Intelligent detection method and device for multivariable roadbed compactness based on knowledge enhancement neural network

Abstract

The invention discloses a multivariable roadbed compactness intelligent detection method based on a knowledge enhancement neural network. The method comprises the following steps: obtaining an original sample data set; screening the original sample data set to output an optimal sample data set, performing label labeling on sample data in the optimal sample data set based on the compactness, and forming a final data set by the optimal sample data set and labels; constructing a corresponding prediction network based on a knowledge enhancement neural network framework; training a prediction network by using the final data set to obtain a compactness prediction model; and inputting sample data collected on the road roller into the compaction degree prediction model to predict the corresponding compaction degree, and combining the predicted compaction degree with the position information of the road roller to output a corresponding compaction quality cloud picture. The invention further provides a multivariable roadbed compactness intelligent detection device. According to the method provided by the invention, the roadbed compaction detection cost can be greatly reduced, and the detection efficiency is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of roadbed compaction detection, and in particular relates to a multivariable roadbed compaction intelligent detection method and device based on knowledge enhanced neural network. Background Art

[0002] Subgrade compaction is one of the most important control indicators in road construction, which directly affects the road's bearing capacity, service life and driving safety. High-quality subgrade compaction can effectively reduce road settlement, deformation and cracks, and improve road durability and safety. However, due to the complex environment and changeable soil conditions at the construction site, the subgrade compaction is often difficult to meet the expected standards, so it is necessary to strictly test and control the subgrade compaction during the subgrade construction process.

[0003] The emergence of intelligent compaction technology has alleviated the shortcomings of the above-mentioned traditional compaction technology to a certain extent. Intelligent compaction realizes real-time monitoring of the compaction degree of the working area by integrating advanced measurement systems, positioning systems, and on-board computers and control systems on the roller. Intelligent compaction measurement values ​​(ICMV) are used in intelligent compaction technology to determine the quality of roadbed compaction, such as compaction value (CMV) and machine vibration power (MDP). However, a single indicator only shows sufficient accuracy within a specific applicable range. For example, the compaction value is mainly applicable to fine-grained soil. Moreover, the factors considered in the establishment of a single indicator are relatively single, and the complex influencing factors of roadbed compaction at the construction site are often not considered, that is, it cannot accurately reflect the actual compaction condition of the roadbed soil at the site. In addition, the potential relationship between intelligent compaction measurement values ​​and compaction quality is still unclear.

[0004] Traditional methods for testing roadbed compaction, such as the sand filling method, the knife ring method, and the core drilling method, have some disadvantages: (1) They require dedicated personnel to operate the sampling, which is inefficient, and the testing method is destructive and will cause certain damage to the roadbed; (2) Traditional testing methods use random sampling, and the limited number of random inspection points cannot reflect the compaction quality of the entire construction area, that is, the sampling points may not be representative; (3) It is impossible to evaluate the compaction quality in real time, and the test samples can only be sent to the laboratory for evaluation after the project is completed. If they are unqualified, the project will be reworked.

[0005] Patent document CN 212956487U discloses a roadbed and pavement intelligent compaction effect test device, comprising a rectangular frame body (1), on which are provided a longitudinal rod (2), a transverse rod (3), a supporting hook (4), a universal wheel (5) and a directional wheel (6), the directional wheel (6) being connected to a rotary encoder (7); a group of connected connecting rods (9) and a test wheel (11) are fixed below the longitudinal rod (2) and the transverse rod (3), the test wheel (11) being connected to an acceleration sensor (12); springs (13) are connected between the connecting rod (9) and the longitudinal rod (2) and the transverse rod (3); a data acquisition terminal (14) is fixed on the frame body (1), and the rotary encoder (7) and the acceleration sensor (12) are respectively connected to the data acquisition terminal (14) for communication.

[0006] This method uses a roller to detect the current roadbed, but the indicators used are single and cannot accurately reflect the complex compaction influencing factors at the construction site. In addition, the indicators considered by the existing multi-index evaluation method are still limited. Summary of the invention

[0007] The purpose of the present invention is to provide a multivariate roadbed compaction intelligent detection method and device based on knowledge enhanced neural network, which can greatly reduce the cost of roadbed compaction detection and improve detection efficiency.

