A large model empowerment graph convolution network aging prediction model automatic construction method

Abstract

The present application relates to the technical field of aging prediction, in particular to a large model enabled graph convolution network aging prediction model automatic construction method, comprising: obtaining an aging prediction original data set, extracting data entities according to column names to form a preliminary set, inputting the preliminary set into a general large language model to complete semantic analysis and relationship reasoning, automatically generating a preliminary triple set, removing unreasonable and excessive triples through verification screening to obtain a finished triple set, constructing a knowledge graph structure, and then inputting the knowledge graph and numerical features of the original data set into a graph convolution network model, training and automatically evaluating and optimizing through preset hyperparameter configuration combination, obtaining an optimized aging prediction model, and forming an optimal model through multiple cycle operations. The method realizes the automation of entity relationship reasoning and triple generation, reduces manual subjective bias, completes model full-process automatic construction by fusing multiple features, and improves entity relationship processing accuracy and model construction process integrity.

Description

Technical Field

[0001] This invention relates to the field of aging prediction technology, and in particular to an automated construction method for aging prediction models using graph convolutional networks with large model empowerment. Background Technology

[0002] In the current aging prediction model construction process, data entity extraction relies on manual screening and definition based on column names in the original dataset, and the relationships between data entities rely on manual inference and summarization using domain knowledge. Triples representing entity relationships are manually constructed and validated before a knowledge graph is built based on the qualified triples. Graph convolutional network model training only uses numerical features from the original dataset as input; model training, performance evaluation, and model optimization all rely on manual parameter adjustment and effect verification. The entire model construction process is manual, lacking an automated construction mechanism.

[0003] Human intervention in entity extraction and relational reasoning is prone to semantic cognitive biases. The rationality and quantity of generated triples cannot be precisely controlled, and the pre-processing of data in knowledge graph construction is highly subjective. Numerical features and knowledge graph structures cannot be integrated into the model, feature utilization is limited, and model training, evaluation, and optimization require significant human intervention, resulting in cumbersome processes with poor adaptability. A general-purpose large language model is needed to automatically complete data entity semantic analysis, relational reasoning, and triple generation, while simultaneously integrating knowledge graph structures and numerical features as input. This would enable automated training, evaluation, and optimization of graph convolutional network models, addressing the problems of high human dependence, inaccurate entity relational reasoning, missing feature fusion, and non-automated model construction in traditional processes. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an automated construction method for aging prediction models of graph convolutional networks empowered by large models.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an automated construction method for a graph convolutional network aging prediction model empowered by a large model, comprising: Obtain the raw dataset related to aging prediction, which contains multiple data columns and the column name corresponding to each data column; Based on the column names of the original dataset, all or some column names are automatically extracted, and each extracted column name is defined as a data entity to form a preliminary data entity set. The preliminary set of data entities is input into a preset general-purpose large language model, which drives the general-purpose large language model to perform semantic analysis and relational reasoning on each data entity in the preliminary set of data entities based on the domain knowledge it has learned internally, and automatically generates a preliminary set of triples describing the relationship between the data entities. A verification and screening process is performed on the preliminary triplet set to remove triplets that do not meet the set reasonableness conditions or exceed the preset quantity constraint range, thereby obtaining the refined triplet set. Using the aforementioned set of triplets, with the data entities contained therein as nodes and the relationships between the data entities as connecting edges, a knowledge graph structure is constructed. The knowledge graph structure is combined with numerical features from the original dataset and input into the graph convolutional network model. The graph convolutional network model is trained based on a preset combination of hyperparameters and automatically evaluated and optimized to generate an optimized graph convolutional network aging prediction model.

[0006] As a further aspect of the present invention, the general-purpose large language model is driven to perform semantic analysis and relational reasoning on each data entity in the preliminary data entity set based on its internally learned domain knowledge, automatically generating a preliminary set of triples describing the relationships between data entities, including: The initial set of data entities and instructions for the entity relationship discovery task are injected into the input prompts of the general large language model. The general large language model is guided to infer and describe the physical, chemical, or statistical relationships between any two or more data entities in the preliminary data entity set, based on the domain knowledge it has learned internally. The system receives the natural language descriptions output by the general large language model, and performs structured parsing and transformation on each natural language statement describing the relationship between entities according to the format of "head entity-relationship description-tail entity" to generate a triple with a uniform format. All generated triples are collected to form the preliminary triple set.

[0007] As a further aspect of the present invention, a verification and screening process is performed on the preliminary triplet set to remove triplets that do not meet the set reasonableness conditions or exceed the preset quantity constraint range, thereby obtaining a refined triplet set, including: Calculate the total number of triplets in the initial set of triplets; Set a threshold range for the number of triples related to the size of the original data and the complexity of the features, and compare the total number of triples with the threshold range; When the total number of triples exceeds the upper limit of the number threshold range, the general large language model is activated to score the relation importance of each triple in the initial triple set. Based on the ranking result of the relation importance scores, the triples with the lowest scores are removed until the number of remaining triples falls within the number threshold range. When the total number of triples is below the lower limit of the number threshold range or within the range, the general big language model is activated to perform a logical rationality check on the preliminary triple set, and triples that are determined by the general big language model to have logical contradictions, do not conform to physical laws, or are weakly related to the aging prediction task are removed.

[0008] As a further aspect of the present invention, the general large language model is activated to score the relation importance of each triple in the initial triple set. Based on the ranking of relation importance scores, the triples with the lowest scores are removed, including: Each triple in the initial set of triples is input into the general large language model, and an instruction is attached to evaluate the importance of the relationship described by the triples to the aging prediction task. Receive a numerical importance score from the output of the general large language model for each input triple; Sort all triples in the initial triplet set from highest to lowest according to their corresponding numerical importance scores; Starting from the end of the sorted list, remove triples one by one, and recalculate the number of remaining triples after each removal. When the number of remaining triples first meets the condition of falling within the threshold range, stop the removal operation. At this time, the remaining triples constitute the filtered set. The general-purpose large language model is mainly used for application, but vertical domain large models can also be selected for triple extraction and filtering.

