Railway subgrade disease evaluation method based on multi-model integration

Abstract

The application belongs to the technical field of high-speed railway engineering and new generation information technology, and specifically relates to a railway subgrade disease evaluation method based on multi-model integration. Existing subgrade disease data is acquired, disease influence factors are screened, a sample set is constructed, and key influence factors are screened through Pearson correlation analysis. Based on the key factors, six machine learning algorithms are used to construct prediction sub-models of four disease types, and the prediction results of the multi-models are fused through an adaptive weight distribution method based on model performance evaluation. The integrated model is used to evaluate the disease proneness of the region to be analyzed, key factors are selected based on a multi-model feature importance voting mechanism, and a structural equation model is used for causal relationship analysis to explain the prediction results of the integrated model. The application provides technical support for intelligent transportation infrastructure management, and significantly improves the accuracy, stability and scientificity of subgrade disease evaluation.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of high-speed railway engineering and new-generation information technology, specifically involving a method for evaluating railway subgrade defects based on multi-model integration. Background Technology

[0002] With the rapid development of high-speed rail networks, the operational mileage of high-speed railways has exceeded 40,000 kilometers, spanning various terrains and climate regions. As a soil structure exposed to the natural environment, the roadbed of high-speed railways faces a variety of risks, such as settlement, frost damage, swelling and uplift, and mud pumping. Especially against the backdrop of global warming and frequent extreme weather events, these risks are becoming increasingly severe.

[0003] Traditional methods for evaluating roadbed defects mainly include qualitative and quantitative approaches. Qualitative methods typically involve experts using their knowledge to assess the likelihood of roadbed defects occurring; however, the evaluation results of such methods are highly dependent on the expert's subjective experience, resulting in poor reliability and consistency. Quantitative methods include those based on physical mechanics and those based on statistical analysis, but these methods have limitations in handling multi-factor nonlinear relationships.

[0004] With the rapid development of new-generation information technology, machine learning methods have been widely used in subgrade distress evaluation. However, existing solutions all face the same problems: First, the prediction accuracy of a single model is limited, failing to fully utilize the advantages of different algorithms; second, there is a lack of scientific methods for screening the importance of features from multiple models; third, existing ensemble models mostly use equal-weighted averaging or fixed-weight allocation, which cannot be dynamically adjusted according to model performance, resulting in poor ensemble performance; fourth, ensemble models suffer from a "black box" problem, lacking interpretability and failing to provide clear theoretical guidance for engineering practice. This has become a significant bottleneck restricting the deep application of artificial intelligence technology in critical infrastructure management. Existing subgrade distress evaluation methods are prone to ignoring important influencing factors or overemphasizing minor factors, and cannot explain the causal mechanisms of prediction results, leading to low accuracy and a lack of scientific explanation in the final subgrade distress evaluation results. Summary of the Invention

[0005] The purpose of this invention is to provide a railway subgrade disease evaluation method based on multi-model integration. By integrating multiple artificial intelligence algorithms and based on an adaptive weight allocation mechanism for model performance evaluation, reasonable prediction of high-speed railway subgrade diseases can be achieved.

[0006] To achieve the above objectives, this invention provides a method for evaluating railway subgrade defects based on multi-model ensemble, comprising the following steps:

[0007] S1. Obtain existing historical data on roadbed defects;

[0008] S2. Divide the study area into grids of a set size, select preliminary influencing factors based on the historical disease data of each grid, and construct a positive sample set of diseases that have occurred and a negative sample set of diseases that have not occurred.

[0009] S3. Based on Pearson correlation analysis, key influencing factors are screened from the preliminary influencing factors to obtain a set of key factors;

[0010] S4. Construct training and validation sets for the positive and negative sample sets according to a preset ratio. Using key influencing factors as input and disease prediction probability as output, construct prediction sub-models for four main disease types: subsidence, frost damage, expansion and heave, and frost heave. For each disease type, construct six prediction sub-models using six machine learning algorithms: Random Forest, XGBoost, LightGBM, GBDT, CatBoost, and AdaBoost.

[0011] S5. Divide the area to be evaluated into grids of a set size, obtain the data of each grid, extract the key influencing factors of each grid, and use the key influencing factors as the input of each prediction sub-model. Then, each disease type will output six sets of disease prediction probabilities. The adaptive weight allocation method based on model performance evaluation is used to fuse the six sets of disease prediction probabilities of each disease type, so as to obtain the probability of each disease type occurring in the area to be evaluated.

[0012] Furthermore, the method for evaluating subgrade defects in high-speed railways also includes the following steps:

[0013] S6. Based on the feature importance ranking of the six machine learning algorithms in step S4 on the key factor set, a voting mechanism is used to select the core influencing factors for structural equation modeling, and a causal relationship analysis model is constructed in combination with structural equation modeling.

[0014] S7. Use the causal relationship analysis model to interpret the prediction results of step S5 as causally.

[0015] Furthermore, in step S4, the construction of the prediction sub-model using six machine learning algorithms specifically includes the following steps:

[0016] Data preprocessing: Standardization was performed on key factors to ensure that the mean was 0 and the standard deviation was 1.

