High-speed railway subgrade compaction quality full-section evaluation method and application thereof

Abstract

A machine learning and interval prediction theory-based high-speed railway roadbed compaction quality full-section evaluation method comprises the following steps: determining the maximum dry density rho dmax based on a dynamic stiffness Krb turning point based on an indoor vibration compaction test; establishing a POA model for predicting rho dmax; introducing an interval prediction theory to quantify the uncertainty of rho dmax prediction, and establishing a BPA interval prediction model of rho dmax; and in combination with the interval prediction model, a spatial interpolation algorithm and a compaction standard, establishing a high-speed railway roadbed compaction quality full-section evaluation model. The method provided by the invention solves the problems of uncertainty, low efficiency, lack of intelligence and the like of the existing mechanism and method for determining the rho dmax. Due to the defects, the compaction quality of the high-speed railway roadbed can be insufficiently evaluated, so that the stability and the strength of the roadbed structure are reduced, diseases such as differential settlement and track irregularity can be possibly caused in the later operation process, and the operation safety of the high-speed railway is seriously influenced.

Description

Technical Field

[0001] The present invention relates to the field of intelligent algorithm technology, and in particular to the field of intelligent prediction and control technology for roadbed vibration compaction, and specifically to a method for evaluating the full-section compaction quality of high-speed railway roadbed based on machine learning and interval prediction theory and its application. Background Art

[0002] The compaction quality of high-speed railway subgrade is controlled by the compaction degree K and the maximum dry density ρ dmax It is the key indicator of its calculation. dmax The mechanisms and methods of high-speed railway subgrade compaction still have problems such as uncertainty, inefficiency and lack of intelligence. These defects will lead to insufficient assessment of the quality of high-speed railway subgrade compaction, thereby reducing the stability and strength of the subgrade structure, and may cause uneven settlement, track unevenness and other diseases in the later operation process, seriously affecting the operational safety of high-speed railways.

[0003] The compaction degree K is the dry density ρ measured by the on-site sand cone method d and ρ measured by indoor vibration compaction method dmax Therefore, the ρ obtained from the indoor experiment dmax It is a standard value for evaluating compaction quality. At present, vibration compaction method, as a popular compaction method, has become the most important method for determining the quality of coarse-grained soil filling materials for high-speed railway subgrades. dmax However, this method is used to accurately determine the filling material ρ of high-speed railway subgrade. dmax There is still a lack of basis for the optimal compaction time. People have concluded that the optimal vibration time is between 60 and 480 seconds, but dmax However, it is impossible to determine it accurately, which leads to errors in the evaluation of compaction quality. Therefore, it is necessary to propose a method to determine ρ in vibration compaction. dmax new method.

[0004] Based on ρ dmax The prediction model can realize the rapid calculation of the compaction degree K, which is an important method for evaluating the compaction quality of high-speed railway subgrade and can save a lot of field test workload. dmax The prediction task of ρ has typical nonlinear characteristics. Many scholars use simple regression models to establish ρ dmax The relationship between ρ and filler parameters is still under discussion, but the accuracy of its prediction results remains to be discussed. Recently, machine learning (ML) methods have been widely used due to their nonlinear mapping, high efficiency and intelligence. dmax However, all existing ML algorithms can only provide a certain ρ dmax The prediction results cannot take into account the errors caused by the uncertainty in the prediction process. For example, ρ dmaxThe random errors caused by the uncertainty of the measurement mechanism and the uncertainty of the experimental process, as well as the systematic errors caused by the uncertainty of human cognitive level and the uncertainty of the prediction model itself, greatly affect ρ. dmax The reliability and credibility of the prediction results. Obviously, based on the error of ρ dmax There is bound to be a large deviation between the compaction quality assessment conclusions drawn from the prediction results and the actual situation. It is necessary to propose a new method to make the compaction quality assessment results more reliable.

[0005] Interval prediction, also known as probability prediction, can quantify the uncertainty of point prediction by constructing prediction intervals. With the increasing attention paid to the uncertainty of prediction models, methods such as Delta, Bayesian, MVE, Bootstrap and LUBE have been proposed for constructing prediction model intervals. These methods have also been successfully applied to many fields such as finance, wind power, and energy. The Bootstrap method is a non-parametric statistical method based on resampling technology. Its advantage is that it only relies on the original observed sample data. Due to the unique advantages of the Bootstrap method, it has been widely used and also provides a new method for statistical distribution modeling under limited data conditions of vibration compaction. Therefore, it is necessary to introduce the Bootstrap method on the basis of the existing high-precision ML algorithm to correct the uncertainty problem and further realize the accurate evaluation of the compaction quality of high-speed railway subgrade.

