CN-121981175-A - Multi-interface data transmission interpolation calculation method based on KAN neural network

CN121981175ACN 121981175 ACN121981175 ACN 121981175ACN-121981175-A

Abstract

The application relates to the technical field of data processing, in particular to a multi-interface data transmission interpolation calculation method based on a KAN (Kan neural network), which aims to solve the technical problem that obvious errors exist when the KAN neural network describes internal rules among multi-interface data. The method comprises the steps of taking complete dimension data acquired by a plurality of data acquisition interfaces as a training set, training a KAN neural network, determining the optimal balance effect expression strength of the KAN neural network based on spline training balance of an intermediate layer of the KAN neural network when training data are processed in the training process, adjusting a function calling strategy of the KAN neural network in a function library in the training process based on the optimal balance effect expression strength, training the KAN neural network based on the adjusted strategy, and performing interpolation calculation on missing values in dimension data acquired by a current multi-interface based on the KAN neural network after training is completed.

Inventors

WANG SHENGMING
SONG YUNQUAN
ZHANG XINZE
HU CHAORAN
FU YINGHAO

Assignees

中国石油大学(华东)

Dates

Publication Date: 20260505
Application Date: 20260124

Claims (10)

1. A multi-interface data transfer interpolation calculation method based on a KAN neural network, the method comprising: taking the complete dimension data acquired by the plurality of data acquisition interfaces as a training set, and training the KAN neural network; In the training process, determining the optimal balance effect expression strength of the KAN neural network based on the spline training balance of the middle layer of the KAN neural network when the training data is processed, wherein the spline training balance is used for representing the balance effect of accuracy and efficiency when the middle layer carries out spline type matching training; Based on the optimal balance effect expression strength, adjusting a calling strategy of the KAN neural network for functions in a function library in a training process, and training the KAN neural network based on the adjusted strategy; And after training is completed, carrying out interpolation calculation on the missing value in the dimension data acquired by the multiple interfaces currently based on the KAN neural network after training is completed.
2. The KAN neural network-based multi-interface data transfer interpolation calculation method of claim 1, wherein determining an optimal balance effect expression level of the KAN neural network based on spline training balance of an intermediate layer of the KAN neural network when processing the training data, comprises: Determining the spline training balance based on structural limiting force generated by data transmission when the KAN neural network middle layer processes the training set and the piecewise approximation ideality of functions in a function library on the function curve to be analyzed of the middle layer, wherein the piecewise approximation ideality is used for representing the appropriateness of candidate functions to approximate the target function curve to be analyzed through the piecewise form; and based on the spline training balance of each middle layer, analyzing the transfer influence of the training effect between layers, and determining the expression strength of the optimal balance effect.
3. The KAN neural network-based multi-interface data transfer interpolation computation method of claim 2, wherein determining the spline training balance based on structural limiting forces generated by data transfer when the KAN neural network middle layer processes the training set, and segment approximation idealities of functions in a function library on the middle layer function curve to be analyzed, comprises: Determining the structure limiting force of the target intermediate layer according to the connecting structure between the target intermediate layer and the previous layer and the data dimension information of the plurality of data acquisition interfaces, wherein the structure limiting force is used for representing the limiting degree of the interface data dimension quantity of the target intermediate layer before spline type training is carried out on each connecting edge; Determining the piecewise approximation ideality of each candidate function in a function library relative to a target function curve to be analyzed based on the structure limiting force, wherein the target function curve to be analyzed is a continuous curve generated according to the mapping relation between the input end and the expected output end of the target function curve to be analyzed when training data passes through the connecting edge between the target middle layer and the upper layer; Determining function application suitability of each candidate function based on the piecewise approximation ideality of each candidate function, wherein the function application suitability is used for representing the comprehensive approximation effect of the candidate function on the function curve to be analyzed; And determining spline training balance of the target middle layer based on application suitability of each function in the function library.
4. The KAN neural network-based multi-interface data transfer interpolation calculation method of claim 3, wherein determining a structural constraint force of the target intermediate layer according to a connection structure between the target intermediate layer and a previous layer and data dimension information of the plurality of data acquisition interfaces comprises: And determining the structure limiting force according to the number of edges between the target middle layer and the previous layer, the number of functions in a function library, the total number of interfaces and the number of dimension data types transferred by each interface.
5. The KAN neural network-based multi-interface data transfer interpolation calculation method of claim 3, wherein determining the piecewise approximation ideality of each candidate function in the function library relative to the target function curve to be analyzed based on the structural constraint force comprises: determining the number of intersection points of the candidate function and the target function curve to be analyzed; calculating the area of the area surrounded by the candidate function and the target function curve to be analyzed between every two adjacent intersection points; And determining the segment approximation ideality according to the number of the intersection points, the sum of the areas and the structure limiting force, wherein the segment approximation ideality is positively correlated with the number of the intersection points and is negatively correlated with the sum of the areas and the structure limiting force.
6. The KAN neural network based multi-interface data transfer interpolation computation method of claim 3, wherein determining a function application suitability of each candidate function based on a piecewise approximation ideality of each candidate function comprises: According to the segment approximation ideality of the candidate function, segment processing is carried out on the target function curve to be analyzed, and the number of data segments corresponding to the candidate function is determined; determining the application suitability of the function according to the segment approximation ideality of the candidate function, the number of data segments and the number of functions traversed before the candidate function is called, wherein the application suitability of the function is positively related to the segment approximation ideality and negatively related to the sum of the number of data segments and the number of traversed functions.
7. The KAN neural network-based multi-interface data transfer interpolation calculation method according to claim 6, wherein the step of performing the piecewise processing on the target function curve to be analyzed according to the piecewise approximation ideality of the candidate function comprises: If the piecewise approximation ideality of the candidate function is smaller than a first preset threshold, uniformly dividing the target function curve to be analyzed into a preset number of data segments; and if the segment approximation ideality of the candidate function is larger than or equal to the first preset threshold value, carrying out cluster analysis based on the distribution of the intersection points on the target function curve to be analyzed, and determining a continuous data area covered by each cluster as a data segment.
8. The KAN neural network based multi-interface data transfer interpolation calculation method of claim 2, wherein analyzing the transfer effect of the training effect between layers based on the spline training balance of each intermediate layer, determining the optimal balance effect expression level comprises: Calculating the multi-interface transmission interpolation interference according to the number of intermediate layers deduced before the target intermediate layer and the spline training balance; And determining the optimal balance effect expression strength based on the multi-interface transmission interpolation interference, wherein the optimal balance effect expression strength is inversely related to the multi-interface transmission interpolation interference.
9. The KAN neural network-based multi-interface data transfer interpolation calculation method of claim 8, wherein calculating the multi-interface transfer interpolation interference based on the number of intermediate layers that have been deduced before the target intermediate layer and spline training balance comprises: Calculating the unsmooth degree of logic deduction according to the number of the deducted intermediate layers and the difference of spline training balance of the adjacent deducted intermediate layers; Calculating the accumulated discomfort interference force according to the number of the deduced middle layers and the spline training balance thereof; and determining the multi-interface transmission interpolation interference according to the degree of unsmooth and the degree of the non-adaptive interference.
10. The multi-interface data transfer interpolation calculation method based on the KAN neural network according to claim 1, wherein adjusting a call policy of the KAN neural network to a function in a function library in a training process based on the best balance effect expression strength comprises: if the expression strength of the optimal balance effect is greater than a first threshold value, maintaining a default calling sequence of functions in the function library to complete training; if the expression strength of the optimal balance effect is smaller than or equal to the first threshold, when the spline type is determined for the connecting edge in the KAN neural network, directly determining a function with the segment approximation ideality larger than the second threshold for the first time as the spline type corresponding to the connecting edge.

