module, the weight distribution module is used to perform weight distribution and overall analysis on the decision map labels based on the learning rate to determine the belonging category;

[0092] A pattern integration module, the pattern integration module is used to screen effective derivation features based on the thinking chain pattern, perform pattern integration on the derivation steps, and determine the thinking chain;

[0093] A classification result acquisition module is used to take the belonging category and the thought chain as the classification result.

[0094] Furthermore, the embodiment of the present application also includes:

[0095] A prediction sample determination module, wherein if the pre-classification task is a newly added defect type, the prediction sample determination module is used to perform standardized prediction on the task information and determine the prediction sample;

[0096] A class incremental learning module, wherein the class incremental learning module is used to perform class incremental learning on the classification large model in combination with the prediction sample to determine a pre-learning large model;

[0097] A verification result determination module, the verification result determination module is used to perform a balance rate and overfitting check on the pre-learning large model to determine a verification result, wherein the balance rate is a balance ratio of the classification effect before and after the incremental learning of the class;

[0098] A loss compensation module is used to identify the verification result and perform loss compensation on the pre-learned large model.

[0099] For a specific embodiment of a large-model-based power label knowledge classification system, please refer to the embodiment of a large-model-based power label knowledge classification method described above, which will not be repeated here. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0100] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. The computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store news data and data such as time attenuation factors. The network interface of the computer device is used to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement a large model-based power label knowledge classification method.

[0101] Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0102] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, steps of a large-model-based power label knowledge classification method are implemented.

[0103] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of a method for classifying power label knowledge based on a large model are implemented.

[0104] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0105] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. 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 application, and these all belong to the protection scope of the present application.

Claims

1. A large-model-based power label knowledge classification method, characterized in that: The method comprises: Collect and preprocess the power source data to determine the effective source data, wherein the preprocessing mode is defect type screening and class parallel processing; the power source refers to various types of power generation facilities, and the power source data includes equipment operating status, power generation, energy conversion efficiency and equipment fault records; the defect type screening includes preliminarily screening out data related to a specific defect type based on the type of power equipment, operating environment and historical fault records, extracting features related to the defect type from the screened data, and labeling the extracted features, that is, assigning one or more defect type labels to each feature; the class parallel processing includes dividing the screened and labeled data into blocks according to the defect type, each data block contains data of the same or similar defect type, processing multiple data blocks at the same time, and merging the processing results of each data block to form a complete preprocessed data set; Classify and determine a hierarchical label map based on the effective source data, and supervise the training of an open source big model based on the hierarchical label map and the effective source data, wherein the training standard includes a classification dimension and a thought chain derivation dimension, the thought chain derivation dimension includes a thought chain mode, and the thought chain mode is an integrated visualization mode of the thought chain; Introduce an unlabeled standard semantic library, perform standardized training on the open source large model, and determine the initialization classification model; Read downstream task features, perform supervised adjustment on the initialized classification model, and determine a large classification model; Receiving a pre-classification task, making a classification decision in combination with the classification model, and determining a classification result, wherein the classification result includes a belonging category and a thinking chain; The classification is determined based on a hierarchical label map of the valid source data, including: Identify the valid source data, perform data clustering based on power tag classes, and determine N data sets; Traversing the N data sets, extracting the first data set and performing class label feature extraction and hierarchical association integration to determine a class of knowledge graphs; Complete the construction of the feature graph of the Nth data set and determine N types of knowledge graphs; Performing hierarchical association integration on the one type of knowledge graph up to the N types of knowledge graphs to determine the hierarchical label graph; The one type of knowledge graph to the N types of knowledge graphs are hierarchically associated and integrated, including: Based on the class relevance, associate nodes between the one class of knowledge graphs up to the N classes of knowledge graphs to determine an aggregate knowledge graph, wherein the aggregate knowledge graph includes at least one; Based on the likelihood probability, the learning rate of the overall knowledge graph is marked, wherein the likelihood probability is negatively correlated with the learning rate; Based on the degree of feature personalization, invalid graph nodes are removed from the overall knowledge graph; The overall knowledge graph after fitting the learning rate identifier and removing the invalid graph nodes is used to determine the hierarchical label graph for label knowledge classification in the power field.

2. The method according to claim 1, characterized in that The determining of valid source data includes: Determining data quality standards, identifying the power source end data, extracting data based on defect types, and determining multiple data classes, wherein the multiple data classes have data overlap; Performing parallel processing on the multiple data classes to determine multiple preprocessing results, wherein the data is processed in parallel based on a preprocessing module, the preprocessing module includes multiple parallel processing branches trained based on a difference preprocessing algorithm; Homologous fitting and regularization are performed on the multiple preprocessing results to determine the effective source data.