[0006] In summary, it is urgent to develop a new method for evaluating the compaction quality of full-section high-speed railway subgrade based on the ML-interval prediction theory. This method can provide an effective technology for evaluating the compaction quality of high-speed railway subgrade and can be extended to the subsequent subgrade service performance evaluation. Summary of the invention

[0007] In view of this, in order to solve the problems in the background technology, the present invention proposes a full-section evaluation method for high-speed railway roadbed compaction quality based on machine learning and interval prediction theory. This method can quantify ρ dmax By determining and acquiring various uncertainties in the process, reliable and accurate compaction quality assessment results can be quickly obtained, providing an effective technology for evaluating the compaction quality of high-speed railway subgrade.

[0008] The technical solution of the present invention comprises the following steps (such as Figure 1 shown):

[0009] Step 1: Determine the dynamic stiffness K based on indoor vibration compaction test rb Maximum dry density ρ at the turning point dmax ;

[0010] Step 2: Build prediction ρ dmaxPSO-Optimal ML-AdaBoost (POA) model;

[0011] Step 3: Introduce interval prediction theory (Bootstrap) to quantify ρ dmax The uncertainty of the prediction, establishing ρ dmax Bootstrap-POA-ANN (BPA) interval prediction model;

[0012] Step 4: Combine the interval prediction model, spatial interpolation algorithm and compaction standard to establish a full-section evaluation model for high-speed railway subgrade compaction quality.

[0013] The present invention also discloses a non-volatile storage medium, characterized in that the non-volatile storage medium includes a stored program, wherein when the program is running, the device where the non-volatile storage medium is located is controlled to execute the above method.

[0014] The present invention also discloses an electronic device, characterized in that it comprises a processor and a memory; the memory stores computer-readable instructions, and the processor is used to execute the computer-readable instructions, wherein the computer-readable instructions execute the above method when they are executed.

[0015] Beneficial effects:

[0016] This method solves the current problem of determining ρ dmax The mechanisms and methods of high-speed railway subgrade compaction still have problems such as uncertainty, inefficiency and lack of intelligence. These defects will lead to insufficient evaluation of the compaction quality of high-speed railway subgrade, thereby reducing the stability and strength of the subgrade structure, and may cause uneven settlement, track unevenness and other diseases in the later operation process, seriously affecting the operation safety of high-speed railway. In the information age, a new method for full-section evaluation of high-speed railway subgrade compaction quality based on machine learning algorithm and interval prediction theory is proposed. Firstly, based on indoor vibration compaction test and multi-parameter collaborative testing method, a method based on dynamic stiffness K is proposed. rb The turning point ρ dmax Secondly, based on three typical machine learning (ML) algorithms (BPNN, SVR and RF), we developed ρ dmax The PSO-ML-AdaBoost hybrid model is proposed. In addition, the PSO-Optimal ML-AdaBoost (POA) model is determined based on the prediction accuracy and error. Third, the interval prediction theory (Bootstrap) is introduced to quantify ρ dmax Finally, based on the Bootstrap-POA-ANN (BPA) interval prediction model and spatial interpolation algorithm, the full section ρ dmaxThe interval distribution of the full-section compaction quality evaluation model was established based on the compaction standard (95%). The test results based on the real test section show that under the action of this new method for full-section evaluation of high-speed railway roadbed compaction quality based on machine learning and interval prediction theory, the following results are obtained: 1) The PSO-BPNN-AdaBoost model performs well in terms of accuracy and error indicators, so it is selected as ρ dmax POA model. 2)ρ dmax The BPA interval prediction model can construct clear and reliable prediction intervals. 3) The compaction quality full-section evaluation model can provide the full-section distribution interval of compaction degree. Compared with the compaction thickness of 50-60cm and 60-70cm, the compaction quality of the compaction thickness of 40-50cm is better. This method can provide an effective technology for evaluating the compaction quality of high-speed railway subgrade, and can be extended to the subsequent subgrade service performance evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 A flow chart of a new method for full-section evaluation of high-speed railway subgrade compaction quality based on machine learning and interval prediction theory;

[0019] Figure 2 Based on K rb ρ of the “inflection point” dmax Determined schematic diagram;

[0020] Figure 3 is dmax Flow chart of PSO-Optimal ML-AdaBoost (POA) prediction model optimization;

[0021] Figure 4 Establish a flow chart for the BPA interval prediction model;

[0022] Figure 5 Process diagram of the full-section assessment model for high-speed railway subgrade compaction quality;

[0023] Figure 6 It is a schematic diagram of Kriging algorithm;

[0024] Figure 7 The dataset constructed;

[0025] Figure 8The following are the prediction results of three typical point prediction models on the test set;

[0026] Fig. 9 The prediction results of the BPA interval prediction model at three different confidence levels;

[0027] Fig.10 This is an overview of the field test section;

