System and method for predicting optimal excitation frequency of vibratory roller based on machine learning

Abstract

The invention discloses a system and a method for predicting the optimal excitation frequency of a vibratory roller based on machine learning, and the method comprises the following steps: S1, collecting the original data of soil under different conditions and the working parameters of the vibratory roller under the conditions, investigating the parameters of different models of vibratory rollers, and processing and storing the data; s2, a machine learning algorithm in MATLAB is selected and trained, a machine learning-based optimal excitation frequency prediction model of the vibratory roller is obtained, and thus the predicted soil inherent frequency can be obtained; and S3, comparing the difference between the predicted soil inherent frequency and the soil inherent frequency measured in actual work, and adjusting and optimizing the prediction model in S2. According to the method, the optimal excitation frequency of the vibratory roller is predicted before construction, the prediction precision is improved by integrating factors such as the environment, the pavement type, the soil density and the compaction degree, and the operation requirements of the vibratory roller under different regions and different pavement conditions can be met.

Description

Technical Field

[0001] The present invention belongs to the field of vibratory compaction technology and equipment, and specifically relates to a system and method for predicting the optimal excitation frequency of a vibratory roller based on machine learning. Background Art

[0002] With the continuous development of global transportation construction and the accelerated advancement of urbanization, people's travel and quality of life have also been greatly improved. With the increasing number of cars, higher requirements are placed on the quality of road surfaces. Good road quality is directly related to the safety and efficiency of transportation and the quality of life of urban residents. High-quality roads can provide better driving surfaces and more reliable vehicle control, thereby reducing the incidence of traffic accidents. Smooth, crack-free, and pothole-free roads can reduce vehicle loss of control and braking distances, and improve driving safety. For the economy, road quality helps to extend the service life of roads. Reducing vehicle wear and maintenance costs, reducing vehicle damage and repair costs caused by road damage, can also improve traffic efficiency, reduce driving time and fuel consumption, and thus save economic costs. The quality of roads is determined by pavement materials and compaction quality. With the in-depth study of vibration compaction theory, vibratory rollers have quickly occupied the world compaction machinery market. Today, vibration compaction has become the main method of roadbed construction.

[0003] Vibratory rollers consist of one or more heavy metal rollers with a vibration mechanism inside. They can compact loose soil, gravel, asphalt concrete and other materials through vibration force, thereby improving the density and bearing capacity of the road surface. The selection of vibration frequency is particularly important for vibratory rollers. Based on the resonance principle, when the system is subjected to forced vibration with the same or similar natural frequency as itself, the system amplitude reaches the maximum, thus producing "resonance", and the compaction effect is best at this time. Therefore, in the vibration compaction and rolling operation, it is very important to select a suitable vibration frequency to improve the compaction efficiency. Too low a vibration frequency may not be able to effectively compact the particles of the material, resulting in insufficient density of the road surface, affecting the bearing capacity and durability of the road surface. Low-frequency vibration may not achieve the compaction effect of the deep layer of the road surface, especially for thicker roadbed or road surface layers, which may cause the surface of the material to be compacted, while the deep layer is not fully compacted, affecting the overall road surface quality. Too high a vibration frequency may cause excessive vibration between the material particles, making them unable to be fully and closely arranged, thus affecting the compaction effect. High-frequency vibrations may cause material dispersion, especially for fine-grained materials (such as asphalt concrete), resulting in uneven or wavy pavement surfaces. Therefore, the selection of the optimal frequency for vibration compaction is a very urgent problem to be solved.

[0004] There are few related technologies available. At present, many studies are mainly aimed at real-time detection of pavement compaction quality, and no method for determining the optimal excitation frequency before construction has been proposed. Among them, patent CN117822667A is a real-time monitoring system and method for rolling quality based on image processing, which can accurately determine the running trajectory of the compaction equipment, thereby making the monitoring more accurate. It can also more intuitively display the ground rolling situation through images, making it easier to find areas that do not meet the quality standards, and also convenient for storage and later restoration of the entire rolling process. There are still many practical problems in roadbed construction. Problem 1: The selection of the frequency of vibrating rollers for different roads in different regions is difficult to control. Problem 2: During the compaction process, the density of the road surface will continue to increase with different compaction times, and its inherent natural frequency will also change accordingly. At this time, the natural frequency of the first pass cannot reach the optimal frequency. Problem 3: There is a lack of big data to process these data to pave the way for more intelligent and efficient road construction in the future. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a system and method for predicting the optimal excitation frequency of a vibratory roller based on machine learning. Compared with traditional methods, the present invention has the characteristics of data-driven, automated and real-time optimization, can adapt to the operating requirements of vibratory rollers in different regions and under different road conditions, and has broad application prospects and economic benefits.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for predicting the optimal excitation frequency of a vibratory roller based on machine learning, comprising the following steps:

[0007] Step S1, collecting raw data of soil under different conditions and working parameters of vibratory rollers under these conditions, investigating parameters of vibratory rollers of different models, processing and storing the data, and providing input data for prediction model training in step S2;

[0008] Step S2, selecting and training a machine learning algorithm in MATLAB to obtain a prediction model for the optimal excitation frequency of a vibratory roller based on machine learning, thereby obtaining a predicted soil natural frequency;

[0009] Step S3: compare the difference between the predicted soil natural frequency and the soil natural frequency measured in actual work, and adjust and optimize the prediction model in step S2.

[0010] A further improvement of the technical solution of the present invention is that the specific steps of step S1 are as follows:

[0011] Step S11, in a laboratory or field operation environment, systematically collect original data of soil under different conditions and corresponding working parameters of the vibratory roller under these conditions, and investigate rollers with different functions, different working parameters, and different usage scenarios, so as to match different models of vibratory rollers in advance for subsequent different parameters or scenarios. The collected original data include the original soil natural frequency of the vibratory roller under different regions, climatic conditions, and road surface types, and related parameters that affect the original soil natural frequency. The influencing factors include road surface type, temperature, water content, environmental conditions, soil density, and compaction degree under different compaction times;

[0012] Step S12: uniformly format the collected raw data to ensure that each record contains complete data information and the quantity of each type of data is equal;

[0013] Step S13: Check each record and fill or delete missing data items to ensure data integrity;

[0014] Step S14: Store the sorted and cleaned data set in a structured database to provide input data for subsequent model training.

[0015] A further improvement of the technical solution of the present invention is that the specific steps of step S2 are as follows:

[0016] Step S21, feature engineering: extracting and selecting features from the original data to extract parameters that affect the original soil natural frequency;

[0017] Step S22, model selection and training: according to the data characteristics and prediction task requirements, select a suitable machine learning algorithm to establish a nonlinear relationship between each parameter and the original soil natural frequency, and realize the prediction of soil natural frequency for different road surface types, different environmental conditions, and different soil density combinations;

[0018] Step S23, model evaluation: divide the data in step S22 into a training set and a test set, and evaluate the model accuracy through the test set;

[0019] Step S24, model application: In actual work, the field data is input into the trained model, and the predicted soil natural frequency is selected by the vibratory roller as the excitation frequency of the vibratory roller. At the same time, the actual soil natural frequency must be measured through the method of the data acquisition stage in the work to provide data for the optimization feedback of step S3.

[0020] A further improvement of the technical solution of the present invention is that in step S21, parameters affecting the original soil natural frequency are extracted from the original data, and for numerical features, standardization or normalization methods are used to scale them so that each feature has the same scale in the model, thereby enhancing the convergence of the model, and features with a greater impact on the vibration frequency are screened out through correlation analysis or principal component analysis methods, thereby reducing data dimensions and reducing computational complexity.

[0021] A further improvement of the technical solution of the present invention is that the prediction task in step S22 is a continuous variable, and the continuous variable is a numerical data that changes continuously with the change of the excitation frequency and the excitation force. The selectable machine learning algorithms include linear regression, decision tree regression, random forest regression, support vector machine regression or neural network.

[0022] A further improvement of the technical solution of the present invention is that in the step S23, for the prediction task of continuous variables, the evaluation indicators include mean square error, root mean square error, mean absolute error, and determination coefficient R2 value; the mean square error is a measure of the degree of deviation of the predicted value by calculating the mean of the sum of squares of the errors between the predicted value and the true value; the root mean square error is the square root of the MSE, providing a standardized measure of the prediction error; the mean absolute error is the mean of the absolute errors between the predicted value and the true value, which is convenient for understanding the error size; the determination coefficient R2 value measures the explanatory power of the model for the data, and the closer the determination coefficient R2 value is to 1, the better the prediction effect of the model.

[0023] A further improvement of the technical solution of the present invention is that: in step S3, when the predicted soil natural frequency and the excitation frequency of the vibratory roller are the same and reach resonance, the predicted soil natural frequency at this time is the optimal excitation frequency of the vibratory roller based on machine learning; when the error between the predicted soil natural frequency and the excitation frequency of the vibratory roller exceeds 5%, the data is added to the training set, step S2 is repeated, and the soil natural frequency measured in actual work is reselected, and the model is retrained, so as to iterate and improve continuously.