[0017] Data partitioning: The training and validation sets are partitioned in a 7:3 ratio. An imbalanced sampling design is used to ensure that the ratio of positive to negative samples in both the training and validation sets is 1:3.

[0018] Hyperparameter optimization: Bayesian optimization is performed using the Optuna optimization framework, and the optimization parameters include the number of trees, tree depth, and node splitting requirements;

[0019] Model training: A dual 5-fold cross-validation strategy was adopted. For each machine learning algorithm, an ensemble model was built through 30 independent resampling operations to obtain predictive sub-models for four disease types: settlement, frost damage, expansion heave, and frost heave.

[0020] The hyperparameter optimization of the six machine learning algorithms adopts a unified Bayesian optimization framework, and the specific parameter settings are as follows:

[0021] Random Forest: Number of estimators ∈ [100, 1000], maximum depth ∈ [3, 20], minimum number of split samples ∈ [2, 10];

[0022] XGBoost: Number of estimators ∈ [100, 1000], maximum depth ∈ [3, 10], learning rate ∈ [0.01, 0.3], subsample ratio ∈ [0.6, 1.0];

[0023] LightGBM: Number of estimators ∈ [100, 1000], maximum depth ∈ [3, 10], learning rate ∈ [0.01, 0.3], feature sampling ratio ∈ [0.6, 1.0];

[0024] GBDT: Number of estimators ∈ [100, 1000], maximum depth ∈ [3, 10], learning rate ∈ [0.01, 0.3];

[0025] CatBoost: Number of iterations ∈ [100, 1000], Depth ∈ [3, 10], Learning rate ∈ [0.01, 0.3];

[0026] AdaBoost: Number of estimators ∈ [50, 500], learning rate ∈ [0.01, 2.0];

[0027] Bayesian optimization was performed using the Optuna framework with 300 trials. The objective function was the average AUC value of the 5-fold cross-validation.

[0028] Furthermore, in step S5, the overall performance index (CPI) of each algorithm is calculated based on 5-fold cross-validation: CPI = 0.4 × AUC + 0.3 × Precision + 0.2 × Recall + 0.1 × F1 score; the model weights are dynamically adjusted according to the CPI value: w i = CPI i / ∑CPI j And set the weight constraint condition: 0.05 ≤ w i ≤ 0.35; Evaluate the susceptibility of roadbed defects in the area to be analyzed.

[0029] Specifically, the adaptive weight allocation method based on model performance evaluation is used to fuse the six sets of disease prediction probabilities for each disease type, including the following steps:

[0030] S51. Performance Evaluation: Based on 5-fold cross-validation, calculate the AUC, precision, recall, and F1 score of the prediction sub-models obtained by the six machine learning algorithms respectively.

[0031] S52. Calculation of Comprehensive Performance Index: CPI = 0.4 × AUC + 0.3 × Precision + 0.2 × Recall + 0.1 × F1 score;

[0032] S53, Weight Allocation: w i = CPI i / ∑CPI j And set constraints to satisfy: w i = max(0.05, min(0.35, w i )); where w i CPI is the weight of the i-th predictive sub-model. i Let ∑CPI be the overall performance index of the i-th predictive sub-model. j The sum of the overall performance indices of all predictive sub-models;

[0033] S54, Ensemble Prediction: P_ensemble = ∑(w i × P i ), where P i P_i represents the prediction probability of the i-th prediction sub-model, and P_ensemble represents the prediction probability of each disease type after fusion.

[0034] Furthermore, the natural breakpoint method is used to divide the predicted probability of each disease type into four levels: low, low-medium, medium-high, and high, with values ​​of 0, 1, 2, and 3 respectively, and a multi-disease comprehensive risk distribution map is generated.

[0035] Furthermore, in step S6, the screening of the core influencing factors specifically includes the following steps:

[0036] S61. Importance Ranking Acquisition: Acquire the feature importance rankings of six algorithms on the key factor set: Random Forest based on Gini importance ranking, XGBoost based on gain importance ranking, LightGBM based on gain importance ranking, GBDT based on feature importance score ranking, CatBoost based on predicted value changes to obtain importance ranking, and AdaBoost based on feature importance ranking.

[0037] S62. Core Factor Screening: A voting mechanism is used to select the core influencing factors for structural equation modeling. The number of times each key factor ranks in the top 80% of the six models is counted, and factors that appear more than or equal to 3 times are selected as core influencing factors.

[0038] S63. Factor Importance Verification: Verify the rationality of the selection results by calculating the average ranking position of the selected core influencing factors in each model.

[0039] Furthermore, the construction of the causal relationship analysis model using structural equation modeling specifically includes the following steps:

[0040] Initial variable screening: Import the core influencing factors into AMOS software to build a preliminary measurement model and exclude variables with P-values ​​greater than 0.05;

[0041] Establish a measurement model: link observed variables with latent variables;

[0042] Constructing structural equation modeling: Analyzing the causal relationships between latent variables based on disaster system theory;

[0043] Model optimization: The model fit is evaluated using standardized root mean square residuals, root mean square approximation error, goodness-of-fit index, and comparison fit index.

[0044] The significance of the mediation effect was verified by performing 4000 resampling operations using the self-administered resampling method.