[0028] Fig.11 It is the dry density measured on site at different thicknesses;

[0029] Fig.12 This is the result diagram of dry density interpolation using Kriging algorithm;

[0030] Fig.13 This is the result diagram of interpolating filling parameters using the Kriging algorithm;

[0031] Fig.14 This is the full section interval prediction result diagram of the maximum dry density at different thicknesses;

[0032] Fig.15 This is the full-section interval distribution result diagram of compaction degree at different thicknesses. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0034] A full-section evaluation method for high-speed railway subgrade compaction quality based on machine learning and interval prediction theory can quantify ρ dmax By determining and acquiring various uncertainties in the process, reliable and accurate compaction quality assessment results can be quickly obtained, providing an effective technology for evaluating the compaction quality of high-speed railway subgrade.

[0035] The technical solution of the present invention comprises the following steps (such as Figure 1 shown):

[0036] Step 1: Determine the dynamic stiffness K based on indoor vibration compaction test rb Maximum dry density ρ at the turning point dmax ;

[0037] This step is based on the dynamic stiffness K rb The maximum dry density ρ at the “inflection point” dmax The determination method is divided into two steps:

[0038] a) Compaction quality multi-parameter collaborative test system; The compaction quality multi-parameter collaborative test system is embedded with dry density ρ d and dynamic stiffness K rb The calculation formula is as follows:

[0039]

[0040] Where: m is the mass of the packing; g is the acceleration due to gravity; D c is the inner diameter of the compaction cylinder; h0 is the virtual paving thickness; S n is the filler sedimentation rate; m e is the deadweight of the eccentric block; r e is the eccentricity; ω is the eccentricity speed; Δφ is the lag phase angle, obtained by the Hall sensor; m p is the mass of the counterweight; m d is the mass of the vibration system; x is the displacement of the vibration system, obtained by the displacement sensor; is the acceleration of the vibration system, obtained by the acceleration sensor.

[0041] The vibration compaction test was carried out using an improved large-scale intelligent vibration compactor. The device has four parameters that can be adjusted simultaneously: vibration frequency f (0-80Hz), static load m p (0~400kg), eccentricity r e (0~8cm), eccentric mass m e (0~2.4kg); at the same time, its compaction mechanism is similar to that of an on-site roller. Compared with conventional vibratory compactors, this device is the first to embed displacement sensors, Hall sensors, and acceleration sensors, and outputs the dry density ρ in real time through equations (1) and (2). d Curve and dynamic stiffness K rb curve.

[0042] b) Based on K rb ρ of the “inflection point” dmax Sure

[0043] Get K rb After compacting the curve, determine K rb The vibration time corresponding to the inflection point is taken as the optimal vibration time. d Determine the ρ of the filler on the compaction curve dmax .(like Figure 2 shown).

[0044] Step 2: Build prediction ρ dmax of PSO-OptimalML-AdaBoost( POA) model;

[0045] Prediction ρ dmaxThe idea of ​​establishing the PSO-Optimize ML-AdaBoost (POA) model is mainly based on three typical machine learning (ML) algorithms: BP neural network (BPNN), support vector machine (SVR) and random forest (RF). The PSO hyperparameter optimization algorithm is used to optimize the hyperparameters of the ML model, and the AdaBoost ensemble algorithm is used to improve the generalization and stability performance of the prediction model. It can be divided into three steps (steps such as Figure 3 shown):

[0046] a) Based on the indoor vibration compaction test, the filler characteristic parameters and maximum dry density are obtained, and a data set is further constructed. The data set is divided into a training set and a test set in a ratio of 7:3. The training set is used to establish the prediction model, and the test set is used to test the performance of the prediction model.

[0047] Among them, the gradation characteristics of the filler (d max , b and m), shape characteristics (EI), crushing characteristics (LAA), water absorption (W ac , W af ) as the input feature of the data set, the maximum dry density ρ dmax As the output features of the data set, further construct the prediction model data set

[0048] b) Establishment of the optimal prediction model ML based on the training set;

[0049] The training set is input into three typical PSO-ML-AdaBoost (PMA) models (PSO-BPNN-AdaBoost, PSO-BPNN-AdaBoost, PSO-BPNN-AdaBoost) to perform iterative training.

[0050] c) Evaluation of the optimal prediction model ML based on the test set;

[0051] The test set was input into three typical PSO-ML-AdaBoost models, and the prediction accuracy and error evaluation indicators were used to select the best PMA model, namely the POA model. The prediction accuracy and error evaluation indicators include: goodness of fit R 2 The prediction performance of the model is evaluated by four indicators: explained model variance score EVS, mean square error MSE and mean absolute error MAE. The specific calculation formulas are shown in formulas (3) to (6).