[0009] As a further aspect of the present invention, the knowledge graph structure is combined with numerical features from the original dataset and input into a graph convolutional network model, including: Extract the historical observation value sequence of the data entity corresponding to each node in the knowledge graph structure from the original dataset to form the original feature vector of the node; The original feature vector of each node is standardized to eliminate the influence of units and generate standardized node features. The normalized features of each node are used as the initial feature vector of the node in the knowledge graph structure, and together with the knowledge graph structure describing the node connection relationship, they are used as the input of the graph convolutional network.

[0010] As a further aspect of the present invention, the training of the graph convolutional network model based on a preset combination of hyperparameters, and the automatic evaluation and optimization, includes: Define a set of hyperparameter configuration combinations for the graph convolutional network to be evaluated, wherein the hyperparameter configuration combinations include graph convolutional layer dimension, graph convolutional network depth, activation function type, and model training learning rate; A portion of the original dataset is used as training data, and the other portion is used as validation data; Using the training data, a graph convolutional network model instance is independently initialized and trained under each graph convolutional network hyperparameter configuration combination. Each graph convolutional network model instance calculates the predicted value through forward propagation and uses the mean squared error between the predicted value and the true value as the loss function for backpropagation optimization. After training is completed, the prediction performance of each graph convolutional network model instance is evaluated using the validation data, and the evaluation metric values ​​of each graph convolutional network model instance on the validation data are recorded. Compare the evaluation index values ​​of all graph convolutional network model instances, select the graph convolutional network hyperparameter configuration combination corresponding to the graph convolutional network model instance with the best evaluation index value as the preferred hyperparameter configuration, and determine the graph convolutional network model instance as the optimized graph convolutional network aging prediction model.

[0011] As a further aspect of the present invention, using the training data, a graph convolutional network model instance is independently initialized and trained for each combination of graph convolutional network hyperparameter configurations. Each graph convolutional network model instance calculates the predicted value through forward propagation, and performs backpropagation optimization using the mean squared error between the predicted value and the true value as the loss function, including the following steps: For each combination of graph convolutional network hyperparameter configurations, an independent graph convolutional network model instance is initialized, including defining the number of layers and the number of neurons per layer according to the graph convolutional layer dimension and graph convolutional network depth, and configuring the activation function of each layer according to the activation function type. From the data composed of the knowledge graph structure and the standardized node features, load the subgraph and node features corresponding to the training data; The subgraphs and node features corresponding to the training data are input into the initialized graph convolutional network model instance; The model instance performs a forward propagation process, where the graph convolution operation of each layer aggregates the features of neighboring nodes and performs nonlinear transformations through activation functions. Finally, the aging prediction value of each node is calculated in the output layer. The aging prediction value obtained by forward propagation is compared with the corresponding real aging label value in the training data, and the mean square error between the aging prediction value and the real aging label value is calculated as the loss value. Using the backpropagation algorithm, the gradients of the parameters of each layer of the model are calculated based on the loss value; Based on the model training learning rate, all parameters in the graph convolutional network model instance are updated using the gradient descent optimizer; Repeat the process of forward propagation to calculate the predicted value, calculate the loss value, and backpropagation to update the parameters until the preset number of training iterations is completed or the loss value no longer decreases significantly.

[0012] As a further aspect of the present invention, the step of selecting the graph convolutional network hyperparameter configuration combination corresponding to the graph convolutional network model instance with the optimal evaluation index value as the preferred hyperparameter configuration, and determining the graph convolutional network model instance as the optimized graph convolutional network aging prediction model, includes: After evaluating all hyperparameter configuration combinations of graph convolutional networks, the evaluation metric values ​​of all graph convolutional network model instances on the validation data are sorted. Find the graph convolutional network model instance with the smallest evaluation index value, and the corresponding graph convolutional network hyperparameter configuration combination is the preferred hyperparameter configuration. Save the network structure and trained parameter weights of the graph convolutional network model instance with the preferred hyperparameter configuration. The saved model is the optimized graph convolutional network aging prediction model.

[0013] As a further aspect of the present invention, after generating the optimized graph convolutional network aging prediction model, the method further includes a full-process iterative optimization step: Without changing the initial set of data entities, the complete process from generating triples from the input of the general large language model to constructing the knowledge graph structure, and then to training and optimizing the graph convolutional network model is re-executed; In each iteration, due to the different triples generated by the general large language model, knowledge graph structures with different connectivity relationships are constructed, and different graph convolutional network model candidates are obtained by training based on different knowledge graph structures. Set a maximum number of iterations, and repeat the entire process until the maximum number of iterations is reached; Among the multiple graph convolutional network model candidates generated after all iterations, the model with the best prediction performance on the independent validation set is selected as the final graph convolutional network aging prediction model for large model empowerment.

[0014] As a further aspect of the present invention, selecting the model with the best prediction performance on the independent validation set from the multiple graph convolutional network model candidates generated after all iterations includes: Before the entire process begins, a portion of the data is pre-divided from the original dataset as a fixed independent validation set, which does not participate in the model training process in any iteration. After each iteration is completed and a graph convolutional network model candidate is obtained, the performance of the graph convolutional network model candidate is immediately evaluated using the fixed independent validation set, and the evaluation results are recorded. Once the set maximum number of iterations is reached and all iterations are completed, compare the evaluation results of all records. The graph convolutional network model candidate with the smallest evaluation result is selected as the final graph convolutional network aging prediction model empowered by the large model.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: The initial set of data entities extracted from the column names of the original dataset is input into a preset general-purpose large language model. Based on the domain knowledge within the model, semantic analysis and relational reasoning are performed on each data entity, automatically generating an initial set of triples representing entity associations. The initial set of triples is then validated and filtered to remove triples that do not meet the reasonableness conditions and quantity constraints, resulting in a refined set of triples. This allows for the automatic derivation of data entity associations, avoiding biases caused by manual semantic reasoning. The generation standards for triples remain consistent, and unreasonable and excessive triples are accurately filtered out. The pre-processing stage of knowledge graph construction is free from manual intervention.