[0008] In order to achieve the first object of the present invention, the following technical solution is provided: a multivariate roadbed compaction intelligent detection method based on knowledge enhanced neural network, comprising the following steps:

[0009] Acquire an original sample data set, wherein the sample data set includes compaction parameters and intelligent compaction measurement parameters;

[0010] The original sample data set is screened to output the best sample data set, and the sample data in the best sample data set is labeled based on the compaction degree, and the best sample data set and the label are combined into a final data set;

[0011] The corresponding prediction network is constructed based on the knowledge-enhanced neural network framework, which includes a feature extraction module, a mapping module and a prediction module:

[0012] The feature extraction module is used to extract data features from input sample data;

[0013] The mapping module converts the extracted data features through a preset display expression to obtain a corresponding mapping result;

[0014] The prediction module performs analysis based on the input mapping result to output a prediction result;

[0015] The prediction network is trained using the final data set to obtain a compaction prediction model;

[0016] The sample data collected on the roller is input into the compaction prediction model to predict the corresponding compaction, and the predicted compaction is combined with the position information of the roller to output the corresponding compaction quality cloud map.

[0017] The present invention utilizes compaction parameters and continuous vibration signals collected during the compaction operation of the roller to predict the roadbed compaction quality in real time, which can significantly reduce the cost of roadbed compaction detection and improve detection efficiency.

[0018] Specifically, the compaction parameters include roller vibration amplitude, vibration frequency, rolling speed, number of rolling passes, loose paving thickness, compacted soil particle grading, soil initial moisture content or soil compaction degree.

[0019] Specifically, the intelligent compaction measurement parameters include one or more of compaction degree value, compaction control value, acceleration intelligent compaction value, total harmonic distortion, vibration peak acceleration, roller vibration acceleration root mean square or roller static driving force.

[0020] Specifically, the process of screening the original sample data set is as follows:

[0021] The original sample data set is input into the pre-built KNN model to obtain the prediction results, and the prediction results are compared with the compaction degree to remove the outliers that do not meet the threshold to obtain a valid sample data set;

[0022] And the effective sample data set is screened by using correlation analysis and feature recursive elimination method to obtain the best sample data set.

[0023] Specifically, the display expression includes linear and nonlinear relationships between input data features and compaction degree, and the linear and nonlinear relationships are constructed by genetic algorithm and character regression algorithm.

[0024] Specifically, during the training process, the data in the final data set is subjected to feature extraction and conversion, and the extracted features are scaled to between 0 and 1 using the maximum and minimum feature scaling technique to help establish a knowledge-enhanced neural network.

[0025] Specifically, during the training process, the loss function is used to update the parameters of the prediction network. The expression of the loss function is as follows:

[0026]

[0027] Where MSE stands for mean square error, which is used to measure the error between the model prediction value and the true value, n is the number of samples, y pred is the predicted value, ytrue is the true value.

[0028] In order to achieve the second purpose of the present invention, the following technical solution is provided: a multivariable roadbed compaction intelligent detection device, used to execute the steps of the above-mentioned multivariable roadbed compaction intelligent detection method based on knowledge enhanced neural network.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] By considering the influence of compaction parameters and intelligent compaction measurement parameters on compaction degree, a multivariate roadbed compaction quality evaluation model was established, which effectively solved the problems of single evaluation index and lack of consideration of soil characteristics in the existing roadbed intelligent compaction evaluation method.

[0031] The distance between samples is calculated by the K-NearestNeighbor model, and the original samples are screened by the threshold method. In the screening process, the original samples with large errors caused by test conditions and human factors can be removed, thus forming a high-quality data set; further, correlation analysis and feature recursive elimination method are used to select the most effective input features, and on this basis, genetic algorithm and symbolic regression algorithm are used to obtain the displayed prediction expression, which can be used to quickly predict the roadbed compaction quality while ensuring a certain accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flow chart of the multivariable roadbed compaction intelligent detection method provided in this embodiment;

[0033] Figure 2 A schematic diagram of the overall framework of the knowledge-enhanced neural network provided in this embodiment;

[0034] Figure 3 The cloud map provided in this embodiment is a layout diagram of the measuring point scheme. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in 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.