[0052] Goodness of fit R 2, explained model variance score EVS, mean square error MSE and mean absolute error MAE four indicators to evaluate the prediction performance of the model, the formula is as follows:

[0053]

[0054] Where N is the number of prediction samples, y t is the actual dmax value, is the predicted dmax Value, Var{} is the variance function; is the actual dmax Mean.

[0055] Step 3: Introduce interval prediction theory (Bootstrap) to quantify ρ dmax The uncertainty of the prediction, establishing ρ dmax Bootstrap-POA-ANN (BPA) interval prediction model;

[0056] The interval prediction model is divided into five aspects: generation of pseudo data sets, model training and storage, prediction mean and cognitive error variance estimation, random error variance estimation, and uncertainty prediction and accuracy assessment. The detailed steps are as follows (the specific process is as follows Figure 4 shown):

[0057] 1) Generation of pseudo datasets

[0058] Using the Bootstrap method, resample the training set to N train Times (N train is the total number of samples in the training set) to obtain the pseudo training set TR*. Repeat the above steps A times to obtain A pseudo training sets, thereby completing the generation of the pseudo training set.

[0059] 2) POA model training and storage

[0060] Input the pseudo training set into the POA model, keep the hyperparameters and structure of the POA model unchanged. Train the model, save the model after each iteration, repeat the above steps A times, and get A trained POA models.

[0061] 3) Estimation of the prediction mean and cognitive error variance

[0062] First, the test set is input into A trained POA prediction models to obtain A ρ dmax The predicted result is shown in formula (7). dmax The predicted result can be obtained under the condition of small deviation. dmax Estimates of the true regression The calculation formula is as shown in formula (8).

[0063] Assuming that the maximum dry density prediction of the POA model is unbiased, the variance of the cognitive error is It can be estimated by the variance of A point prediction values, and the calculation method is shown in formula (9).

[0064] Input the test set into A trained POA prediction models to obtain A ρ dmax The predicted results are as follows:

[0065]

[0066] In the formula, is the predicted value of the ath maximum dry density of the nth sample, Represents the mean of the A maximum dry density prediction values ​​of the nth sample.

[0067] Average A ρ dmax The predicted result can be obtained under the condition of small deviation. dmax Estimates of the true regression The calculation formula is as follows:

[0068]

[0069] In the formula, Represents the maximum dry density prediction mean of the nth sample A prediction models after eliminating errors.

[0070] 3) Estimation of cognitive error variance.

[0071] Assuming that the maximum dry density prediction of the POA model is unbiased, the variance of the cognitive error is The variance of the predicted values ​​at A points can be used to estimate the value, and the calculation method is as follows:

[0072]

[0073] 4) Random error variance estimation

[0074] ① Determine the variance of the model’s random error The estimation formula is as follows:

[0075]

[0076] In the formula, is the variance of cognitive error, and E is the mathematical expectation formula.

[0077] ②Construct a square residual data set.

[0078] To better estimate Based on formula (10), a set of square residual sequence r is reconstructed 2 (x n ) to fit The calculation formula is shown in formula (11):

[0079]

[0080] Combine it with the predicted input feature vector x n Combination can create a squared residual data set Refer to the data set division method in the POA model module, and the square residual training set can be obtained after dividing it in a ratio of 7:3 Sum squared residual test set

[0081] ③ Improvement of ANN model loss function and Prediction. Based on An ANN model can be trained to predict During the training process, the goal is to maximize the r_train Observed in The probability of the sample. The concept of maximum likelihood estimation is introduced to enhance the loss function Loss of the ANN model. ε :

[0082]

[0083] In the formula, ln() is the logarithm operation with base e, r 2 (x n ) is the square residual sequence of the maximum dry density prediction of the sample.

[0084] by Based on calculable The ANN model used for prediction, during the training process, the prediction goal is to make the observed samples exist The probability of occurrence in the ANN model is the largest, and the traditional mean square error MSE or mean absolute error MAE cannot achieve this goal. Therefore, the idea of ​​maximum likelihood estimation (MLE) can be introduced to calculate the loss function Loss of the ANN model. ε Improvements are made to obtain the improved ANN model loss function Loss ε As shown in formula (12). It should be noted that in order to ensure the prediction output of the ANN model is always positive, the activation function of the output layer must be set to an exponential function (such as exponential function, softplus function, etc.). This method uses the softplus activation function. After training, the squared residual test set Input the model and get the result

[0085] 5) Uncertainty prediction and accuracy assessment

[0086] ① Uncertainty prediction

[0087] Assuming the significance level is α, ρ dmax The construction of the prediction interval can be expressed as

[0088]

[0089] In the formula, is the upper limit of the maximum dry density prediction interval, is the lower limit of the maximum dry density prediction interval, C confidence is the confidence level of the maximum dry density prediction interval, and its value depends on the confidence level (1-α)*100%. When the confidence level is 90%, 95%, and 99%, C confidence They are 1.65, 1.96, and 2.58 respectively. For specific solutions, please refer to the quantile table of standard normal distribution.