[0016] A knowledge graph structure is constructed using data entities in the triplet set as nodes and relationships between entities as connecting edges. This knowledge graph structure is then combined with the numerical features of the original dataset and input into a graph convolutional network model. Model training is carried out based on hyperparameter configurations, and automatic evaluation and optimization operations are performed simultaneously to generate an optimized graph convolutional network aging prediction model. This achieves joint input of knowledge features and numerical features, fully supplementing the input feature dimensions of the graph convolutional network model. The entire process of model training, performance evaluation, and model selection is completed autonomously without manual intervention for adjustment and selection. The overall construction process of the aging prediction model forms a fully automated operation mode. Attached Figure Description

[0017] Figure 1 The flowchart illustrates an automated construction method for a graph convolutional network aging prediction model empowered by a large model, as described in this invention. Figure 2 A flowchart for the initial triplet set verification and screening process; Figure 3 The optimal GCN aging prediction model training and validation loss convergence curve; Figure 4 A performance trend chart for the multi-round iterative optimization model; Figure 5 For full-process iterative optimization - model prediction of performance change trends. Detailed Implementation

[0018] See Figure 1 An automated method for constructing a graph convolutional network aging prediction model empowered by a large model is as follows: Obtain the original dataset related to aging prediction, which contains multiple data columns and their corresponding column names. Based on the column names in the original dataset, automatically extract all or some column names, defining each extracted column name as a data entity, forming a preliminary data entity set. Input the preliminary data entity set into a preset general-purpose language model, driving the model to perform semantic analysis and relational reasoning on each data entity in the preliminary data entity set based on its internally learned domain knowledge, automatically generating a preliminary set of triples describing the relationships between data entities. Perform a validation and filtering process on the preliminary triple set, removing triples that do not meet the set of reasonableness conditions or exceed the preset quantity constraints, obtaining a refined triple set. Using the refined triple set, with the data entities contained within as nodes and the relationships between data entities as connecting edges, construct a knowledge graph structure. The knowledge graph structure is combined with numerical features from the original dataset and input into the graph convolutional network model. The graph convolutional network model is trained based on the preset hyperparameter configuration combination, and automatic evaluation and optimization are performed to generate an optimized graph convolutional network aging prediction model.

[0019] In one embodiment of the present invention, a preliminary set of data entities and instructions specifying the entity relationship discovery task are injected into the input prompts of the general large language model. The general large language model is guided to infer and describe the physical, chemical, or statistical relationships between any two or more data entities in the preliminary data entity set, based on its internally learned domain knowledge. The natural language descriptions output by the general large language model are received, and each natural language statement describing the relationship between entities is structured, parsed, and transformed according to the format of "head entity - relationship description - tail entity," generating a uniformly formatted triple. All generated triples are collected to form a preliminary triple set.

[0020] In the specific implementation, the original dataset related to aging prediction is obtained. The original dataset contains multiple data columns and their corresponding column names. Based on the column names of the original dataset, all column names are automatically extracted, and each extracted column name is defined as a data entity, forming a preliminary data entity set. The preliminary data entity set contains five data entities: "operating temperature," "voltage fluctuation," "vibration frequency," "relative humidity," and "capacitance decay." The input prompt for the general large language model injects the preliminary data entity set and the instruction for the specified entity relationship discovery task. The input instruction text is: "Based on your knowledge in the fields of physics and electrical engineering, analyze and list the possible direct relationships between the following entities: 'operating temperature,' 'voltage fluctuation,' 'vibration frequency,' 'relative humidity,' and 'capacitance decay.'" This guides the general large language model to infer and describe the physical, chemical, or statistical relationships between any two or more data entities in the preliminary data entity set, based on its internally learned domain knowledge. The general large language model performs inference based on its understanding of the aging mechanism of electronic components. It receives natural language descriptions output by a general large language model. The natural language descriptions include statements such as "Higher operating temperatures will accelerate the drying of the electrolyte in the capacitor, which may lead to capacitance decay", "Excessive ambient humidity may cause metal ion migration, affecting the insulation performance of the capacitor and indirectly leading to capacitance decay", and "Voltage fluctuations will generate additional current stress, which, together with the operating temperature, will accelerate the aging of the capacitor".

[0021] In specific implementations, each natural language statement describing the relationship between entities is structured and parsed according to the format of "head entity - relation description - tail entity," generating a uniformly formatted triple. For the natural language description "Higher operating temperatures accelerate the drying of the capacitor's electrolyte, which may lead to capacitance decay," the parsing process identifies the head entity "operating temperature," the relation description "accelerated electrolyte drying leads to," and the tail entity "capacitance decay," thus generating the triple (operating temperature, accelerated electrolyte drying leads to, capacitance decay). For the statement "Voltage fluctuations generate additional current stress, which, together with operating temperature, exacerbates capacitor aging," the parsing generates the triple (voltage fluctuations, exacerbated by working temperature, capacitance decay). In some embodiments, the structured parsing and conversion process is implemented through predefined relation keyword matching and syntactic analysis rules, mapping free text to predefined relation description templates. It can be understood that in other embodiments, the parsing process may introduce a relation confidence scoring function to quantitatively evaluate the conversion quality from natural language to structured triples. This scoring function considers the frequency of the relational predicate in the domain corpus and the grammatical dependency distance between the head and tail entities in the sentence. Its formal expression is as follows:

[0022] in: Represents the relations in the generated triples Confidence score, In the original natural language description sentence, the head entity is represented. With tail entity Word distance between corresponding words The total number of words in the sentence. Representation of relation description Frequency of occurrence in domain literature corpora and To adjust the model's weighting coefficients for syntactic tightness and domain priors, all triples generated through parsing, regardless of their confidence scores, are aggregated at this stage to form a preliminary triple set. Optionally, the preliminary triple set is stored in memory or a file as a list or graph structure for subsequent validation and filtering steps.