[0045] Furthermore, in step S7, the explanation of the causal relationship specifically includes:

[0046] Global interpretive analysis: Through standardized path coefficients of structural equation modeling, the influence intensity of each core influencing factor on different disease types is quantitatively described;

[0047] Local explanatory analysis: For regions classified as high, analyze the consistency of prediction results from different models;

[0048] Model prediction consistency analysis: Calculate the standard deviation and correlation coefficient of the prediction results of six machine learning algorithms to evaluate the reliability of the integrated prediction;

[0049] Causal path explanation: Combining engineering practice, provide causal explanations and prevention and control suggestions for each type of disease;

[0050] Mediating mechanism identification: Identify key mediating and moderating variables to explain the transmission mechanism of disease occurrence.

[0051] Furthermore, in step S3, Pearson correlation analysis is used to screen key factors, specifically including the following steps: calculating the Pearson correlation coefficient among each influencing factor using the formula... Redundant factors with a correlation greater than 0.8 are removed, where r is the Pearson correlation coefficient. and These are the i-th data points of variables X and Y, respectively. and These are the means of variables X and Y, respectively; the remaining factors are subjected to variance inflation factor test to ensure that the VIF value is less than 10; finally, the statistically significant and mutually independent key influencing factors are retained as the key factor set.

[0052] Furthermore, in step S2, the preliminary influencing factors include climate variables, geomorphological variables, geological variables, high-speed rail operation characteristics, economic variables, and spatial lag variables. Standardization processing methods are used to normalize the data of different dimensions.

[0053] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages:

[0054] 1. The railway subgrade disease evaluation method based on multi-model integration provided by this invention uses new-generation information technology to scientifically screen key influencing factors through Pearson correlation analysis, avoiding multicollinearity problems and the subjectivity of factor selection in traditional methods; and constructs a high-precision intelligent prediction model based on the screened key factor set through various artificial intelligence machine learning algorithms, realizing accurate prediction of four major disease types.

[0055] 2. An innovative adaptive weight allocation mechanism based on model performance evaluation is proposed, overcoming the technical shortcomings of traditional ensemble methods that use equal weights or fixed weights. By evaluating the performance of each model and assigning corresponding weights during the training phase, the advantages of each model can be better utilized, the predictive power of high-performance models can be fully leveraged, and the accuracy and stability of roadbed disease evaluation can be significantly improved.

[0056] 3. This invention innovatively uses structural equation modeling to analyze the causal relationships of the importance factors in the integrated model, thus solving the "black box" problem of artificial intelligence algorithms, achieving model interpretability, and providing clear theoretical guidance for intelligent transportation infrastructure management. Through explicit causal path analysis, engineers can clearly understand how various environmental factors affect different types of roadbed defects, thereby formulating targeted prevention and control measures. This invention deeply integrates advanced information technology with traditional engineering management, possessing not only high prediction accuracy but, more importantly, excellent interpretability and engineering guidance value. It provides important technical support for promoting intelligent management of transportation infrastructure, truly achieving a highly reliable, accurate, and objectively scientific intelligent evaluation of roadbed defects. Attached Figure Description

[0057] Figure 1This is a flowchart illustrating the railway subgrade disease evaluation method based on multi-model integration provided by the present invention.

[0058] Figure 2 The following is a schematic diagram of the distribution of roadbed defects in high-speed railways according to the method of the present invention: (a) roadbed settlement; (b) frost damage; (c) swelling and heave; (d) mud pumping and mud bubbling.

[0059] Figure 3 This is a schematic diagram of the Pearson correlation coefficient for the method of the present invention.

[0060] Figure 4 The following is a schematic diagram illustrating the use of the AUC (area under the curve) curve to evaluate the random forest model using the method of this invention: (a) roadbed settlement; (b) frost damage; (c) swelling and heave; (d) mud pumping.

[0061] Figure 5 The diagram illustrates the risk prediction of high-speed railway subgrade defects using the method of this invention: (a) subgrade settlement; (b) frost damage; (c) swelling and heave; (d) frost heave and mudslide. The figure shows the mean of 30 random forest (RF) model ensembles.

[0062] Figure 6 This is a schematic diagram illustrating the comprehensive risk of high-speed railway subgrade defects according to the method of the present invention.

[0063] Figure 7 This diagram illustrates the causal path of high-speed railway subgrade defects according to the method of this invention. (The diagram shows the final structural equation model (SEM), which assesses the impact of geomorphic variables, land use composition, climate variables, geological variables, economic variables, spatial lag, high-speed railway operation characteristics, and high-speed railway proximity index on subgrade defects. Standardized path coefficients are indicated next to unidirectional arrows, representing hypothetical causal relationships. Significant path coefficients are marked with an asterisk, indicating the significance level (*p<.05, **p<.01, ***p<.001). Gray arrows indicate insignificant paths. Small double-headed arrows next to endogenous variables represent covariance, while large arrows between latent variables represent correlation coefficients.)

