CN-120878002-B - Multi-index comprehensive performance evaluation method for road material

CN120878002BCN 120878002 BCN120878002 BCN 120878002BCN-120878002-B

Abstract

The invention relates to a multi-index comprehensive performance evaluation method for a road material, and belongs to the technical field of data processing. The method comprises the steps of constructing a road material multi-index data set, carrying out data marking, calculating attribute correlation weight and noise robustness factors, dynamically adjusting normalization scales to obtain normalized road material multi-index data, constructing a characteristic enhancement mapping module, extracting local structures of the normalized data through a radial basis function, enhancing nonlinear expression capacity through attribute interaction items, extracting high-order nonlinear characteristics to obtain final characteristic enhancement mapping output, constructing a road material multi-index comprehensive performance evaluation model, training the model through the final characteristic enhancement mapping output to obtain a trained model, and inputting new road material data samples into the trained model after processing in steps S2 and S3 to obtain a performance grade evaluation result. The method can improve the accuracy of multi-index comprehensive performance evaluation of the road material.

Inventors

Long Fuzhong
ZHANG XIAOYING
Long Fuchong
AN BIN
YU YONGWEI
JIANG FEILONG
DU LIWEI
YUAN JIAN
WANG WEIDA
LI XIANGKUN
WANG LU

Assignees

山东省路桥集团有限公司

Dates

Publication Date: 20260508
Application Date: 20250926

Claims (7)