[0023] In one embodiment of the present invention, see [reference] Figure 2 The process involves calculating the total number of triples in the initial triplet set. A threshold range for the number of triples is set, related to the original data size and feature complexity. The total number of triples is then compared to this threshold range. When the total number of triples exceeds the upper limit of the threshold range, the generalized large language model is activated to score the relational importance of each triple in the initial triplet set. Based on the ranking of relational importance scores, the triples with the lowest scores are removed until the number of remaining triples falls within the threshold range. When the total number of triples is below the lower limit of the threshold range or within the range, the generalized large language model is activated to perform a logical rationality check on the initial triplet set, removing triples that the generalized large language model determines to have logical contradictions, contradict physical laws, or are weakly correlated with the aging prediction task.

[0024] Each triple in the initial triplet set is input into the general large language model, along with instructions to evaluate the importance of the relation described by the triplet to the aging prediction task. The general large language model outputs a numerical importance score for each input triplet. All triples in the initial triplet set are sorted from highest to lowest according to their corresponding numerical importance scores. Starting from the end of the sorted list, triples are removed sequentially, and the number of remaining triples is recalculated after each removal. The removal operation stops when the number of remaining triples first falls within a threshold range; the remaining triples at this point constitute the filtered set.

[0025] In practical implementation, the process of performing verification and screening on the initial triplet set and obtaining the refined triplet set in an automated construction method for graph convolutional network aging prediction models empowered by large models is defined in detail. The system calculates the total number of triplets in the initial triplet set. In practice, the system sets a threshold range for the number of triples that is related to the size and complexity of the original data. The system will count the total number of triples. With quantity threshold range In comparison, the comparison operation triggers different subsequent processing branches in practice. This depends on the total number of triples. Exceeding the upper limit of the quantity threshold range At this point, the system activates the general large language model to score the relation importance of each triple in the initial triple set. This can be understood as the system removing the triples with the lowest scores based on the ranking of relation importance scores, until the number of remaining triples falls within a threshold range. When the total number of triples Below the lower limit of the quantity threshold range If the condition is within the specified range, the system activates the general large language model to perform a logical rationality check on the initial triplet set. Triplets that are judged by the general large language model to have logical contradictions, do not conform to physical laws, or are weakly related to the aging prediction task are removed. Under this branch, the triplets that are checked and retained directly constitute the refined triplet set.

[0026] In some embodiments, the specific process of activating the general large language model to score the relational importance of the initial set of triples includes the following steps. The system will assign each triple from the initial set of triples... The inputs are fed into a general large language model, which means the system appends instructions to the input to evaluate the importance of the relationships described by the triples to the aging prediction task. The system receives triples from the general large language model for each input. Output a numerical importance score In practice, the system scores all triples in the initial triplet set according to their importance in numerical form. Sort the data from highest to lowest to generate an ordered list. The system starts from an ordered list. Starting from the end, remove triples one by one, and recalculate the number of remaining triples after each removal. When the number of remaining triples The first time the threshold range of the number of falls is met When the condition is met, the removal operation stops, and the remaining triples form the filtered set, i.e., the triple set. Optionally, the removal process satisfies the relation ,in The number of triples removed; the removal operation continues until the condition is met. The needs are met.

[0027] In one embodiment of the present invention, historical observation numerical sequences of data entities corresponding to each node in the knowledge graph structure are extracted from the original dataset to form the original feature vector of the node. The original feature vector of each node is standardized to eliminate the influence of dimensions, generating standardized node features. The standardized node features of each node are used as the initial feature vector of the node in the knowledge graph structure, and together with the knowledge graph structure describing the node connections, are used as the input to the graph convolutional network.

[0028] A set of hyperparameter configurations for graph convolutional networks (GCNNs) to be evaluated is defined. These configurations include the dimension of the GCNN layers, the depth of the GCNN, the type of activation function, and the model training learning rate. A portion of the original dataset is used as training data, and the other portion as validation data. Using the training data, a GCNN model instance is independently initialized and trained for each hyperparameter configuration. Each GCNN model instance calculates its prediction value using forward propagation, and uses the mean squared error between the predicted and true values ​​as the loss function for backpropagation optimization. After training, the prediction performance of each GCNN model instance is evaluated using the validation data, and the evaluation metric values ​​for each GCNN model instance on the validation data are recorded. The evaluation metric values ​​for all GCNN model instances are compared, and the hyperparameter configuration corresponding to the GCNN model instance with the best evaluation metric value is selected as the preferred hyperparameter configuration. This GCNN model instance is then designated as the optimized GCNN aging prediction model.

[0029] In a specific implementation, a method for automatically constructing a graph convolutional network aging prediction model empowered by a large model is defined in detail. This method combines the knowledge graph structure with the numerical features of the original dataset and trains and evaluates the graph convolutional network model. Historical observation numerical sequences of data entities corresponding to each node in the knowledge graph structure are extracted from the original dataset to form the original feature vector of the node. In practice, the original feature vector of each node is standardized to eliminate the influence of dimensions, generating standardized node features. These standardized features, as the initial feature vector of the node in the knowledge graph structure, are used as input to the graph convolutional network along with the knowledge graph structure itself. In some embodiments, the standardization process uses the Z-score standardization method, which involves calculating the mean and standard deviation of each dimension of the original node feature vector and performing a linear transformation. Table 1 shows a simplified example of node feature standardization, including a comparison of the feature values ​​of three data entities before and after standardization.