[0064] Figure 8 This diagram illustrates the standardized effects of various variables on high-speed railway subgrade defects in the method of this invention. (The diagram summarizes the total, direct, and indirect effects of the key variables affecting subgrade defects. The total effect is the sum of the direct and indirect effects. The direct effect is represented by the path coefficients between variables, while the indirect effect is calculated by multiplying multiple path coefficients.) Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0066] This invention provides a method for evaluating railway subgrade defects based on multi-model ensemble, comprising the following steps:

[0067] S1. Obtain existing historical data on roadbed defects;

[0068] S2. Divide the study area into grids of a set size, select preliminary influencing factors based on the historical disease data of each grid, and construct a positive sample set of diseases that have occurred and a negative sample set of diseases that have not occurred.

[0069] S3. Based on Pearson correlation analysis, key influencing factors are screened from the preliminary influencing factors to obtain a set of key factors;

[0070] S4. Using the positive and negative sample sets as training and validation sets, with key influencing factors as input and disease prediction probability as output, predictive sub-models for four main disease types—settlement, frost damage, expansion heave, and frost heave—are constructed respectively. For each disease type, six predictive sub-models are constructed using six machine learning algorithms: Random Forest, XGBoost, LightGBM, GBDT, CatBoost, and AdaBoost.

[0071] S5. Divide the area to be evaluated into grids of a set size, obtain the data of each grid, extract the key influencing factors of each grid, and use the key influencing factors as the input of each prediction sub-model. Then, each disease type will output six sets of disease prediction probabilities. The adaptive weight allocation method based on model performance evaluation is used to fuse the six sets of disease prediction probabilities of each disease type, so as to obtain the probability of each disease type occurring in the area to be evaluated.

[0072] Furthermore, the method for evaluating subgrade defects in high-speed railways also includes the following steps:

[0073] S6. Based on the feature importance ranking of the six machine learning algorithms in step S4 on the key factor set, a voting mechanism is used to select the core influencing factors for structural equation modeling, and a causal relationship analysis model is constructed in combination with structural equation modeling.

[0074] S7. Use the causal relationship analysis model to interpret the prediction results of step S5 as causally.

[0075] Specifically, such as Figure 1 As shown, the intelligent evaluation method for high-speed railway subgrade defects based on multi-model ensemble and structural equation modeling provided by this invention includes the following steps:

[0076] Step S1: Obtain existing roadbed defect data.

[0077] Through extensive literature review, historical data on roadbed defects of high-speed railways were collected, including records of defect events with multiple geographical references. These data cover four main types of defects: subsidence, frost damage, expansion and uplift, and frost heave, spanning different administrative levels such as provinces, cities, counties, and townships.

[0078] Step S2: Preliminary selection of impact factors and construction of sample set

[0079] Since the roadbed of high-speed railways crosses various terrain and climate regions, we will select climate variables, land use variables, geomorphological variables, high-speed railway operation variables, economic variables, high-speed railway proximity index, spatial lag, and geological variables as the factor set. The specific factors are shown in Table 1.

[0080] Table 1 Detailed Description of Variables

[0081]

[0082] The study area was divided into grids of a set size. Preliminary influencing factors were selected based on the historical disease data of each grid, including: climate variables (annual average rainfall, 5-day consecutive rainfall, number of days with maximum temperature exceeding 35°C, number of days of freezing per year, wind speed), geomorphological variables (elevation, slope, aspect), geological variables (rock hardness, soil composition, effective water storage capacity, soil depth, drainage grade), high-speed rail operation characteristics (operation frequency, construction length, design speed, operation cycle), economic variables (GDP, population, county area), etc.

[0083] Spatial lag variable calculation: Moran's I index was used for spatial autocorrelation analysis. The conceptual spatial relationship was defined as inverse distance, and the distance method was set to Euclidean distance. A spatial weight matrix was constructed, and a 100-meter threshold was set to define spatial neighbors. Adjacent polygons were assigned a weight of 1, and non-adjacent polygons were assigned a weight of 0, and row standardization was performed. The spatial lag value of each disease type was calculated using the maximum likelihood estimation method.

[0084] Construct positive and negative sample sets: Construct a positive sample set based on historical disease coordinates, and randomly select grid cells that have not experienced disease to construct a negative sample set.

[0085] Step S3: Screening of key influencing factors based on Pearson correlation analysis

[0086] Pearson correlation analysis was used to screen key factors:

[0087] Data preprocessing: Examine the data distribution, identify outliers that exceed the interquartile range (IQR), check for missing values ​​and process them;

[0088] Pearson correlation analysis: The correlation coefficients among the influencing factors were calculated using the following formula:

[0089] (1)

[0090] Where r is the Pearson correlation coefficient. and These are the i-th data points of variables X and Y, respectively. and These are the means of variables X and Y, respectively. Redundant factors with a Pearson correlation coefficient (r) greater than 0.8 are removed, such as... Figure 3 As shown, this avoids the problem of multicollinearity.

[0091] Variance inflation factor test: Perform a variance inflation factor test on the remaining factors to ensure that the variance inflation factor (VIF) value is less than 10, further eliminating the problem of multicollinearity;

[0092] Finally, a set of statistically significant and independent key influencing factors was selected, providing a scientific variable basis for subsequent machine learning modeling.

[0093] Step S4: Construction of prediction sub-models based on multiple machine learning algorithms

[0094] A predictive sub-model is constructed using six machine learning algorithms based on a set of key factors.