[0024] A further improvement of the technical solution of the present invention is that the optimal excitation frequency prediction system for a vibratory roller based on machine learning includes a data measurement module, a database establishment module, and an optimization feedback module; the data measurement module is used to collect the original data of the soil under different conditions and the working parameters of the vibratory roller under these conditions, investigate the parameters of vibratory rollers of different models, and process and store the data; the database establishment module is used to select and train the machine learning algorithm in MATLAB to obtain the optimal excitation frequency prediction model for a vibratory roller based on machine learning; the optimization feedback module is used to compare the difference between the predicted soil natural frequency and the soil natural frequency measured in actual work, and adjust and optimize the optimal excitation frequency prediction model for a vibratory roller based on machine learning.

[0025] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is:

[0026] 1. The present invention provides a method for predicting the optimal excitation frequency of a vibratory roller based on machine learning, which can predict the optimal excitation frequency of a vibratory roller before construction, reduce the cost of trial and error, and improve the prediction accuracy by integrating factors such as the environment and road surface type, soil density and compaction degree, providing technical support for intelligent and efficient road construction. Compared with traditional methods, the present invention has the characteristics of data-driven, automated and real-time optimization, can adapt to the operation requirements of vibratory rollers in different regions and under different road conditions, and has broad application prospects and economic benefits;

[0027] 2. The machine learning model of the present invention can predict the optimal construction vibration frequency of the current road surface before construction, so that the entire construction process can be optimized according to this frequency. This allows the construction team to better formulate a construction plan before actual work to avoid potential problems; it can also integrate multiple data sources (such as road surface type, environmental conditions, soil density and compaction, etc.), rather than relying solely on image data. This makes the prediction results more comprehensive and takes into account more factors that may affect the compaction effect. By predicting the optimal frequency of the vibrator, the machine learning model helps reduce the number and time of trial and error, which directly affects the efficiency and speed of construction. The machine learning model can be trained and adjusted for different working conditions and conditions, and is suitable for a variety of different engineering environments and needs. With the accumulation of more data, the machine learning model can continue to learn and improve, improving its prediction accuracy and adaptability. Therefore, in comparison, the real-time monitoring system based on image processing has many shortcomings, and the machine learning model prediction can just make up for these shortcomings;

[0028] 3. The present invention detects the changes in soil density and the corresponding soil natural frequency during the change of compaction times, optimizes the predicted frequency in real time, and dynamically adjusts the model output, thereby realizing adaptive selection of the optimal excitation frequency in each compaction, so as to improve the compaction effect and construction efficiency;

[0029] 4. The present invention introduces machine learning technology to establish a prediction system for selecting the optimal excitation frequency of a vibratory roller based on experimental data. Through the synergy of the data measurement module, the database establishment module, and the optimization feedback module, the system can accurately predict the optimal vibration frequency, thereby improving the compaction quality of the roadbed. The data measurement module is responsible for collecting the original data of the soil under different conditions and the working parameters of the vibratory roller under these conditions, investigating the parameters of vibratory rollers of different models, processing and storing the data, ensuring the accuracy and quality of the data, and providing a high-quality data foundation for subsequent machine learning modeling. The database establishment module automatically extracts features and trains models from a large amount of data without manual processing and analysis of data. In this way, large-scale data can be effectively processed and the accuracy and efficiency of the prediction model can be improved. The optimization feedback module evaluates the performance of the model and provides directions for improvement by continuously collecting and analyzing actual operation feedback information. This enables the prediction model to remain efficient and accurate in practical applications and adapt to changing working conditions and requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flow chart of the method for predicting the optimal excitation frequency of a vibratory roller based on machine learning of the present invention;

[0031] Figure 2 It is a flow chart of establishing a database of the present invention;

[0032] Figure 3 It is the optimization feedback flow chart of the present invention;

[0033] Figure 4 It is a schematic diagram of the effective area of ​​the database of the present invention. DETAILED DESCRIPTION

[0034] The present invention is further described in detail below in conjunction with embodiments:

[0035] like Figure 1 As shown, the present invention provides a method for predicting the optimal excitation frequency of a vibratory roller based on machine learning, comprising the following steps:

[0036] Step S1, collecting raw data of soil under different conditions and working parameters of vibratory rollers under these conditions, investigating parameters of vibratory rollers of different models, processing and storing the data, and providing input data for prediction model training in step S2;

[0037] The specific steps are as follows:

[0038] Step S11: In a laboratory or field operation environment, the original data of the soil under different conditions and the corresponding working parameters of the vibratory roller under these conditions are systematically collected, and rollers with different functions, different working parameters, and different usage scenarios are investigated, so that different models of vibratory rollers can be matched in advance with different parameters or scenarios in the future. The collected original data include the natural frequency of the soil of the vibratory roller under different regions, climatic conditions, and pavement types (such as sand, clay, gravel, asphalt, etc.), and related parameters that affect the natural frequency of the soil. The influencing factors include pavement type, temperature, moisture content, environmental conditions (temperature, humidity, moisture content, etc.), soil density, and compaction degree under different compaction times.