[0030] Table 1: Examples of Node Feature Standardization Processing

[0031] A set of hyperparameter configurations for graph convolutional networks (GCNNs) to be evaluated is defined. These configurations include the dimension of the GCNN layers, the depth of the GCNN, the type of activation function, and the model training learning rate. A portion of the original dataset is used as training data, and the other portion as validation data. Using the training data, a GCNN model instance is independently initialized and trained for each GCNN hyperparameter configuration. Each GCNN model instance calculates its prediction value using forward propagation and optimizes it using the mean squared error between the predicted and true values ​​as the loss function. After training, the prediction performance of each GCNN model instance is evaluated using the validation data, and the evaluation metric values ​​for each GCNN model instance on the validation data are recorded. The evaluation metric values ​​of all GCNN model instances are compared, and the GCNN hyperparameter configuration corresponding to the GCNN model instance with the optimal evaluation metric value is selected as the preferred hyperparameter configuration. This GCNN model instance is then designated as the optimized GCNN aging prediction model. Optionally, in graph convolutional network regression prediction tasks, the optimal evaluation metric value usually refers to the minimum mean squared error value.

[0032] In some embodiments, the mean squared error metric used to evaluate the predictive performance of a graph convolutional network model instance on validation data is calculated using the following formula:

[0033] Where: symbol This represents the mean squared error evaluation index calculated on the validation dataset, with the symbol... Represents the total number of sample nodes in the validation dataset, symbol Represents the first in the validation dataset Each sample node, symbol Represents sample nodes The corresponding true aging label value, symbol This represents a graph convolutional network model instance for sample nodes. The calculated aging prediction values. It can be understood that the summation operation is performed on all sample nodes in the validation dataset.

[0034] See Figure 3During the training and validation of the optimal GCN aging prediction model, the loss curves visually reflect the model's convergence performance and generalization ability. The vertical axis represents the mean squared error loss (MSE), used to quantify the deviation between the model's predicted values ​​and the true aging label values; the horizontal axis represents the number of training iterations (Epochs), recording the complete process of model parameter updates. The dashed line represents the training set loss, which shows a continuous downward trend, gradually converging from approximately 0.79 initially to near 0, indicating that the model has fully learned the feature patterns on the training data and has not exhibited underfitting. The solid line represents the validation set loss, which, although showing slight fluctuations, generally decreases in sync with the training set loss, eventually converging to a low level similar to the training set loss, indicating that the model has good generalization ability on unknown data and has not experienced significant overfitting. Both curves stabilize in the low loss range in the later stages of iteration, verifying the effectiveness of the GCN model under this optimized hyperparameter configuration and its ability to support high-precision aging prediction tasks.

[0035] In one embodiment of the present invention, for each combination of graph convolutional network hyperparameter configurations, an independent graph convolutional network model instance is initialized. This includes defining the number of layers and the number of neurons per layer based on the graph convolutional layer dimension and graph convolutional network depth, and configuring the activation function for each layer according to the activation function type. Subgraphs and node features corresponding to the training data are loaded from the data consisting of the knowledge graph structure and standardized node features. The subgraphs and node features corresponding to the training data are input into the initialized graph convolutional network model instance. The model instance performs a forward propagation process, where the graph convolutional operations of each layer aggregate the features of neighboring nodes and perform nonlinear transformations through activation functions, finally calculating the aging prediction value for each node at the output layer. The aging prediction value calculated by forward propagation is compared with the corresponding true aging label value in the training data, and the mean squared error between the aging prediction value and the true aging label value is calculated as the loss value. The gradient of the parameters of each layer of the model is calculated using the backpropagation algorithm based on the loss value. Based on the model training learning rate, all parameters in the graph convolutional network model instance are updated using a gradient descent optimizer. Repeat the process of forward propagation to calculate predicted values, calculating loss values, and backpropagation to update parameters until the preset number of training iterations is completed or the loss value no longer decreases significantly. After evaluating all graph convolutional network (GCNN) hyperparameter configurations, sort all GCNN model instances on the validation data based on their evaluation metric values. Identify the GCNN model instance with the smallest evaluation metric value; its corresponding GCNN hyperparameter configuration is the optimal hyperparameter configuration. Save the network structure and trained parameter weights of the GCNN model instance with the optimal hyperparameter configuration; this saved model is the optimized GCNN aging prediction model.

[0036] In practical implementation, the training details and model selection process of the graph convolutional network (GCNN) model instance are defined in detail in an automated construction method for a large-scale model-enabled graph convolutional network aging prediction model. For each combination of graph convolutional network hyperparameters, an independent graph convolutional network model instance is initialized. In practice, the initialization operation includes defining the number of layers and the number of neurons per layer based on the graph convolutional layer dimension and graph convolutional network depth, and configuring the activation function for each layer according to the activation function type. From the data composed of knowledge graph structure and standardized node features, the subgraphs and node features corresponding to the training data are loaded. It can be understood that the subgraphs corresponding to the training data are a subset of the complete knowledge graph structure. The subgraphs and node features corresponding to the training data are input into the initialized graph convolutional network model instance. The graph convolutional network model instance performs a forward propagation process. The graph convolutional operations of each layer aggregate the features of neighboring nodes and perform nonlinear transformations through activation functions. Finally, the aging prediction value of each node is calculated at the output layer. The aging prediction value calculated by forward propagation is compared with the corresponding true aging label value in the training data, and the mean squared error between the aging prediction value and the true aging label value is calculated as the loss value. Using the backpropagation algorithm, the gradients of the parameters of each layer of the graph convolutional network model instance are calculated based on the loss value. Based on the model training learning rate, all parameters of the graph convolutional network model instance are updated using the gradient descent optimizer. The process of forward propagation to calculate the prediction value, calculating the loss value, and backpropagation to update the parameters is repeated until the preset number of training iterations is completed or the loss value no longer decreases significantly.