[0036] like Figure 1 As shown, a multivariable roadbed compaction intelligent detection method provided by this embodiment specifically includes:

[0037] Step 1: According to the laboratory compaction test results, a multivariate original data set Data_ori is established; the multivariate includes a compaction parameter X1 and an intelligent compaction measurement value parameter X2;

[0038] The compaction parameters X1 include the vibration amplitude a of the roller, the vibration frequency f, the rolling speed v, the number of rolling passes n, the loose paving thickness h, the compacted soil particle size distribution d, the initial soil moisture content w and the soil compaction degree R;

[0039] Intelligent compaction measurement parameter X2: compaction value CMV, compaction control value CCV, acceleration intelligent compaction value AICV, total harmonic distortion THD, vibration peak acceleration a p , RMS vibration acceleration of roller a rms , static driving force MDP of roller.

[0040] S1.1. Set up a simulated rolling test on the roadbed in the laboratory, and measure the compacted soil parameters in advance before compaction, including the loose thickness of the roadbed h, the particle size distribution of the compacted soil d, and the initial moisture content of the soil w; during the compaction process, collect the acceleration signal of the roller vibration wheel in real time according to the set sampling frequency and record the compaction parameters of the roller, including the vibration amplitude a, vibration frequency f, rolling speed v, and number of rolling passes n;

[0041] S1.2. Perform Fourier transform on the collected acceleration signal to obtain the vibration wheel acceleration frequency domain signal, and then calculate CMV, CCV, AICV, THD, and a based on the acceleration time domain characteristic diagram and frequency domain characteristic diagram. p 、a rms , MDP.

[0042] The method for collecting compaction parameters X1 in the above S1.1 is: setting up a roadbed on-site simulation rolling test in the laboratory, and determining the compacted soil parameters in advance before compaction, including the loose thickness of the roadbed h and the initial moisture content of the soil w; during the compaction process, the roller compaction parameters are collected and recorded through the vibrating roller control system and the on-site roadbed compaction test plan, including the roller vibration amplitude a, vibration frequency f, rolling speed v, and rolling number n; the soil compaction degree R and the compacted soil particle size distribution d are determined by sand filling method, drying method, compaction test and screening test in accordance with the provisions of the "Highway Geotechnical Test Code of the People's Republic of China Industry Standard JTG 3430-2020".

[0043] The method for obtaining the intelligent compaction measurement value parameter X2 in the above S1.2 is: installing an acceleration sensor on the vibration wheel and collecting the vibration wheel acceleration signal in real time according to the set sampling frequency, using a bandpass filter to filter the noise generated by the on-site construction environment and the roller, and obtaining the roller vibration wheel acceleration time domain signal; then using Fourier transform to convert the vibration wheel acceleration time domain signal to obtain the vibration wheel acceleration frequency domain signal; then based on the acceleration time domain signal and frequency domain signal, calculate CMV, CCV, AICV, THD, a p 、a rms and MDP.

[0044] The intelligent compaction measurement parameters X2 in step 1 include CMV, CCV, AICV, THD, a p 、a rms The specific calculation formula of MDP is as follows:

[0045] The calculation formula of compaction value CMV is:

[0046]

[0047] In formula (1), C CMV A is the calibration coefficient of the vibration compaction value, usually 300; ω and A 2ω are the fundamental amplitude and the second harmonic amplitude respectively.

[0048] The calculation formula of compaction control value CCV is:

[0049]

[0050] In formula (2), C CCV A is the field calibration coefficient of the compaction control value, usually taken as 100; kω is k times the harmonic amplitude, and X is 0.5, 1.5, 2.5 and 3 respectively.

[0051] The calculation formula of acceleration intelligent compaction value AICV is:

[0052]

[0053] In formula (3), C AICV is a constant, take 100.

[0054] The calculation formula for total harmonic distortion THD is:

[0055]

[0056] In formula (4), C THD is a constant, take 100; A Mω is M times the harmonic amplitude, where M is a positive integer greater than or equal to 2.