[0095] Data preprocessing: Standard scaler is used to standardize the key factors to ensure that the mean of all features is 0 and the standard deviation is 1, thus eliminating the influence of feature scale differences.

[0096] Data partitioning: The dataset is randomly divided into training and validation sets in a 7:3 ratio. An imbalanced sampling design is used to ensure that the ratio of positive to negative samples is 1:3, thereby improving the model's ability to identify the minority class.

[0097] Hyperparameter optimization for six algorithms: Bayesian optimization was performed using the Optuna optimization framework, and 300 trials were conducted to find the optimal parameter combination. Specific parameter settings are as follows:

[0098] Random Forest Algorithm: As a classic ensemble learning algorithm, it constructs multiple decision trees and uses voting to determine the final result. Hyperparameter settings include the number of trees between 100 and 1000, the maximum depth between 3 and 20, and the minimum number of split samples between 2 and 10. Feature importance is evaluated using Gini importance and permutation importance.

[0099] XGBoost algorithm: An ensemble algorithm based on gradient boosting, possessing strong predictive power and anti-overfitting capabilities. Hyperparameter settings include the number of estimators between 100 and 1000, maximum depth between 3 and 10, learning rate between 0.01 and 0.3, and subsample ratio between 0.6 and 1.0. Feature importance is evaluated using three metrics: weight, gain, and coverage.

[0100] LightGBM algorithm: A gradient boosting framework developed by Microsoft, characterized by fast training speed and low memory usage. Hyperparameter settings include the number of estimators between 100 and 1000, maximum depth between 3 and 10, learning rate between 0.01 and 0.3, and feature sampling ratio between 0.6 and 1.0.

[0101] GBDT algorithm: A classic gradient boosting decision tree algorithm that optimizes model performance by progressively adding weak learners. Hyperparameter settings include the number of estimators between 100 and 1000, the maximum depth between 3 and 10, and the learning rate between 0.01 and 0.3.

[0102] CatBoost algorithm: A gradient boosting algorithm developed by Yandex, specifically optimized for categorical features. Hyperparameter settings include the number of iterations between 100 and 1000, depth between 3 and 10, and learning rate between 0.01 and 0.3.

[0103] AdaBoost algorithm: An adaptive boosting algorithm that improves classification accuracy by adjusting sample weights. Hyperparameter settings include the number of estimators between 50 and 500, and the learning rate between 0.01 and 2.0.

[0104] Double cross-validation: The first 5-fold cross-validation is used for hyperparameter optimization, and the second 5-fold cross-validation is used to evaluate the generalization ability of the optimized model. In each cross-validation, the training set is divided into 5 subsets: 4 subsets are used for training, and 1 subset is used for validation.

[0105] Model ensemble: An ensemble model was constructed through 30 independent resampling operations, establishing six predictive sub-models using machine learning algorithms for each of the four disease types: settlement, frost damage, expansion heave, and mud pumping. The model's fit was assessed using the area under the curve (AUC). Figure 4 As shown in the figure. Precision, recall, F1 score, and other metrics are calculated using the validation set in the 7:3 ratio described above.

[0106] Specifically, for each machine learning algorithm, 30 models are fitted through 30 independent resampling operations. Each model yields a set of AUC, precision, recall, and F1 scores. The average of the 30 sets is calculated to obtain the AUC, precision, recall, and F1 scores corresponding to that machine learning algorithm, which are then used for subsequent weight allocation calculations.

[0107] Step S5: Evaluation of Subgrade Disease Susceptibility (Adaptive Weighted Integration)

[0108] An adaptive weight allocation method based on model performance evaluation is used to evaluate the susceptibility of roadbed defects.

[0109] Obtain target area data: Based on the key factor set, extract the corresponding environmental variable data for the target area, including climate, landform, geology, high-speed rail operation characteristics, economic variables and spatial lag variables;

[0110] Data preprocessing: Standardize the target region data to ensure consistency with the dimensions and distribution of the training data;

[0111] Six machine learning methods were used to predict four types of diseases, obtaining the probability of disease occurrence for each grid cell. According to feature importance analysis, subsidence disease is mainly affected by factors such as elevation, slope, bare land, construction length, and operation frequency; frost damage disease is mainly affected by climate variables such as the number of frozen days per year and average annual rainfall; expansion and uplift disease is mainly affected by factors such as bare land and construction length; and mud pumping disease is mainly affected by factors such as average annual rainfall and operation frequency.

[0112] Overall Performance Index Calculation: The weighted average method is used to calculate the overall performance index (CPI) of each algorithm, with the formula: CPI = 0.4 × AUC + 0.3 × Precision + 0.2 × Recall + 0.1 × F1 score. Where, CPI is the overall performance index used to evaluate the overall performance of a single model; AUC is the area under the ROC curve, ranging from [0,1]; precision is the proportion of correctly predicted positive samples out of all correctly predicted positive samples; recall is the proportion of correctly predicted positive samples out of all actual positive samples; and F1 score is the harmonic mean of precision and recall. Weighting coefficients of 0.4, 0.3, 0.2, and 0.1 respectively reflect the importance of each performance index in disease prediction.