[0037] After evaluating all hyperparameter configurations of graph convolutional networks (GCNNs), the evaluation metrics of all GCNN model instances on the validation data are sorted. The GCNN model instance with the smallest evaluation metric value is identified, and its corresponding hyperparameter configuration is the optimal hyperparameter configuration. The network structure and trained parameter weights of the GCNN model instances with the optimal hyperparameter configuration are saved; this saved model is the optimized GCNN aging prediction model. It can be understood that the parameter update process is performed after each backpropagation. Refer to Table 2, which shows a simplified example of model evaluation and selection, comparing the evaluation results of GCNN model instances under different hyperparameter configurations.

[0038] Table 2: Example Table of Hyperparameter Configuration Evaluation Results for Graph Convolutional Networks

[0039] In specific implementation, the mean squared error of the validation set for all configurations in Table 2 is compared, and the graph convolutional network model instance with the smallest evaluation metric value is identified. In Table 2, the mean squared error of the validation set for configuration HP-2, at 0.038, is the smallest. Therefore, the graph convolutional network hyperparameter configuration combination corresponding to configuration HP-2 (graph convolutional layer dimension 128, graph convolutional network depth 2, activation function ReLU, model training learning rate 0.001) is selected as the preferred hyperparameter configuration. Optionally, saving the graph convolutional network model instance with the preferred hyperparameter configuration means persistently storing the model structure and trained parameter weights corresponding to configuration HP-2 to a disk file. In some embodiments, the criterion for the loss value no longer decreasing significantly is that the change in the loss value over multiple consecutive training epochs is less than a preset minimum positive threshold.

[0040] See Figure 4 During the full-process iterative optimization phase, model performance was evaluated using the validation set mean squared error as the core metric, showing a convergence trend that continuously decreased with increasing iterations. Specifically, from the 1st to the 8th iteration, the validation set mean squared error gradually decreased from an initial 0.042 to 0.032, with the 8th iteration being marked as the optimal iteration, corresponding to the validation set mean squared error reaching its minimum value throughout the entire process. Although a slight rebound occurred in the 5th iteration (mean squared error of 0.036), the overall trend remained stable and declining. This reflects the continuous enhancement of the generalization ability of the graph convolutional network aging prediction model through dynamic adjustment of the knowledge graph structure and model hyperparameter configuration. The trend intuitively verifies the effectiveness of the full-process iterative optimization strategy: under the premise of a fixed initial data entity set, different triples are generated through a large language model to construct a differentiated knowledge graph, thereby training multiple sets of model candidates, and performance optimization is completed based on independent validation sets, ultimately achieving significant convergence and optimization of model prediction error.

[0041] In one embodiment of the invention, without changing the initial data entity set, the complete process from generating triples from the input general-purpose large language model to constructing the knowledge graph structure, and then training and optimizing the graph convolutional network model, is re-executed. In each iteration, because the triples generated by the general-purpose large language model are different, knowledge graph structures with different connectivity relationships are constructed, and different graph convolutional network model candidates are obtained based on different knowledge graph structures. A total iteration limit is set, and the complete process is repeated until the iteration limit is reached. Among the multiple graph convolutional network model candidates generated after all iterations, the model with the best prediction performance on the independent validation set is selected as the final determined graph convolutional network aging prediction model for large model empowerment.

[0042] Before the entire process begins, a portion of the original dataset is pre-defined as a fixed independent validation set. This validation set does not participate in any iteration of the model training process. After each iteration, once a graph convolutional network (GCNN) model candidate is obtained, its performance is immediately evaluated using the fixed independent validation set, and the evaluation results are recorded. Once the set iteration limit is reached and all iterations are completed, all recorded evaluation results are compared. The GCNN model candidate with the smallest evaluation result is selected as the final GCNN aging prediction model for large-scale model empowerment.

[0043] In its implementation, the automated construction method for a graph convolutional network (GCNN) aging prediction model empowered by a large model includes a full-process iterative optimization step after generating the optimized GCNN aging prediction model. Without changing the initial data entity set, it re-executes the complete process from generating triples from the input general large language model to constructing the knowledge graph structure, and then training and optimizing the GCNN model. In each iteration, because the triples generated by the general large language model are different, knowledge graph structures with different connectivity relationships are constructed. Different GCNN model candidates are obtained by training based on different knowledge graph structures. A total upper limit on the number of iterations is set. Repeating the entire process until the maximum number of iterations is reached. In specific implementation, symbols This indicates the maximum preset number of iterations. Among the multiple graph convolutional network model candidates generated after all iterations, the model with the best prediction performance on the independent validation set is selected as the final graph convolutional network aging prediction model for large-scale model empowerment.

[0044] In some embodiments, the process of selecting the model with the best prediction performance on the independent validation set includes the following steps: Before the entire process begins, a portion of the data is pre-allocated from the original dataset as a fixed independent validation set. This independent validation set does not participate in the model training process during any iteration. After each iteration is completed and a candidate graph convolutional network model is obtained, the performance of the candidate model is immediately evaluated using the fixed independent validation set, and the evaluation results are recorded. When the set maximum number of iterations is reached... After all iterations are completed, the evaluation results of all records are compared. The graph convolutional network (GCNN) model candidate with the smallest evaluation result is selected as the final GCNN aging prediction model for large-scale model empowerment. It can be understood that the independent validation set remains unchanged throughout the entire iteration process, providing a consistent and fair performance comparison benchmark for the GCNN model candidates generated in all iteration rounds. Optionally, the split ratio of the independent validation set is a fixed percentage of the total original dataset. In practice, each iteration is an independent automated construction process, starting with the same initial set of data entities. However, due to the inherent randomness or diversity of the triple generation process of the general large language model, the initial triple set generated in each iteration will differ, leading to the diversity of subsequent knowledge graph structures and trained GCNN model candidates. It can be understood that the full-process iterative optimization steps utilize the diversity of the general large language model's generation results, exploring different potential relational structures by running the complete process multiple times, and selecting the optimal model from multiple candidate models.