1. The multi-index comprehensive performance evaluation method for the road material is characterized by comprising the following steps of: S1, constructing a multi-index data set of road materials and marking the data; S2, calculating attribute correlation weight and noise robustness factors for the multi-index data of the road material, and dynamically adjusting a normalization scale to obtain the multi-index data of the normalized road material, wherein the specific process is as follows: Calculating attribute correlation weight through a Sigmoid function based on the absolute value of the average pearson correlation coefficient of each attribute and other attributes, calculating local noise robustness factor through an exponential decay function based on the absolute deviation of each attribute value and the median of the samples thereof, dynamically calculating to obtain normalized road material multi-index data by combining basic scaling of global average and standard deviation, attribute correlation weight and noise robustness factor; s3, constructing a characteristic enhancement mapping module, wherein the module extracts the local structure of the normalized multi-index data of the road material through a radial basis function, and the specific process is that the K-means clustering algorithm is used for carrying out clustering analysis on the multi-index data of the normalized road material, K cluster centers are generated as center points of the radial basis function, and local distribution characteristics of the data are extracted; and the attribute interaction item is adopted to enhance the nonlinear expression capability, the high-order nonlinear characteristics are extracted, and the final characteristic enhancement mapping output is obtained, wherein the specific process is as follows: calculating attribute interaction terms of corresponding attribute values in the sample based on the predefined first attribute pair set, specifically calculating product terms of corresponding attribute values in the sample based on the predefined first attribute pair set , wherein, Represents the normalized mth attribute of the ith sample, Representing normalized ith sample nth attribute, interaction term Attribute pairs of (2) Belonging to a predefined first set of attribute pairs Comprising attribute pairs requiring capture of interaction relationships, a first set of attribute pairs The method comprises the steps of carrying out vector splicing on radial basis function values and attribute interaction items to form final characteristic enhancement mapping output; s4, constructing a multi-index comprehensive performance evaluation model of the road material, wherein the model adopts a deep feed-forward network design and comprises an input layer, a hidden layer and an output layer; S5, after the new road material data samples are processed in the steps S2 and S3, the new road material data samples are input into the trained model, and a performance grade evaluation result is obtained.
2. The method of claim 1, wherein the model calculates the initial weight based on final feature enhancement map output statistics and intra-class distribution compactness: based on the sample class labels, respectively calculating the characteristic mean value vector of each class as a representative characteristic vector of the class, and calculating a distribution divergence correction term based on the variance and covariance matrix characteristic vectors of the characteristic mean value vector, wherein the distribution divergence correction term is expressed as follows: , in the formula, Indicating the distribution divergence correction term of the first layer neural network, wherein l is the index of the neural network layer and the value range is To the point of ; Is the total layer number of the neural network; indicating the initializing adjustment rate super parameter, and controlling the amplitude of the correction term; Representing the total category number; As a logarithmic function; Characteristic variance representing the y-th class; feature vectors representing covariance matrices of the y-th class; Is that Is a transpose of (2); Representing dimensions as Is a full 1 column vector of (2); Representing a characteristic dimension of a first layer of the neural network; and adding the weighted average of the characteristic mean vectors of each class with the distribution divergence correction term to obtain an initialization weight matrix of each layer of the neural network.
3. The method for evaluating the multi-index comprehensive performance of the road material according to claim 1, wherein the hidden layer weights the characteristic channel through a gating mechanism and dynamically enhances key attribute response by combining attribute importance feedback: The method comprises the steps of calculating attention scores of all-dimensional features through a Softmax function based on absolute values of input features, generating gating vectors through a Sigmoid function based on the attention scores, applying an adaptive activation function to the features after linear transformation, combining linear transformation with nonlinear transformation to obtain adaptive activation output vectors, and multiplying the gating vectors with the adaptive activation output vectors element by element to generate output feature vectors of all layers of a neural network.
4. The method for evaluating the multi-index comprehensive performance of the road material according to claim 1, wherein the output layer dynamically fuses high-order features and physical priors through attribute attention gating and class prototype collaborative calculation to generate prediction output conforming to the scientific rule of the material: Based on the category characteristic mean value vector, the correlation weight vector and the characteristic vector of the neural network output layer, constructing attribute-level attention weight, controlling sensitivity through the correlation strength coefficient, and obtaining a correlation vector, wherein the correlation vector is expressed as: , in the formula, Represent the first Characteristic channel and first sample The association vector of the category has the dimension of ; As a Softmax function; is the correlation intensity coefficient; a feature mean vector representing a y-th class as a prototype representation of that class; feature dimension of the layer L of the neural network; Is that Is a transpose of (2); As the weight vector of the correlation, ; A relevance weight representing the jth attribute, For the correlation weight of attribute 1, For the relevance weight of attribute 2, A relevance weight for the D-th attribute; representing a transpose operation; is the neural network of Layer number Output feature vectors of the individual samples; Representing element-by-element multiplication operations; Based on the category characteristic mean vector and the interaction item splicing vector of the predefined attribute pair, generating a gating value through a Sigmoid function, and then combining the characteristic mean vector to obtain a gating prototype vector with enhanced physical constraint, wherein the gating prototype vector is expressed as follows: , in the formula, The average value of interaction terms of all samples in the samples of the y class is obtained; A gating prototype vector representing a y-th class; Is a weight matrix for gating, is a trainable parameter, and has the dimension of ; A first set of attribute pairs; Is a Sigmoid function; Representation of The vector of (2) is spliced to obtain a vector with the dimension of Is used for the vector of (a), Representing the number of elements of the first set of attribute pairs; based on the corresponding product of the associated vector element and the gating prototype vector element, a hyperbolic tangent function and a probability scaling factor are applied, and a predictive probability is output through a Softmax function.
5. The method for evaluating the multi-index comprehensive performance of the road material according to claim 1, wherein the model for evaluating the multi-index comprehensive performance of the road material constructs the total loss function by integrating the classification loss term and the physical consistency loss term.
6. The method for evaluating the multi-index comprehensive performance of the road material according to claim 5, wherein the iterative training process of the multi-index comprehensive performance evaluation model of the road material adopts an optimization algorithm based on gradient descent, and aims at minimizing a total loss function, each iteration comprises two stages of forward propagation and reverse propagation, wherein in the forward propagation, an input training sample is calculated through each layer of a neural network to obtain a prediction output, and in the reverse propagation, the gradient of the loss function relative to network parameters is calculated, and the weight and bias parameters are updated according to the learning rate.
7. The method for evaluating the multi-index comprehensive performance of the road material according to claim 6, wherein the stopping iteration condition of the model is based on the performance monitoring of the verification set, and training is stopped in advance when the loss of the verification set is no longer reduced or starts to rise for 10 consecutive iteration cycles.