3. The method according to claim 1, characterized in that After determining the classification model, it includes: Integrate batch training data and randomly select test samples; Inputting the test sample into the classification model for convergence verification, and determining the loss amount based on the sample verification result; If the loss amount is greater than the loss threshold, the classification model is optimized and learned in combination with the cross entropy loss function.

4. The method according to claim 1, characterized in that Determining the classification result includes: Determine a decision map label, wherein the decision map label is multiple; Based on the learning rate, weight distribution and overall analysis are performed on the decision graph labels to determine the belonging category; Based on the thinking chain model, effective derivation features are screened, the derivation steps are integrated in a pattern, and the thinking chain is determined; The belonging category and the thought chain are taken as the classification result.

5. The method according to claim 1, characterized in that If the classification result of the pre-classification task is empty, receiving a newly added defect type, and performing class incremental learning of the classification large model; The step of performing incremental learning of the classification model includes: If the pre-classification task is a newly added defect type, perform standardized prediction on the task information to determine the prediction sample; In combination with the prediction samples, performing class incremental learning on the classification large model to determine a pre-learning large model; Performing a balance rate and overfitting check on the pre-learning large model to determine a check result, wherein the balance rate is a balance ratio of the classification effect before and after the incremental learning; The verification result is identified, and loss compensation is performed on the pre-learned large model.

6. A large-model-based power label knowledge classification system, characterized in that: The system comprises: An effective source end data determination module, the effective source end data determination module is used to collect power source end data and pre-process it to determine the effective source end data, wherein the pre-processing mode is defect type screening and class parallel processing; the power source end refers to various types of power generation facilities, and the power source end data includes equipment operating status, power generation, energy conversion efficiency and equipment fault records; the defect type screening includes preliminarily screening out data related to a specific defect type based on the type of power equipment, operating environment and historical fault records, extracting features related to the defect type from the screened data, and labeling the extracted features, that is, assigning one or more defect type labels to each feature; the class parallel processing includes dividing the screened and labeled data into blocks according to the defect type, each data block contains data of the same or similar defect type, processing multiple data blocks at the same time, and merging the processing results of each data block to form a complete pre-processed data set; An open source big model training module, the open source big model training module is used to classify and determine a hierarchical label map based on the effective source end data, and supervise the training of the open source big model based on the hierarchical label map and the effective source end data, wherein the training standard includes a classification dimension and a thinking chain derivation dimension, the thinking chain derivation dimension includes a thinking chain mode, and the thinking chain mode is an integrated visualization mode of the thinking chain; A standardized learning module, wherein the standardized learning module is used to introduce an unlabeled standard semantic library, perform standardized training on the open source large model, and determine an initialization classification model; A classification large model determination module, wherein the classification large model determination module is used to read downstream task features, perform supervised adjustment on the initialized classification model, and determine the classification large model; A classification result determination module, the classification result determination module is used to receive the pre-classification task, make a classification decision in combination with the classification model, and determine the classification result, wherein the classification result includes the belonging category and the thinking chain; A data clustering module, the data clustering module is used to identify the valid source data, perform data clustering based on power tag classes, and determine N data sets; A knowledge graph determination module, the knowledge graph determination module is used to traverse the N data sets, extract the first data set and perform class label feature extraction and hierarchical association integration to determine a class of knowledge graphs; A feature graph construction module, wherein the feature graph construction module is used to complete the feature graph construction of the Nth data set and determine N types of knowledge graphs; A hierarchical association integration module, the hierarchical association integration module is used to perform hierarchical association integration on the one type of knowledge graph up to the N types of knowledge graphs to determine the hierarchical label graph; An assembly knowledge graph determination module, the assembly knowledge graph determination module is used to associate nodes between the one type of knowledge graphs to the N types of knowledge graphs based on class relevance to determine an assembly knowledge graph, the assembly knowledge graph including at least one; A learning rate identification module, wherein the learning rate identification module is used to identify the learning rate of the overall knowledge graph based on the likelihood probability, wherein the likelihood probability is negatively correlated with the learning rate; An invalid graph node elimination module, which is used to eliminate invalid graph nodes from the overall knowledge graph based on the degree of feature personalization; A hierarchical label graph determination module is used to fit the overall knowledge graph after the learning rate identifier and invalid graph nodes are removed, and determine the hierarchical label graph for label knowledge classification in the power field.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images