[0045] See Figure 5 In the full-process iterative optimization stage of the graph convolutional network aging prediction model empowered by a large model, the trend of model prediction performance with each iteration can be quantitatively analyzed using two core indicators: mean squared error (MSE) and mean absolute error (MAE). Specifically, during the evolution from iteration 1 to 10, both error indicators showed a continuous decline and gradual convergence: Mean Squared Error (MSE): The initial value was approximately 2.85, which decreased rapidly with each iteration, converging to approximately 1.48 after the 8th iteration, with minimal subsequent fluctuations, indicating that the squared deviation between the model's predicted values ​​and the actual values ​​was effectively controlled, and the prediction stability was significantly improved. Mean Absolute Error (MAE): The initial value was approximately 1.75, which showed a generally stable downward trend, dropping to approximately 0.84 in the 8th iteration. Although there was a slight rebound in the 9th iteration, it finally fell back to approximately 0.87 in the 10th iteration, indicating that the absolute deviation of the model's prediction continued to narrow, and the prediction accuracy was gradually optimized. From the perspective of error evolution logic, the downward trend of the two indicators is highly consistent with the full-process iterative optimization mechanism: each iteration generates different triples through the large language model, constructing diverse knowledge graph structures, providing richer relational features and topological information for the graph convolutional network model, prompting the model to continuously learn better aging prediction patterns in multiple rounds of training, ultimately achieving error convergence and performance stability. At the parameter level, a fixed partition of the independent validation set is maintained during the iteration process, ensuring the fairness and comparability of the performance evaluation of candidate models in each round, providing a reliable basis for selecting the optimal model.

Claims

1. An automated construction method for aging prediction models of graph convolutional networks empowered by large models, characterized in that, The method includes: Obtain the raw dataset related to aging prediction, which contains multiple data columns and the column name corresponding to each data column; Based on the column names of the original dataset, all or some column names are automatically extracted, and each extracted column name is defined as a data entity to form a preliminary data entity set. The preliminary set of data entities is input into a preset general-purpose large language model, which drives the general-purpose large language model to perform semantic analysis and relational reasoning on each data entity in the preliminary set of data entities based on the domain knowledge it has learned internally, and automatically generates a preliminary set of triples describing the relationship between the data entities. A verification and screening process is performed on the preliminary triplet set to remove triplets that do not meet the set reasonableness conditions or exceed the preset quantity constraint range, thereby obtaining the refined triplet set. Using the aforementioned set of triplets, with the data entities contained therein as nodes and the relationships between the data entities as connecting edges, a knowledge graph structure is constructed. The knowledge graph structure is combined with numerical features from the original dataset and input into the graph convolutional network model. The graph convolutional network model is trained based on a preset combination of hyperparameters and automatically evaluated and optimized to generate an optimized graph convolutional network aging prediction model.

2. The automated construction method for aging prediction model of graph convolutional network empowered by large model as described in claim 1, characterized in that, The general-purpose large language model, based on its internally learned domain knowledge, performs semantic analysis and relational reasoning on each data entity in the preliminary data entity set, automatically generating a preliminary set of triples describing the relationships between data entities, including: The initial set of data entities and instructions for the entity relationship discovery task are injected into the input prompts of the general large language model. The general large language model is guided to infer and describe the physical, chemical, or statistical relationships between any two or more data entities in the preliminary data entity set, based on the domain knowledge it has learned internally. The system receives the natural language descriptions output by the general large language model, and performs structured parsing and transformation on each natural language statement describing the relationship between entities according to the format of "head entity-relationship description-tail entity" to generate a triple with a uniform format. All generated triples are collected to form the preliminary triple set.

3. The automated construction method for aging prediction model of graph convolutional network empowered by large model as described in claim 2, characterized in that, The initial set of triples is subjected to a verification and screening process to remove triples that do not meet the set of reasonableness conditions or exceed the preset quantity constraints, thereby obtaining the refined set of triples, including: Calculate the total number of triplets in the initial set of triplets; Set a threshold range for the number of triples related to the size of the original data and the complexity of the features, and compare the total number of triples with the threshold range; When the total number of triples exceeds the upper limit of the number threshold range, the general large language model is activated to score the relation importance of each triple in the initial triple set. Based on the ranking result of the relation importance scores, the triples with the lowest scores are removed until the number of remaining triples falls within the number threshold range. When the total number of triples is below the lower limit of the number threshold range or within the range, the general big language model is activated to perform a logical rationality check on the preliminary triple set, and triples that are determined by the general big language model to have logical contradictions, do not conform to physical laws, or are weakly related to the aging prediction task are removed.

4. The automated construction method for aging prediction model of graph convolutional network empowered by large model as described in claim 3, characterized in that, The generalized large language model is activated to score the relation importance of each triple in the initial triple set. Based on the ranking of relation importance scores, the triples with the lowest scores are removed, including: Each triple in the initial set of triples is input into the general large language model, and an instruction is attached to evaluate the importance of the relationship described by the triples to the aging prediction task. Receive a numerical importance score from the output of the general large language model for each input triple; Sort all triples in the initial triplet set from highest to lowest according to their corresponding numerical importance scores; Starting from the end of the sorted list, remove triples one by one, and recalculate the number of remaining triples after each removal. When the number of remaining triples first meets the condition of falling within the threshold range, stop the removal operation. At this time, the remaining triples constitute the filtered set. The general-purpose large language model is mainly used for application, but vertical domain large models can also be selected for triple extraction and filtering.