[0039] This step is to investigate and count the various working parameters, different functions, and different use conditions of the vibratory roller, and to measure the natural frequency of the soil at different locations, different road surfaces, and different compaction times. There are several ways to test the natural frequency of the soil:

[0040] (1) Laser Doppler vibrometer: This method uses laser technology to measure the vibration velocity and displacement of the roadbed in a non-contact manner, and analyzes the vibration frequency from these data. This method has high accuracy and sensitivity and is suitable for fragile or difficult-to-reach surfaces. However, the equipment is relatively expensive and is easily affected by external light and weather conditions, which affects the measurement accuracy.

[0041] (2) Rebound method: Use an elastic hammer or a heavy hammer to hit the roadbed surface and collect vibration data through a vibration sensor. The equipment is relatively simple, low cost, and easy to operate. However, the data may be greatly affected by the operating skills and have poor repeatability. It is affected by surface conditions, such as slippery or uneven surfaces.

[0042] (3) Acoustic wave testing: The vibration frequency is estimated by emitting sound waves and measuring their propagation time in the roadbed material. It can be operated remotely, is suitable for harsh environments, and can conduct deep underground testing. However, it requires specific environmental conditions, such as the propagation speed of sound waves in different media. Complex mathematical processing of the results may be required.

[0043] (4) Measure ground parameters in real time during operation by installing sensors on the vibrating wheel: With the development of intelligent compaction technology, its basic principle is to use the vibrating roller as a detection device, install sensors on the vibrating wheel, obtain the real-time status of the soil, and then judge the degree of compaction of the soil. These are four methods for testing the natural frequency of the roadbed. Appropriate methods and equipment can be selected for testing based on their advantages and disadvantages and the properties of the road surface. However, the above methods cannot directly measure the value of the natural frequency of the soil, but can only obtain signals such as displacement, velocity, and acceleration. Therefore, it is necessary to process the above signals through Fourier transform to obtain the natural frequency. The following are the main steps for calculating the test signal using Fourier transform:

[0044] (1) Data collection

[0045] The collected signals are x(t), v(t), a(t), and the sampling frequency is f s , the sampling time is t=0,Δt,2Δt……,(N-1)Δt, where

[0046] (2) Data preprocessing, which means removing the DC offset of displacement, velocity and acceleration signals.

[0047]

[0048] in: are the mean values ​​of the signals respectively;

[0049] (3) Frequency domain analysis: Fourier transform is performed on the displacement, velocity and acceleration signals respectively to convert the time domain signals into frequency domain signals.

[0050] Fourier transform formula:

[0051]

[0052] For discrete signals, the Fast Fourier Transform (FFT) is usually used to calculate its Discrete Fourier Transform (DFT):

[0053]

[0054] Similarly, for the velocity and acceleration signals, their FFTs are calculated separately:

[0055]

[0056] (4) Spectrum amplitude calculation

[0057]

[0058] (5) Frequency calculation

[0059] The actual value corresponding to the frequency:

[0060]

[0061] (6) Extracting the main frequency components

[0062] Find the frequency corresponding to the maximum amplitude in the spectrum:

[0063]

[0064] After the results are obtained, the data can be initially screened and processed to select valid data to lay a foundation for the establishment of the database. The data normalization formula can be used in data processing:

[0065]

[0066] Where: x′ is the normalized data; x is the original data; min(x)max(x) is the minimum and maximum values ​​of the data. There are many advantages to using the normalization formula. The normalized data is usually in a smaller range (such as 0 to 1 or -1 to 1), which helps to speed up the convergence of the gradient descent algorithm. Unnormalized data may cause the gradient descent to have too large or too small steps in some directions, thereby slowing down the training process. It can help the algorithm find the global optimal solution more easily and avoid falling into the local optimum. This is because the normalized data is more evenly distributed, making the shape of the loss function more regular. Moreover, when performing matrix operations, the mixture of large and small values ​​may cause numerical instability or overflow problems. Normalization can scale the data to a similar range and reduce the instability in numerical calculations. Data normalization is a key step in machine learning data processing. By scaling the data to a similar range, the training efficiency, stability and accuracy of the model can be significantly improved, ensuring that the influence of each feature on the model is balanced, thereby obtaining better prediction results.