[0090] ② Accuracy evaluation

[0091] PICP, MPIW and CWC are used to evaluate the performance of the prediction interval. The calculation formula is as follows:

[0092]

[0093] CWC=MPIW[1+γ(PICP)*e η(PINC-PICP) ] (16)

[0094] Where N test is the total number of samples in the test set; c t It is a Boolean value. If the output target falls within the prediction interval, it takes 1, otherwise it takes 0. is the upper limit of the maximum dry density prediction interval; is the lower limit of the prediction interval of the maximum dry density; γ(PI CP) is a Boolean variable determined by the PICP value; PINC is the nominal confidence of the prediction interval, PINC=(1-α)*100%; η is the penalty parameter. If PI CP<PI NC, γ(PI CP)=1, otherwise it is 0; PINC is the nominal confidence of the prediction interval, PINC=(1-α)*100%; η is the penalty parameter, which determines the degree of penalty together with PINC, and its function is to amplify the small difference between PICP and PI NC.

[0095] Step 4: Combine the interval prediction model, spatial interpolation algorithm and compaction standard to establish a full-section evaluation model for high-speed railway subgrade compaction quality.

[0096] Combined with the BPA interval prediction model to predict the distribution interval of the maximum dry density, the Kriging interpolation algorithm is used to realize full-section data collection, and combined with the 95% compaction standard specified in the high-speed railway subgrade design specification, a full-section evaluation model for high-speed railway subgrade compaction quality is established. It includes the following steps (the specific process is as follows Figure 5 shown):

[0097] 1) Obtain various material source parameters (including filler gradation characteristics (d max , b and m), shape characteristics (EI), crushing characteristics (LAA), water absorption (W ac , W af )) and dry density ρ d data;

[0098] 2) Combine the measured material source parameters and dry density data with the spatial coordinate data of the test points, and use the Kriging interpolation algorithm to interpolate the data of the unknown test points to achieve full-section data collection. The schematic diagram of the Kriging algorithm is shown in Figure 6 As shown;

[0099] The full-section data collection mainly adopts the Kriging interpolation algorithm, and its core calculation formula is as follows:

[0100]

[0101] In the formula, Z * (x0) is the estimated value at the point (x0, y0), λ i is the weight corresponding to the i-th surrounding observation.

[0102] 3) After obtaining the full-section material source parameter data, input the BPA interval prediction model, keep its model parameters unchanged, and implement the prediction to achieve ρ dmax Full section interval prediction;

[0103] 4) Based on the compaction degree K calculation formula (20), the compaction quality evaluation of the entire section is realized; the compaction degree K calculation formula is used to realize the compaction quality evaluation of the entire section:

[0104]

[0105] Where K is the compaction degree, ρ d is the dry density, ρ dmax is the maximum dry density. The ρ of the whole section is obtained d and ρ dmax Then, based on the compaction degree K calculation formula (10), the compaction degree of the entire section is calculated, thereby realizing the compaction quality assessment of the entire section.

[0106] The embodiment and implementation process of the complete method according to the present invention are as follows:

[0107] 1) Using the self-developed multi-parameter collaborative testing compaction quality system to carry out indoor vibration compaction tests on roadbed fillers, and based on the dynamic stiffness K rb The maximum dry density determination method of the turning point is obtained as ρ dmax data.

[0108] Since the surface layer of the roadbed is most affected by high-speed train loads and external environmental factors. Generally, the surface layer of China's high-speed railway roadbed is compacted with graded crushed stone filler with good strength and deformation characteristics. Insufficient compaction of graded crushed stone filler will greatly reduce the stability and strength of the roadbed structure. Therefore, the present invention takes graded crushed stone filler as the research object for subsequent experiments and analysis. According to the optimal vibration compaction parameters determined in the previous experiments, a large number of vibration compaction tests were carried out to establish the characteristic parameters of graded crushed stone filler and ρ dmax Database, such as Figure 7 As shown. It can be seen that the ρ of graded crushed stone filler dmax With d max , m, EI and LAA were negatively correlated, but were positively correlated with b, W ac and W af Positive correlation. dmax The properties (including d max ,m,b,EI,LAA,W ac and W af ) as the input parameter of the prediction model, ρ dmax as an output parameter.

[0109] 2) Establish prediction ρ dmax The PMA model is then used to obtain the POA model.

[0110] a) Input the training set into three typical PSO-ML-AdaBoost (PMA) models (PSO-BPNN-AdaBoost, PSO-BPNN-AdaBoost, PSO-BPNN-AdaBoost) and perform iterative training.