[0113] Dynamic weight allocation: The weights of each algorithm are calculated based on the CPI value, using the formula: w i = CPIi / ∑CPI j , where w i CPI is the weight of the i-th model. i Let ∑CPI be the overall performance index of the i-th model. jThis represents the sum of the overall performance indices of all models. To avoid any model having an excessively large or small weight, the following constraint is set: 0.05 ≤ wi ≤ 0.35, where 0.05 is the lower bound for the weights to ensure that each model has a minimum level of participation, and 0.35 is the upper bound for the weights to prevent any model from having an excessively large weight, which could lead to a degradation in the ensemble performance. Taking settlement disorder as an example, the calculated weights are: w_RF=0.18, w_XGB=0.20, w_LGB=0.22, w_GBDT=0.15, w_Cat=0.21, w_Ada=0.04.

[0114] Ensemble prediction: The final prediction result is obtained using a weighted average method: P_ensemble = ∑(w i × P i ), where P_ensemble is the final predicted probability of the ensemble model, w i P represents the dynamic weights of the i-th model. i Let P be the prediction probability of the i-th model. This method can allocate weights according to the actual performance of each model, making better use of the prediction capabilities of high-performance models. Specifically, during prediction, the data of the region to be tested is input into 30 sets of models for each trained machine learning algorithm, resulting in 30 prediction probabilities. The average of these probabilities is taken as the prediction probability of that machine learning algorithm, corresponding to P. i Then, by combining the weights, the final prediction probability is calculated based on the prediction probabilities of the six machine learning algorithms.

[0115] The natural breakpoint method was used to classify the final predicted probabilities of the four diseases into four levels: low, low-medium, medium-high, and high. Figure 5 As shown.

[0116] A comprehensive risk map of railway subgrade defects was drawn, and the four types of defects were spatially coupled. First, regions with an integrated probability value of 0.6 or greater for each defect were selected and marked as 1, while regions with a probability value less than 0.6 were marked as 0.

[0117] Subsequently, the regions in the map were divided into low, medium, high, and very high risk levels based on risk levels, assigned values ​​of 0, 1, 2, and 3 respectively (where 0 indicates that the predicted probability of all four diseases in a region is less than 0.6, and 3 indicates that the predicted probability of three diseases in a region is greater than or equal to 0.6), representing different levels of risk. This was achieved by generating a comprehensive risk distribution map of multiple diseases, such as... Figure 6 As shown, high-risk sections such as the Menyuan-Minle section of the Lanzhou-Xinjiang High-Speed ​​Railway and the Harbin-Dalian High-Speed ​​Railway were identified, providing quantitative basis for the scientific management and risk prevention of high-speed railway subgrade.

[0118] Step S6: Causal relationship analysis based on structural equation modeling

[0119] Structural equation modeling based on factor importance from an ensemble model:

[0120] Voting Mechanism Screening: A voting mechanism is used to select core influencing factors for structural equation modeling. The cutoff value K is calculated as 80% of the total number of features in the key factor set. The frequency of each key factor ranking in the top K positions across the six models is counted. Factors appearing at least three times are selected as core influencing factors, ensuring that the selected factors have high importance in most models.

[0121] Initial variable screening: Import the core influencing factors into AMOS software to build a preliminary measurement model, exclude variables with P-values ​​greater than 0.05, and ensure that only statistically significant factors are retained;

[0122] Establish a measurement model: By linking observed variables and latent variables, the following measurement equation is established:

[0123] (1)

[0124] (2)

[0125] Where (1) and (2) are the measurement models for exogenous and endogenous variables, respectively, and ξ and η represent the exogenous and endogenous latent variables, respectively. and Here, δ and ε represent the load matrix, and δ and ε represent the measurement errors.

[0126] Constructing a structural equation model: Based on disaster system theory, analyze the causal relationships between latent variables and establish structural equations.

[0127] (3)

[0128] Where B is the endogenous latent variable coefficient matrix, Γ is the direct effect coefficient matrix of exogenous latent variables on endogenous latent variables, and ζ is the structural equation error. This is the error vector of the structural equation.

[0129] Model Fit Optimization: Multiple fit metrics are used to evaluate the model quality. Specifically, a Standardized Root Mean Square Residual (SRMR) less than 0.05 indicates an excellent model fit; a Root Mean Square Approximation Error (RMSEA) less than 0.05 indicates a very good fit; a Goodness-of-Fit Index (GFI) greater than 0.95 indicates an excellent fit; and a Comparison Fit Index (CFI) greater than 0.95 indicates an excellent model fit. These metrics provide a comprehensive evaluation of the model's fit performance.

[0130] Mediation effect test: 4000 resampling operations were performed using the bootstrap method, with a 95% confidence level. The significance of the mediation effect was verified using the bias correction method and the percentile method.

[0131] Step S7: Interpretability analysis of integrated prediction results

[0132] The causal analysis results from structural equation modeling are used to interpret the ensemble prediction results:

[0133] Causal relationship explanation: Standardized path coefficients are used to quantitatively describe the intensity of the influence of each factor on different disease types. For example... Figure 7 and 8 As shown, topographic variables have a positive direct impact on subgrade defects (standardized path coefficient is 0.36); land use has a positive direct impact on subgrade defects (standardized path coefficient is 0.30); high-speed rail operation characteristics have a positive direct impact on subgrade defects (standardized path coefficient is 0.42); and spatial lag factors have a strong positive impact on subgrade defects (standardized path coefficient is 0.92).