[0057] Peak acceleration of roller vibration a p The calculation formula is:

[0058] a p =max{a i}(i=1,2,3,…,n) (5)

[0059] In formula (5), a i is the vibration acceleration of the ith sample, and n is the number of samples collected.

[0060] Roller vibration acceleration root mean square a rms The calculation formula is:

[0061]

[0062] The calculation formula of the static driving force MDP of the roller is:

[0063]

[0064] In formula (7), P g is the power of the mobile machine; W is the weight of the vibrating wheel; a is the acceleration of the vibrating wheel; g is the acceleration of gravity; v is the moving speed of the vibrating wheel; θ is the slope; m and b are the internal loss coefficients of the roller.

[0065] Step 2: Sample screening and processing of the original data set;

[0066] S2.1, collect sample data of 15 parameters obtained in step 1; after the sample data is collected, remove outliers through the K-NearestNeighbor (KNN) model to obtain a valid sample data set Data_val;

[0067] The specific steps of removing outliers from the original data set through the KNN model are as follows: training the KNN model to predict the compaction degree, comparing the ratio of the model prediction value to the actual compaction value with the threshold, iteratively removing outliers, and obtaining the valid sample data set Data_val.

[0068] S2.2, using correlation analysis and feature recursive elimination method to select the most useful input features for the knowledge enhanced neural network model to obtain the best sample data set Data_best;

[0069] The most effective input features are selected by using correlation analysis and feature recursive elimination method. Specifically, the correlation coefficient r between the 15 indicators is calculated, and the 14 indicators except the soil compaction degree R are screened by the correlation coefficient r. The indicators with redundant information and small correlation coefficient r with the soil compaction degree R are screened out, and the m most relevant features are obtained. Then, the most useful feature subset is determined by feature recursive elimination method, and the best sample data set Data_best is obtained.

[0070] The above S2.1 is specifically:

[0071] A KNN model is established to predict compaction, and the model's hyperparameters are selected by default. The model's predicted value is compared with the actual compaction value. Data points with a ratio less than 0.67 or greater than 1.5 are considered outliers. These outliers are removed and the K-NearestNeighbor model is retrained using a new data set. This is repeated until no new abnormal data points are found, and finally a valid sample data set Data_val is obtained.

[0072] The above S2.2 specifically means:

[0073] S2.2.1. Calculate the correlation coefficient matrix between the 14 indicators and the soil compaction degree K, screen the 14 indicators by the correlation coefficient, and screen out the indicators whose correlation coefficient r between the 14 indicators is greater than the threshold and whose correlation coefficient r with the soil compaction degree R is smaller, and finally obtain the m most useful input features.

[0074] S2.2.2. Take the screened m features as the initial feature subset, use the XGboost model to calculate the importance of each feature, and use the ten-fold cross validation method to get the prediction accuracy of the initial feature subset; then remove one or more worst features from the current feature subset to get a new feature subset; use the XGboost model again to calculate the importance of each feature in the new feature subset, and use the ten-fold cross validation method to get the classification accuracy of the new feature subset; repeat the previous process until all features are exhausted; finally, get m feature subsets with different numbers of features, and select the feature subset with the highest classification accuracy as the final feature subset, the best sample data set Data_best.

[0075] In S2.2.1, the correlation is measured by the Pearson coefficient, and the calculation formula is shown in formula (8):

[0076]

[0077] In formula (8), xi and yi represent the i-th sample data values ​​of the two features whose correlation is to be evaluated; b is the total number of samples, and its threshold is 0.75. When r≤0.75, it indicates that there is no information redundancy between the two features and they can be retained; when r≥0.75, it indicates that there is information redundancy between the two features and the feature with a lower correlation with the soil compaction degree K should be screened out.

[0078] Step 3: Use regression analysis to establish an explicit relationship between the most effective input features and compaction. The specific process is as follows:

[0079] S3.1, the best sample data set obtained in step 2 is divided into a training set and a test set according to a certain ratio; the training set is used to establish the display expression, and the test set is used to test the performance of the expression;

[0080] S3.2. Based on the most relevant features extracted in S2.2, the genetic algorithm and character regression algorithm are used on the training set to establish the linear and nonlinear relationship between the most effective input features and the compaction degree. The most relevant features will be exhaustively input in the form of free combinations; then the test set is used to test the performance of all prediction expressions, and finally the expression with the best performance is determined.