Description

Multi-index comprehensive performance evaluation method for road material Technical Field The invention belongs to the technical field of data processing, and particularly relates to a multi-index comprehensive performance evaluation method for a road material. Background With the rapid development of road traffic infrastructure, the performance quality of road materials is directly related to the service life and safety of road engineering. Conventional road material evaluation methods rely on single or few indicators, such as compressive strength, density or durability, for grading by laboratory testing and empirical determination. The method has obvious limitations that on one hand, the performance of the road material is determined by various physical, mechanical, chemical and environmental sensitivity indexes, a single index cannot comprehensively reflect the whole quality of the material, and on the other hand, complex nonlinear coupling relations and physical constraints often exist between different indexes, such as the product of porosity and water absorption rate can have obvious influence on the strength, and the traditional linear method cannot effectively capture the deep correlations. The existing method has defects in the data processing link. Common normalization methods such as min-max or z-score often adopt unified scaling rules for all attributes, neglect correlation and noise difference among the attributes, and easily cause attenuation of contribution of key indexes and amplification of noise signals, so that stability and classification accuracy of a model are affected. Even if machine learning or neural network methods are introduced, the model may still suffer from the problem of "prediction accuracy but violation of the rules of the materials science" if there is a lack of feature enhancement mechanisms and physical consistency constraints for the road material properties, such as outputting unreasonable attribute-related trends. In addition, the road material data generally has the characteristics of high dimension, sparsity and unbalance, and the conventional neural network is easy to be in local optimum or gradient vanishing in the weight initialization and training process, so that classification boundaries are fuzzy, and the method is difficult to popularize in engineering practice. The problems that the existing min-max or z-score method is cut and scaled at one time on all the attributes, correlation and noise difference among the attributes are ignored, key attributes are weakened, noise is amplified, a principal component analysis method, linear projection or simple polynomial expansion can only model shallow linear relation and cannot effectively capture nonlinear interaction and coupling effect in material performance, a conventional network is insensitive to the key attributes depending on a fixed activation function, gradient disappearance or characteristic contribution is easily averaged are easy to occur, the existing method only pursues classification precision, physical consistency among the material attributes is ignored, and a result against scientific rules can be output. Disclosure of Invention The invention aims to achieve the aim, and the aim is achieved by the following technical scheme: The invention provides a multi-index comprehensive performance evaluation method of a road material, which comprises the following steps: S1, constructing a multi-index data set of road materials and marking the data; s2, calculating attribute correlation weight and noise robustness factors for the multi-index data of the road material, and dynamically adjusting a normalization scale to obtain the multi-index data of the normalized road material; S3, constructing a feature enhancement mapping module, wherein the module extracts a local structure of the normalized multi-index data of the road material through a radial basis function, enhances the nonlinear expression capacity by adopting an attribute interaction item, extracts high-order nonlinear features and obtains final feature enhancement mapping output; s4, constructing a multi-index comprehensive performance evaluation model of the road material, wherein the model adopts a deep feed-forward network design and comprises an input layer, a hidden layer and an output layer; S5, after the new road material data samples are processed in the steps S2 and S3, the new road material data samples are input into the trained model, and a performance grade evaluation result is obtained. Further, the problems of high attribute dimension, large scale difference, complex correlation among attributes and remarkable noise interference exist in the multi-index data of the road material, and the conventional normalization technologies such as min-max or z-score are used for uniformly scaling all attributes, so that the correlation among the attributes and the noise distribution difference are ignored, the weight of key attributes cannot be