[0111] b) Input the test set into three typical PSO-ML-AdaBoost models and obtain the prediction results of each model as follows Figure 8 The prediction performance evaluation results are shown in Table 1.

[0112] Table 1 Point prediction accuracy evaluation results

[0113]

[0114] Depend on Figure 8 It is clear that all three PMA models havedmax The PBA model showed good performance in all aspects, and the predicted values ​​were in good agreement with the overall trend of the measured values. 2 The PBA model is most closely around the measured curve, followed by the PSA model (0.9453) and then the PRA model (0.9330). In addition, Table 1 shows the comparison of the accuracy evaluation results. It is obvious that the PBA model outperforms the other models in both prediction accuracy and error indicators. Therefore, the PBA model is selected as ρ dmax POA model.

[0115] 3) Refer to the process of step 3) in the manual. On the basis of determining the POA model, introduce the interval prediction theory (Bootstrap) to quantify ρ dmax The uncertainty of the prediction, establishing ρ dmax The Bootstrap-POA-ANN (BPA) interval prediction model is used. Furthermore, we obtain ρ based on the BPA model. dmax The interval prediction results are as follows Fig. 9 The accuracy evaluation results are shown in Table 2.

[0116] like Fig. 9 As shown, it is clear that the prediction interval effectively contains ρ dmax The measured curve, and ρ dmax The measured values ​​are basically within the prediction interval, which shows that the reliability of the interval prediction results is very high. In addition, the overall width of the prediction interval is uniform and increases with the increase of confidence. As shown in Table 2, as the confidence level increases, both PICP and CWC increase. In addition, at three different confidence levels, the PICP value not only exceeds the corresponding confidence level, but its average value also exceeds 95%, which shows that the interval prediction results obtained by the BPA model are highly reliable. It is worth noting that the higher the PICP, the smaller the MPIW, and the larger the CWC, indicating that the prediction results are more accurate and reliable. At a confidence level of 95%, PICP, MPIW, and CWC are 100%, 0.4690 g / cm 3 and 0.4690g / cm 3 Although PICP is higher than the 90% confidence level and equal to the 99% confidence level, both MPIW and CWC are lower than the 99% confidence level. Therefore, the 95% confidence level can be selected as the core parameter of the BPA model for subsequent compaction quality assessment.

[0117] Table 2 Maximum dry density interval prediction accuracy evaluation

[0118]

[0119] 4) Carry out experiments based on specific engineering cases. Combined with the BPA interval prediction model to predict the distribution interval of the maximum dry density, the Kriging interpolation algorithm is used to realize full-section data collection, and combined with the 95% compaction standard specified in the high-speed railway subgrade design specification, a full-section evaluation model for the compaction quality of the high-speed railway subgrade is established.

[0120] a) On-site data acquisition

[0121] Before large-scale paving and compaction, it is necessary to conduct a test section construction to optimize the construction process parameters. The optimal paving thickness H0 is a key construction process parameter, which affects both the improvement of construction efficiency and the precise control of compaction quality. Fig.10 As shown, this section takes the compaction test section of a station in the southwest region as an example to verify the superiority of the proposed method. Three typical test points (10m long and 3m wide) were selected in the test section, and different paving thicknesses were designed: low thickness (40-50cm), medium thickness (50-60cm) and high thickness (60-70cm). The vibration compaction process parameters in the test are as follows: dead weight 23t, working width 2.15m, vibration frequency 32Hz, amplitude 1.03mm. After compacting 4 to 5 times in different test sections, the surface settlement of the test section tended to be stable. Test pits were selected every 1m at each test point to measure the filler parameters and ρ d .

[0122] b) Full-section data acquisition based on Kriging algorithm

[0123] ① Measured dry density results

[0124] Fig.11 shows ρ for three different paving thicknesses d Measurement results. It can be seen that under the same vibration compaction energy, reducing the paving thickness can significantly increase ρ d When the paving thickness is low, the measured ρ d The distribution is relatively uniform. However, as the paving thickness increases, ρ d will decrease, and the measured ρ d The distribution also becomes uneven, and the ρ at the edge of the test section d Significantly lower than the internal ρ d Therefore, in the on-site compaction quality inspection, it is necessary to strictly control the "weak" area of ​​the compaction edge.

[0125] ②Interpolation dry density results

[0126] like Fig.12 As shown in the figure, on the test section with a paving thickness of 40-50 cm, the full-section measurement value ρ was obtained using the Kriging interpolation algorithm. d spatial distribution.

[0127] ③ Full-section surface source parameter interpolation distribution results

[0128] In a similar way, the filler in the test pit was subjected to screening test, water absorption test, particle shape rapid scanning test and Losangele abrasion test to obtain d max , b, m, W ac , W af , EI and LAA parameters. Then, the full cross-sectional distribution of each filler parameter was obtained using the Kriging interpolation algorithm, as shown in the following figure. Fig.13 shown.