[0081] The above S3.2 specifically means:

[0082] S3.2.1. Sort the final feature subsets obtained in step 2 according to feature importance, and then filter out the features with the lowest importance one by one to form multiple subsets. Each time, select one subset to display the candidate features of expression fitting; the display expression is divided into linear and nonlinear expressions, and the form of multivariate linear function is y=a1x1+a2x2+…+a u x u ; Wherein ai and xi are the coefficients to be determined and the corresponding features respectively. The genetic algorithm is used to find the optimal coefficients. The initial population (i.e., the initial coefficients) is first generated, and then the initial population is cross-mutated according to the calculation results of the fitness function to form a new population. Then the fitness function is calculated again, and this cycle is repeated until the termination condition is met to determine the final coefficients. The symbolic regression method based on genetic programming is used to fit the nonlinear function. This method starts with a random combination of operation symbols, input features and constants to generate the initial population. The initial population is represented as a symbolic model of a tree structure. The subsequent process is similar to the genetic algorithm, and it is repeated until the termination condition is met, and the optimal tree structure (i.e., the optimal expression) is output.

[0083] S3.2.2. Use the samples of the training set in step S3.1 to fit the display expression, and then use the test set to test the performance of each expression. The expression with the best performance is used as the final display prediction expression, and the determination coefficient (R 2 ) calculates the performance of the expression in S3.2.2, and the calculation formula is shown in formula (9):

[0084]

[0085] In formula (9), N ei and N pi Respectively represent the actual compaction degree and predicted compaction degree of the i-th sample; N ea is the average value of the actual compaction degree.

[0086] Step 4: Using the compaction degree K as a label, the predicted compaction degree of the display relationship and the most relevant feature obtained in step 2 are input into the knowledge-enhanced neural network for training to obtain a compaction degree prediction model;

[0087] S4.1. First, extract and convert the data, and use the maximum and minimum feature scaling technology to scale the input features to between 0 and 1 to help establish the knowledge-enhanced neural network;

[0088] S4.2. Set the hyperparameter optimization interval of the knowledge-enhanced neural network model. The parameters will be adjusted through hyperparameter optimization to determine the optimal hyperparameters that maximize the performance of the model. Define the loss function to measure the gap between the predicted results of the model and the actual results. The hyperparameter optimization and model training are performed using the training set; the expression used by the loss function is: Where MSE stands for mean square error, which is used to measure the error between the model prediction value and the true value, n is the number of samples, y pred is the predicted value, y true is the true value;

[0089] Then, the performance of the model is evaluated on the validation set, the performance of the model on unseen data is checked, the model parameters are adjusted based on the performance of the validation set, the performance of the final model is verified using the test set, the accuracy and generalization ability of the model are evaluated, and finally the establishment of the knowledge-enhanced neural network is completed;

[0090] The calculation formula for the maximum and minimum feature scaling in step S4.1 above is shown in formula (9):

[0091]

[0092] In formula (9), x norm and xori are the values ​​of the feature before and after scaling, respectively; μ and σ are the mean and standard deviation of the feature, respectively.

[0093] The above S4.2 specifically means:

[0094] like Figure 2As shown in Figure 1, the knowledge-enhanced neural network is a model that embeds domain knowledge into an artificial neural network. It can combine the existing prediction formula and the neural network to combine the advantages of the two methods. The knowledge-enhanced neural network contains a back propagation neural network (BPNN). The output of the knowledge-enhanced neural network can be expressed as shown in formula (10):

[0095] y pred =y e +α1f L (x,y e )+α2f n (x,y e ) (10)

[0096] In formula (10), ye is the predicted value of the display expression; x is the input parameter; f L (-) and f n (-) are input features x and y e Linear and nonlinear functions of . Specifically, f n is a BPNN, and f L is a single neuron without an activation function. α1 and α2 are two trainable parameters that are initialized to 0.0 at the beginning of training.