Description

Multi-interface data transmission interpolation calculation method based on KAN neural network Technical Field The application relates to the technical field of data processing, in particular to a multi-interface data transmission interpolation calculation method based on a KAN neural network. Background In big data application scenarios such as credit card risk assessment and equipment maintenance prediction, tabular dimension data from multiple data acquisition interfaces is usually required to be used as input to predict key target values, and the processing of such structured data often adopts a multi-layer perceptron neural network to learn complex nonlinear mapping preferentially. However, the decision process of the multi-layer perceptron neural Network has low transparency and presents black box characteristics, which restricts the interpretation and analysis of data decision logic by background personnel, and therefore interpolation calculation in data transmission is currently supported through a transparent reasoning process based on a Kolmogorov-Arnold neural Network (KAN). When the existing KAN neural network processes data transmitted by each interface, the type and the number of the splines in the training process can be determined in an artificial preset mode, but in the actual process, the preset mode usually causes the condition that the originally preset type and number of the splines are too much or too little, so that obvious errors exist when the KAN neural network describes the internal rules among the multi-interface data. Disclosure of Invention In order to solve the technical problem that obvious errors exist when KAN neural network describes internal rules among multi-interface data, the application aims to provide a KAN neural network-based multi-interface data transmission interpolation calculation method, and the adopted technical scheme is as follows: taking the complete dimension data acquired by the plurality of data acquisition interfaces as a training set, and training the KAN neural network; In the training process, determining the optimal balance effect expression strength of the KAN neural network based on the spline training balance of the middle layer of the KAN neural network when training data are processed, wherein the spline training balance is used for representing the balance effect of accuracy and efficiency when the middle layer carries out spline type matching training; based on the optimal balance effect expression strength, adjusting a calling strategy of the KAN neural network to functions in a function library in the training process, and training the KAN neural network based on the adjusted strategy; after training is completed, interpolation calculation is carried out on missing values in the dimension data acquired by the current multi-interface based on the trained KAN neural network. In one possible implementation, the method for determining the optimal balance effect expression level of the KAN neural network based on spline training balance of the middle layer of the KAN neural network when training data are processed comprises determining spline training balance based on structural limiting force generated by data transmission of the middle layer of the KAN neural network when a training set is processed and piecewise approximation ideality of functions in a function library on function curves to be analyzed of the middle layer, analyzing transfer influence of training effects among layers based on spline training balance of each middle layer, and determining the optimal balance effect expression level. In one possible implementation, the spline training balance is determined based on structural limiting force generated by data transfer of the KAN neural network middle layer when the training set is processed and piecewise approximation ideality of functions in the function library on a middle layer function curve to be analyzed, wherein the piecewise approximation ideality is used for representing the suitability of the target function curve to be analyzed in a piecewise manner according to a connecting structure between the target middle layer and the last layer and data dimension information of a plurality of data acquisition interfaces, the structural limiting force is used for representing the limiting degree of the target middle layer by the number of interface data dimensions before spline type training is carried out on each connecting side of the target middle layer, the piecewise approximation ideality of each candidate function in the function library relative to the target function curve to be analyzed is determined based on the structural limiting force, the target function curve to be analyzed is a continuous curve generated according to a mapping relation between an input end and an expected output end of the function library when training data passes through the connecting side between the target middle layer and the last layer, th