[0129] c)ρ dmax Full section interval prediction

[0130] The obtained filler parameters are input into the PBA model to obtain the ρ under different paving thicknesses. dmax The full cross-sectional distribution of Fig.14 It is obvious that under different paving thickness, ρ dmax The upper limit is between 2.30 and 2.45 g / cm 3 In the test section with H0 = 40-50 cm, due to the smaller thickness, it is easier to achieve uniform paving, resulting in ρ dmax The lower limit is between 2.2 and 2.3 g / cm 3 However, the greater the thickness, the more challenging it is to control the uniformity of the packing, resulting in serious deviations of the packing parameters from the design criteria, ρ dmax The lower limit is between 2.0 and 2.3 g / cm 3 between.

[0131] d) Full-section compaction quality assessment

[0132] In addition, by converting the obtained ρ dmax The full-section distribution interval results are integrated with the full-section dry density distribution results to determine the full-section distribution interval results of compaction under different paving thicknesses, such as Fig.15As shown. Obviously, the proposed method can intuitively, accurately and comprehensively evaluate the compaction quality of the entire section of the base. In the test section with H0 = 40-50 cm, the upper and lower limits of the compaction degree exceeded 95% of the compaction quality standard, which shows that selecting a paving thickness in the range of 40-50 cm can obtain a good roadbed structure during the on-site compaction process. However, when a paving thickness of 50-60 cm is selected, the lower limit of the compaction degree cannot reach the compaction quality standard of 95%. This shows that within this thickness range, there may be areas with insufficient compaction, and the compaction quality of the roadbed structure cannot be guaranteed. Similarly, when a paving thickness of more than 60 cm is selected, the upper limit of the compaction degree does not reach the compaction quality standard of 95%. This shows that within this thickness range, there are many places where the compaction degree is insufficient, and it is difficult to guarantee the performance of the roadbed.

[0133] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A full-section evaluation method for high-speed railway subgrade compaction quality based on machine learning and interval prediction theory, It is characterized in that The following steps are involved: Step 1: Determine the dynamic stiffness K based on indoor vibration compaction test rb Maximum dry density ρ at the turning point dmax ; Step 2: Build prediction ρ dmax of PSO-OptimalML-AdaBoost , namely the POA model; Step 3: Introduce the interval prediction theory Bootstrap quantification ρ dmax The uncertainty of the prediction, establishing ρ dmax Bootstrap-POA-ANN, i.e., BPA interval prediction model; Step 4: Combine the interval prediction model, spatial interpolation algorithm and compaction standard to establish a full-section evaluation model for high-speed railway subgrade compaction quality.

2. According to claim 1, a method for evaluating the full section of high-speed railway subgrade compaction quality based on machine learning and interval prediction theory, Features: Described step 1 comprises the following steps: 1) Compaction quality multi-parameter collaborative testing system; 2) Dry density ρ d and dynamic stiffness K rb Curve drawing; 3) Based on K rb ρ of the "inflection point" dmax Sure.

3. The evaluation method according to claim 2 is characterized in that: the compaction quality multi-parameter collaborative test system is embedded with the dry density ρ d and dynamic stiffness K rb The calculation formula is as follows: Where: m is the mass of the filler; D c h is the inner diameter of the compaction cylinder; 0 S is the virtual paving thickness; n is the filler sedimentation rate; m e is the deadweight of the eccentric block; r e is the eccentricity; ω is the eccentricity speed; Δφ is the lag phase angle, obtained by the Hall sensor; m d is the mass of the vibration system; x is the displacement of the vibration system, obtained by the displacement sensor; is the acceleration of the vibration system, obtained by the acceleration sensor, m p is the mass of the counterweight, and g is the acceleration due to gravity.

4. The evaluation method according to claim 1, Features: In step 2, the prediction ρ is established dmax The PSO-OptimalML-AdaBoost, or POA model, includes the following steps: 1) Establishment and division of data sets; 2) ML model establishment based on training set; 3) Evaluation of the optimal ML model based on the test set.

5. The evaluation method according to claim 4, Features: The optimal ML model evaluation based on the test set adopts the goodness of fit R 2 , explained model variance score EVS, mean square error MSE and mean absolute error MAE four indicators to evaluate the prediction performance of the model, the formula is as follows: Where N is the number of prediction samples, y t is the actual dmax value, is the predicted dmax value, Var{} is the variance function, y t is the actual dmax Mean.

6. The evaluation method according to claim 1, Features: In step 3), establish dmax The Bootstrap-POA-ANN, i.e., the BPA interval prediction model, includes the following steps: 1) Generation of pseudo datasets; 2) Model training and preservation; 3) Estimation of the prediction mean and cognitive error variance; 4) Random error variance estimation; 5) Uncertainty prediction and accuracy assessment.