[0097] The establishment of a knowledge-enhanced neural network requires training and learning, and the actual measured compaction degree is used as the label of the training sample. The optimal feature subset determined in step 2 is used as the input variable. The established neural network model uses Xavier to initialize the weights and biases, and the loss function is determined by forward propagation of the input, and the mean square error is used. Measures the difference between the model output and the actual compaction. According to the loss function, the back propagation algorithm is used to calculate the gradient of each parameter to the loss, and then the optimization algorithm is used to update the network parameters to reduce the loss.

[0098] During the training process, grid search and ten-fold cross validation methods were used to determine the best hyperparameters of the knowledge-enhanced neural network: first, the training set was evenly divided into ten subsets, nine of which were used to train the model, and the remaining one was used to test the performance of the model, which was measured by the coefficient of determination; then, based on the optimization interval of the hyperparameters, the grid search method was used to find the hyperparameters that gave the neural network model the highest coefficient of determination on the test subset. Finally, the entire training set was used to establish the best compaction prediction model, and the trained model was tested using the test set to evaluate its performance on unseen data.

[0099] Among them, the hyperparameters to be optimized for the knowledge-enhanced neural network in S4.2 are the number of hidden layers, the number of neurons in each hidden layer, and the learning rate.

[0100] The Adam optimizer is used as the optimization algorithm for the knowledge-enhanced neural network in S4.2.

[0101] Step 5: Input the best input features obtained by real-time measurement of the instrument during the current construction process of the roller into the compaction prediction model, predict the real-time compaction of the road, and draw a cloud map of the on-site soil compaction quality of the entire working area of ​​the roadbed according to the location information;

[0102] Further, the step 5 comprises:

[0103] S5.1, collect compaction vibration signals through sensors installed on the roller, calculate the best feature subset obtained in step 2 according to step 1, and input the calculation results into the knowledge enhanced neural network to predict the compaction degree in real time;

[0104] S5.2. After each compaction, the location information is obtained through the positioning system, and the location information and the real-time compaction degree are combined to draw a compaction quality cloud map;

[0105] like Figure 3 As shown in the figure, the rolling section is divided according to the set height (H) and width (W), and each rolling area of ​​H*B is taken as a specific pixel. The measurement parameters (GPS position information, acceleration amplitude signal, compaction parameters) are collected at the signal collection point, and the parameters except GPS information are converted into the optimal features determined in step 2.2, and input into the established optimal compaction prediction model to predict the on-site compaction. Then, the compaction quality cloud map is drawn based on the GPS position information combined with the compaction prediction results of the on-site collection points.

[0106] Among them, the height of S5.2 is determined by the width of the vibrating wheel of the roller used, and the width can be determined according to the conditions of the on-site construction area or the requirements of visualization.

[0107] This embodiment also provides a multivariable roadbed compaction intelligent detection device, which is used to execute the steps of the multivariable roadbed compaction intelligent detection method based on knowledge enhanced neural network provided in the above embodiment.

[0108] Compared with the prior art, the above technical solution conceived by this embodiment can achieve the following

[0109] Beneficial effects:

[0110] (1) This technical solution aims to address the shortcomings of existing intelligent roadbed compaction technology and proposes an intelligent roadbed compaction quality detection method based on knowledge-enhanced neural network technology. It uses compaction parameters and continuous vibration signals collected during the compaction operation of the roller to predict the roadbed compaction quality in real time, which can significantly reduce the cost of roadbed compaction detection and improve detection efficiency.

[0111] (2) This technical solution comprehensively considers the influence of compaction parameter X1 and intelligent compaction measurement value parameter X2 on compaction degree, and establishes a multivariable roadbed compaction quality evaluation model, which effectively solves the problems of single evaluation index and lack of consideration of soil characteristics in the existing roadbed intelligent compaction evaluation method, thereby realizing accurate prediction of intelligent compaction quality of roadbed under multiple working conditions.

[0112] (3) This technical solution calculates the distance between samples through the K-NearestNeighbor model and screens the original samples through the threshold method. During the screening, the original samples with large errors caused by test conditions and human factors can be removed, thereby forming a high-quality data set; further, correlation analysis and feature recursive elimination method are used to select the most effective input features, and on this basis, genetic algorithms and symbolic regression algorithms are used to obtain the displayed prediction expression, which can be used to quickly predict the roadbed compaction quality while ensuring a certain accuracy.