5. The automated construction method for aging prediction model of graph convolutional network empowered by large model as described in claim 4, characterized in that, The knowledge graph structure is combined with numerical features from the original dataset and input into a graph convolutional network model, including: Extract the historical observation value sequence of the data entity corresponding to each node in the knowledge graph structure from the original dataset to form the original feature vector of the node; The original feature vector of each node is standardized to eliminate the influence of units and generate standardized node features. The normalized features of each node are used as the initial feature vector of the node in the knowledge graph structure, and together with the knowledge graph structure describing the node connection relationship, they are used as the input of the graph convolutional network.

6. The method for automatically constructing a graph convolutional network aging prediction model empowered by a large model as described in claim 5, characterized in that, The process of training the graph convolutional network model based on a preset combination of hyperparameters, and then automatically evaluating and optimizing it, includes: Define a set of hyperparameter configuration combinations for the graph convolutional network to be evaluated, wherein the hyperparameter configuration combinations include graph convolutional layer dimension, graph convolutional network depth, activation function type, and model training learning rate; A portion of the original dataset is used as training data, and the other portion is used as validation data; Using the training data, a graph convolutional network model instance is independently initialized and trained under each graph convolutional network hyperparameter configuration combination. Each graph convolutional network model instance calculates the predicted value through forward propagation and uses the mean squared error between the predicted value and the true value as the loss function for backpropagation optimization. After training is completed, the prediction performance of each graph convolutional network model instance is evaluated using the validation data, and the evaluation metric values ​​of each graph convolutional network model instance on the validation data are recorded. Compare the evaluation index values ​​of all graph convolutional network model instances, select the graph convolutional network hyperparameter configuration combination corresponding to the graph convolutional network model instance with the best evaluation index value as the preferred hyperparameter configuration, and determine the graph convolutional network model instance as the optimized graph convolutional network aging prediction model.

7. The method for automatically constructing a graph convolutional network aging prediction model empowered by a large model as described in claim 6, characterized in that, Using the training data, for each combination of graph convolutional network hyperparameters, an independent graph convolutional network model instance is initialized and trained. Each graph convolutional network model instance calculates the predicted value through forward propagation, and performs backpropagation optimization using the mean squared error between the predicted and true values ​​as the loss function. The process includes the following steps: For each combination of graph convolutional network hyperparameter configurations, an independent graph convolutional network model instance is initialized, including defining the number of layers and the number of neurons per layer according to the graph convolutional layer dimension and graph convolutional network depth, and configuring the activation function of each layer according to the activation function type. From the data composed of the knowledge graph structure and the standardized node features, load the subgraph and node features corresponding to the training data; The subgraphs and node features corresponding to the training data are input into the initialized graph convolutional network model instance; The model instance performs a forward propagation process, where the graph convolution operation of each layer aggregates the features of neighboring nodes and performs nonlinear transformations through activation functions. Finally, the aging prediction value of each node is calculated in the output layer. The aging prediction value obtained by forward propagation is compared with the corresponding real aging label value in the training data, and the mean square error between the aging prediction value and the real aging label value is calculated as the loss value. Using the backpropagation algorithm, the gradients of the parameters of each layer of the model are calculated based on the loss value; Based on the model training learning rate, all parameters in the graph convolutional network model instance are updated using the gradient descent optimizer; Repeat the process of forward propagation to calculate the predicted value, calculate the loss value, and backpropagation to update the parameters until the preset number of training iterations is completed or the loss value no longer decreases significantly.

8. The method for automatically constructing a graph convolutional network aging prediction model empowered by a large model as described in claim 7, characterized in that, The selection of the graph convolutional network hyperparameter configuration combination corresponding to the graph convolutional network model instance with the optimal evaluation index value is taken as the preferred hyperparameter configuration, and the graph convolutional network model instance is determined as the optimized graph convolutional network aging prediction model, including: After evaluating all hyperparameter configuration combinations of graph convolutional networks, the evaluation metric values ​​of all graph convolutional network model instances on the validation data are sorted. Find the graph convolutional network model instance with the smallest evaluation index value, and the corresponding graph convolutional network hyperparameter configuration combination is the preferred hyperparameter configuration. Save the network structure and trained parameter weights of the graph convolutional network model instance with the preferred hyperparameter configuration. The saved model is the optimized graph convolutional network aging prediction model.

9. The method for automatically constructing a graph convolutional network aging prediction model empowered by a large model as described in claim 8, characterized in that, After generating the optimized graph convolutional network aging prediction model, the method also includes a full-process iterative optimization step: Without changing the initial set of data entities, the complete process from generating triples from the input of the general large language model to constructing the knowledge graph structure, and then to training and optimizing the graph convolutional network model is re-executed; In each iteration, due to the different triples generated by the general large language model, knowledge graph structures with different connectivity relationships are constructed, and different graph convolutional network model candidates are obtained by training based on different knowledge graph structures. Set a maximum number of iterations, and repeat the entire process until the maximum number of iterations is reached; Among the multiple graph convolutional network model candidates generated after all iterations, the model with the best prediction performance on the independent validation set is selected as the final graph convolutional network aging prediction model for large model empowerment.

10. The method for automatically constructing a graph convolutional network aging prediction model empowered by a large model as described in claim 9, characterized in that, Among the multiple graph convolutional network model candidates generated after all iterations, the model with the best prediction performance on the independent validation set is selected, including: Before the entire process begins, a portion of the data is pre-divided from the original dataset as a fixed independent validation set, which does not participate in the model training process in any iteration. After each iteration is completed and a graph convolutional network model candidate is obtained, the performance of the graph convolutional network model candidate is immediately evaluated using the fixed independent validation set, and the evaluation results are recorded. Once the set maximum number of iterations is reached and all iterations are completed, compare the evaluation results of all records. The graph convolutional network model candidate with the smallest evaluation result is selected as the final graph convolutional network aging prediction model empowered by the large model.

Images