7. The evaluation method according to claim 6, Its characteristics are: The prediction mean and cognitive error variance estimation includes two steps: prediction mean calculation and cognitive error variance estimation: 1) Forecast mean calculation Input the test set into A trained POA prediction models to obtain A ρ dmax The predicted results are as follows: In the formula, is the predicted value of the ath maximum dry density of the nth sample, represents the mean of the A maximum dry density prediction values ​​of the nth specimen; Average A ρ dmax The predicted result is ρ under small deviation conditions dmax Estimates of the true regression The calculation formula is as follows: In the formula, Represents the maximum dry density prediction mean of the nth sample A prediction models after eliminating errors; 2) Cognitive Error Variance Estimation Assuming that the maximum dry density prediction of the POA model is unbiased, the variance of the cognitive error is Use the variance of A point prediction values ​​to estimate, and the calculation method is as follows:

8. The evaluation method according to claim 6 is characterized in that the random error variance estimation comprises the following steps: 1) Determine the variance of the model’s random error The estimation formula is as follows: In the formula, is the variance of cognitive error, E is the mathematical expectation formula; 2) Construct a squared residual data set To better estimate Based on formula (10), a set of square residual sequences is reconstructed to fit In the formula, and From equations (8) and (9), we can get r 2 (x n ) is the square residual sequence of the maximum dry density prediction of the sample, which is combined with the prediction input feature vector x n Combination can create a squared residual data set Refer to the data set division method in the POA model module to obtain the square residual training set Sum squared residual test set 3) Improvement of ANN model loss function and predict based on Train an ANN model to predict During the training process, the goal is to maximize the r_train Observed in The probability of the sample; the concept of maximum likelihood estimation is introduced to enhance the loss function Loss of the ANN model ε : In the formula, ln( ) is the logarithm with base e, r 2 (x n ) is the square residual sequence of the maximum dry density prediction of the sample.

9. The evaluation method according to claim 6, characterized in that the uncertainty prediction and accuracy evaluation comprises the following steps: 1) Uncertainty prediction Assuming the significance level is α, ρ dmax The construction of the prediction interval can be expressed as In the formula, is the upper limit of the maximum dry density prediction interval, is the lower limit of the maximum dry density prediction interval, C confidence is the confidence level of the maximum dry density prediction interval; 2) Accuracy evaluation PICP, MPIW and CWC are used to evaluate the performance of the prediction interval. The calculation formula is as follows: CWC=MPIW[1+γ(PICP)*e η(PINC-PICP) ] (16) Where N test is the total number of samples in the test set; c t It is a Boolean value. If the output target falls within the prediction interval, it takes 1, otherwise it takes 0. is the upper limit of the maximum dry density prediction interval; is the lower limit of the maximum dry density prediction interval; γ(PICP) is a Boolean variable determined by the PICP value; PINC is the nominal confidence of the prediction interval, PINC = (1-α) * 100%; η is the penalty parameter.

10. The evaluation method according to claim 1, Features: The step 4) in which a high-speed railway subgrade compaction quality full-section assessment model is established comprises the following steps: 1) Obtain the material source parameters and dry density ρ on site through field tests d data; 2) Full-section data acquisition based on spatial interpolation algorithm; 3) After obtaining the full-section material source parameter data, input the BPA interval prediction model to achieve ρ dmax Full cross-section interval prediction; 4) Based on the compaction degree K calculation formula, the full-section compaction quality assessment is achieved.

11. The evaluation method according to claim 10, Its characteristics are: The full-section data collection based on the spatial interpolation algorithm adopts the Kriging interpolation algorithm (Kriging), and its core calculation formula is as follows: In the formula, Z * (x 0 ) is the point (x 0 ,y 0 ), λ i is the weight corresponding to the i-th surrounding observation.

12. The evaluation method according to claim 10, characterized in that: the calculation formula of the compaction degree K is as follows: Where K is the compaction degree, ρ d is the dry density, ρ dmax is the maximum dry density, and the ρ of the whole section is obtained d and ρ dmax Then, based on the compaction degree K calculation formula (10), the compaction degree of the entire section is calculated, thereby realizing the compaction quality assessment of the entire section.

13. A non-volatile storage medium, It is characterized in that The non-volatile storage medium includes a stored program, wherein when the program is executed, the device where the non-volatile storage medium is located is controlled to execute the method according to any one of claims 1 to 12.

14. An electronic device, It is characterized in that It comprises a processor and a memory; the memory stores computer-readable instructions, and the processor is used to execute the computer-readable instructions, wherein the computer-readable instructions execute the method described in any one of claims 1 to 12 when executed.

Images