[0113] (4) This technical solution can draw a compaction quality cloud map of the roadbed soil in real time, visualize the compaction quality of the roadbed in the compaction work area, solve the problem of the difficulty of evaluating the compaction quality of the entire work area after the traditional compaction degree, and help construction personnel quickly locate areas with over-pressure, under-pressure and leakage.

[0114] In addition, the terms "upper", "lower", "inner", "outer", "front", "rear" are used for descriptive purposes only and should not be understood as indicating or implying relative importance. Unless otherwise specifically stated, the relative steps, numerical expressions and values ​​of the components and steps described in these embodiments do not limit the scope of the present invention.

[0115] Of course, the above description is only a specific embodiment of the present invention and is not intended to limit the scope of implementation of the present invention. All equivalent changes or modifications made according to the structure, characteristics and principles described in the patent application scope of the present invention should be included in the patent application scope of the present invention.

[0116] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A multivariate intelligent detection method for roadbed compaction based on knowledge-enhanced neural network, characterized in that: The steps include: Acquire an original sample data set, wherein the sample data set includes compaction parameters and intelligent compaction measurement parameters; The original sample data set is screened to output the best sample data set, and the sample data in the best sample data set is labeled based on the compaction degree, and the best sample data set and the label are combined into a final data set; The corresponding prediction network is constructed based on the knowledge-enhanced neural network framework, which includes a feature extraction module, a mapping module and a prediction module: The feature extraction module is used to extract data features from input sample data; The mapping module converts the extracted data features through a preset display expression to obtain a corresponding mapping result; The prediction module performs analysis based on the input mapping result to output a prediction result; The prediction network is trained using the final data set to obtain a compaction prediction model; The sample data collected on the roller is input into the compaction prediction model to predict the corresponding compaction, and the predicted compaction is combined with the position information of the roller to output the corresponding compaction quality cloud map.

2. The multivariable roadbed compaction intelligent detection method based on knowledge enhanced neural network according to claim 1 is characterized in that: The compaction parameters include roller vibration amplitude, vibration frequency, rolling speed, number of rolling passes, loose paving thickness, compacted soil particle grading, soil initial moisture content or soil compaction degree.

3. The multivariable roadbed compaction intelligent detection method based on knowledge enhanced neural network according to claim 1 is characterized in that: The intelligent compaction measurement parameters include one or more of a compaction degree value, a compaction control value, an acceleration intelligent compaction value, a total harmonic distortion, a vibration peak acceleration, a root mean square of roller vibration acceleration, or a roller static driving force.

4. The multivariable roadbed compaction intelligent detection method based on knowledge enhanced neural network according to claim 1 is characterized in that: The process of screening the original sample data set is as follows: The original sample data set is input into the pre-built KNN model to obtain the prediction results, and the prediction results are compared with the compaction degree to remove the outliers that do not meet the threshold to obtain a valid sample data set; And the effective sample data set is screened by using correlation analysis and feature recursive elimination method to obtain the best sample data set.

5. The multivariable roadbed compaction intelligent detection method based on knowledge enhanced neural network according to claim 1 is characterized in that: The display expression includes linear and nonlinear relationships between input data features and compaction degree, and the linear and nonlinear relationships are constructed by genetic algorithm and character regression algorithm.

6. The multivariable roadbed compaction intelligent detection method based on knowledge enhanced neural network according to claim 1 is characterized in that: During the training process, the data in the final data set is feature extracted and transformed, and the maximum and minimum feature scaling technology is used to scale the extracted features to between 0 and 1 to help establish a knowledge-enhanced neural network.

7. The multivariable roadbed compaction intelligent detection method based on knowledge enhanced neural network according to claim 1 is characterized in that: During the training process, the loss function is used to update the parameters of the prediction network. The expression of the loss function is as follows: Where MSE stands for mean square error, which is used to measure the error between the model prediction value and the true value, n is the number of samples, y pred is the predicted value, y true is the true value.

8. A multivariable roadbed compaction intelligent detection device, characterized in that: Used to execute the steps of the multivariate roadbed compaction intelligent detection method based on knowledge enhanced neural network as described in any one of claims 1 to 7.

Images