[0134] Direct and Indirect Effects Analysis: This section analyzes the direct and indirect effects among various influencing factors to reveal complex causal networks. For example, geomorphological variables not only directly affect roadbed defects but also indirectly influence their occurrence by affecting the operational characteristics of high-speed railways; climate variables can indirectly affect the occurrence of roadbed defects through land use.

[0135] Model prediction consistency analysis: The standard deviation and correlation coefficient of the prediction results of the six algorithms are calculated to evaluate the reliability of the integrated prediction. A smaller standard deviation indicates that the prediction results of each model are more consistent, and a higher correlation coefficient indicates that the prediction trends among the models are similar.

[0136] Mediating mechanism identification: Identifying key mediating and moderating variables to explain the transmission mechanism of disease occurrence. For example, economic factors influence roadbed defects by affecting the operational characteristics of high-speed railways;

[0137] Engineering Guidance Significance: Based on practical engineering scenarios, this document provides targeted explanations of the causes and prevention recommendations for each type of disease. For subsidence diseases, the focus should be on coordinating topographical conditions with operational loads; for frost damage diseases, the main focus should be on controlling climatic factors, especially the impact of freeze-thaw cycles; for expansion and uplift diseases, the focus should be on monitoring land use changes and construction intensity; and for mud pumping diseases, the main focus should be on controlling the coordination between rainfall infiltration and operational frequency.

[0138] Overcoming the black box problem: Through explicit causal analysis, engineers can clearly understand the cause mechanism of the prediction results of the ensemble model, avoiding the problem that traditional machine learning methods cannot explain the prediction results, and achieving an organic combination of high-precision prediction and strong interpretability.

[0139] Implementation examples demonstrate that the adaptive weight allocation mechanism of this invention effectively integrates the predictive advantages of each model across four disease types, improving prediction stability and reliability through dynamic weight adjustment. This intelligent evaluation method not only accurately identifies high-risk areas but, more importantly, fully leverages the predictive capabilities of each model through the adaptive weight allocation mechanism, providing engineers with more reliable decision support.

[0140] In summary, this invention provides a highly reliable, accurate, and interpretable intelligent evaluation method for high-speed railway subgrade defects based on multi-model ensemble and structural equation modeling. By combining next-generation information technology, artificial intelligence algorithms, and statistical causal analysis, and particularly by innovatively proposing an adaptive weight allocation mechanism, this method better utilizes the predictive capabilities of high-performance models through dynamic weight allocation, effectively improving the accuracy and stability of ensemble prediction. Accurate defect prediction is achieved through multi-model feature importance voting and artificial intelligence algorithms. Simultaneously, by leveraging structural equation modeling to reveal causal relationships, it overcomes the limitations of single models and the black-box problem of ensemble models, significantly improving the accuracy and interpretability of the model and providing technical support for intelligent transportation infrastructure management.

[0141] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for evaluating railway subgrade defects based on multi-model ensemble, characterized in that, Includes the following steps: S1. Obtain existing historical data on roadbed defects; S2. Divide the study area into grids of a set size, select preliminary influencing factors based on the historical disease data of each grid, and construct a positive sample set of diseases that have occurred and a negative sample set of diseases that have not occurred. S3. Based on Pearson correlation analysis, key influencing factors are screened from the preliminary influencing factors to obtain a set of key factors; S4. Construct training and validation sets for the positive and negative sample sets according to a preset ratio. Using key influencing factors as input and disease prediction probability as output, construct prediction sub-models for four main disease types: subsidence, frost damage, expansion and heave, and frost heave. For each disease type, construct six prediction sub-models using six machine learning algorithms: Random Forest, XGBoost, LightGBM, GBDT, CatBoost, and AdaBoost. S5. Divide the area to be evaluated into grids of a set size, obtain the data of each grid, extract the key influencing factors of each grid, and use the key influencing factors as the input of each prediction sub-model. Then, each disease type will output six sets of disease prediction probabilities. The adaptive weight allocation method based on model performance evaluation is used to fuse the six sets of disease prediction probabilities of each disease type, so as to obtain the probability of each disease type occurring in the area to be evaluated. The method of using adaptive weight allocation based on model performance evaluation to fuse the six sets of disease prediction probabilities for each disease type specifically includes the following steps: S51. Performance Evaluation: Based on 5-fold cross-validation, calculate the AUC, precision, recall, and F1 score of the prediction sub-models obtained by the six machine learning algorithms respectively. S52. Calculation of Comprehensive Performance Index: CPI = 0.4 × AUC + 0.3 × Precision + 0.2 × Recall + 0.1 × F1 score; S53, Weight Allocation: w i = CPI i / ∑CPI j And set constraints to satisfy: w i = max(0.05, min(0.35, w i )); where w i CPI is the weight of the i-th predictive sub-model. i Let ∑CPI be the overall performance index of the i-th predictive sub-model. j The sum of the overall performance indices of all predictive sub-models; S54, Ensemble Prediction: P_ensemble = ∑(w i × P i ), where P i P_i represents the prediction probability of the i-th prediction sub-model, and P_ensemble represents the prediction probability of each disease type after fusion.

2. The railway subgrade distress evaluation method based on multi-model ensemble as described in claim 1, characterized in that, The evaluation method for subgrade defects of high-speed railways also includes the following steps: S6. Based on the feature importance ranking of the six machine learning algorithms in step S4 on the key factor set, a voting mechanism is used to select the core influencing factors for structural equation modeling, and a causal relationship analysis model is constructed in combination with structural equation modeling. S7. Use the causal relationship analysis model to interpret the prediction results of step S5 as causally.

3. The railway subgrade distress evaluation method based on multi-model ensemble as described in claim 1 or 2, characterized in that, Step S4, which involves constructing a predictive sub-model using six machine learning algorithms, includes the following steps: Data preprocessing: Standardization was performed on key factors to ensure that the mean was 0 and the standard deviation was 1. Data partitioning: The training and validation sets are partitioned in a 7:3 ratio. An imbalanced sampling design is used to ensure that the ratio of positive to negative samples in both the training and validation sets is 1:

3. Hyperparameter optimization: Bayesian optimization is performed using the Optuna optimization framework, and the optimization parameters include the number of trees, tree depth, and node splitting requirements; Model training: A dual 5-fold cross-validation strategy was adopted. For each machine learning algorithm, an ensemble model was built through 30 independent resampling operations to obtain predictive sub-models for four disease types: settlement, frost damage, expansion heave, and frost heave.

4. The railway subgrade distress evaluation method based on multi-model ensemble as described in claim 1, characterized in that, The natural breakpoint method is used to classify the predicted probability of each disease type into four levels: low, low-medium, medium-high, and high, and assign values ​​of 0, 1, 2, and 3 respectively, and generate a comprehensive risk distribution map of multiple diseases.

5. The railway subgrade distress evaluation method based on multi-model ensemble as described in claim 2, characterized in that, In step S6, the screening of the core influencing factors specifically includes the following steps: S61. Importance Ranking Acquisition: Acquire the feature importance rankings of six algorithms on the key factor set: Random Forest based on Gini importance ranking, XGBoost based on gain importance ranking, LightGBM based on gain importance ranking, GBDT based on feature importance score ranking, CatBoost based on predicted value changes to obtain importance ranking, and AdaBoost based on feature importance ranking. S62. Core Factor Screening: A voting mechanism is used to select the core influencing factors for structural equation modeling. The number of times each key factor ranks in the top 80% of the six models is counted, and factors that appear more than or equal to 3 times are selected as core influencing factors. S63. Factor Importance Verification: Verify the rationality of the selection results by calculating the average ranking position of the selected core influencing factors in each model.

6. The railway subgrade distress evaluation method based on multi-model ensemble as described in claim 5, characterized in that, The construction of the causal relationship analysis model by combining structural equation modeling specifically includes the following steps: Initial variable screening: Import the core influencing factors into AMOS software to build a preliminary measurement model and exclude variables with P-values ​​greater than 0.05; Establish a measurement model: link observed variables with latent variables; Constructing structural equation modeling: Analyzing the causal relationships between latent variables based on disaster system theory; Model optimization: The model fit is evaluated using standardized root mean square residuals, root mean square approximation error, goodness-of-fit index, and comparison fit index. The significance of the mediation effect was verified by performing 4000 resampling operations using the self-administered resampling method.

7. The railway subgrade distress evaluation method based on multi-model ensemble as described in claim 5, characterized in that, In step S7, the explanation of the causal relationship specifically includes: Global interpretive analysis: Through standardized path coefficients of structural equation modeling, the influence intensity of each core influencing factor on different disease types is quantitatively described; Local explanatory analysis: For regions classified as high, analyze the consistency of prediction results from different models; Model prediction consistency analysis: Calculate the standard deviation and correlation coefficient of the prediction results of six machine learning algorithms to evaluate the reliability of the integrated prediction; Causal path explanation: Combining engineering practice, provide causal explanations and prevention and control suggestions for each type of disease; Mediating mechanism identification: Identify key mediating and moderating variables to explain the transmission mechanism of disease occurrence.

8. The method for evaluating railway subgrade defects based on multi-model ensemble according to claim 1 or 2, characterized in that, In step S3, Pearson correlation analysis is used to screen key factors, specifically including the following steps: calculating the Pearson correlation coefficient among each influencing factor using the formula... Redundant factors with a correlation greater than 0.8 are removed, where r is the Pearson correlation coefficient. and These are the i-th data points of variables X and Y, respectively. and These are the means of variables X and Y, respectively; The remaining factors were subjected to variance inflation factor test to ensure that the VIF value was less than 10; finally, the statistically significant and independent key influencing factors were retained as the key factor set.

9. The railway subgrade distress evaluation method based on multi-model ensemble according to claim 1 or 2, characterized in that, In step S2, the preliminary influencing factors include climate variables, geomorphological variables, geological variables, high-speed rail operation characteristics, economic variables, and spatial lag variables. The data of different dimensions are normalized using a standardization processing method.

Images