[0067] Step S12: uniformly format the collected raw data to ensure that each record contains complete data information and the quantity of each type of data is equal;

[0068] Step S13: Check each record and fill or delete missing data items to ensure data integrity;

[0069] Step S14: Store the sorted and cleaned data set in a structured database to provide input data for subsequent model training.

[0070] Step S2, select and train the machine learning algorithm in MATLAB to obtain the optimal excitation frequency prediction model of the vibratory roller based on machine learning, thereby obtaining the predicted soil natural frequency. This step is to collect data, including the material, thickness, weather conditions, traffic flow, etc. of different road surfaces, and store these data in the database, which can be used as the basic data for training and testing models.

[0071] The specific steps are as follows:

[0072] Step S21, feature engineering: extract and select features for the original data, and extract the parameters that affect the original soil natural frequency; for numerical features, use standardization or normalization methods to scale them so that each feature has the same scale in the model to enhance the convergence of the model; use correlation analysis or principal component analysis to screen out features that have a greater impact on the vibration frequency, reduce data dimensions, and reduce computational complexity.

[0073] Before establishing a prediction model, it is necessary to determine which factors will affect the prediction results, and then convert or create new features through feature engineering. This step can increase the prediction accuracy of the model. Some conditions will also affect the natural frequency of the road surface after interaction. For example, different combinations of road surface materials and climate environment, temperature and humidity will produce new influencing factors. At the same time, when inputting data, the parts with excessive errors or missing data fragments should be filtered out to improve the accuracy of the model. This process can use the correlation coefficient method for feature selection. The correlation coefficient method selects the most useful features for prediction by calculating the correlation between each feature and the target variable. Calculate the Pearson Correlation Coefficient between each feature and the target variable. The calculation formula of the Pearson correlation coefficient is as follows:

[0074]

[0075] Where: r xy is the relevant data of features x and y; i ,y i are the values ​​of features x and y; is the mean of features x and y. According to the calculated correlation coefficient value, select the features with higher correlation with the target variable. Usually a correlation threshold is set, for example, features with an absolute value of correlation coefficient greater than 0.5 are selected. Feature selection can significantly improve the performance and reliability of the vibration compaction prediction model, ensuring that the model has good performance in practical applications.

[0076] Step S22, model selection and training: According to data characteristics and prediction task requirements, select a suitable machine learning algorithm to establish a nonlinear relationship between each parameter and the original soil natural frequency, and realize the prediction of soil natural frequency for different road types, different environmental conditions, and different soil density combinations; if the prediction task is a continuous variable, the selectable machine learning algorithms include linear regression, decision tree regression, random forest regression, support vector machine regression, or neural network. These algorithms are mature models in MATLAB and can be directly imported into data for use. These models each have their own advantages and disadvantages and are suitable for different conditions and can achieve different functions. The appropriate model can be selected according to the complexity of the data, the prediction accuracy requirements, and the linear relationship of the data.

[0077] (1) Linear regression model

[0078] Model description: Linear regression is used to establish a linear relationship between input features and output continuous variables. It is suitable for situations where the data is relatively linear and the noise is small. If the relationship between the expected natural frequency and the road surface characteristics is relatively linear, the linear regression model is a simple and easy-to-interpret choice. The model is simple, computationally efficient, and easy to understand and implement.

[0079] Training method: Define loss function: Use mean square error (MSE) as the loss function to quantify the difference between the predicted value and the actual value.

[0080] Minimize loss: Adjust the model parameters (i.e., regression coefficients) through the gradient descent method to minimize the loss function and thus optimize the prediction accuracy.

[0081] (2) Decision Tree Regression Model

[0082] Model introduction: Decision tree regression predicts the target variable by continuously segmenting the feature space, and has a good effect on modeling nonlinear relationships. When the data structure is complex, the relationship between features is nonlinear, or contains many categorical variables, decision trees are intuitive, easy to understand and implement, and can naturally handle nonlinear relationships.

[0083] Training method: Build a tree structure and build a decision tree based on the training data. Select the split features and split points so that the mean square error (MSE) of the child nodes after the split is the lowest.

[0084] Pruning: To prevent the model from overfitting, pruning techniques are used (such as setting the maximum depth of the tree, the minimum number of sample splits, etc.) to limit the depth or complexity of the tree.

[0085] (3) Random Forest Regression Model

[0086] Model introduction: Random forest is an ensemble learning method consisting of multiple decision trees. It integrates the prediction results of multiple trees through a "voting" mechanism to improve the generalization ability of the model. Random forest is an ensemble learning method that improves the stability and accuracy of a single decision tree, reduces the risk of overfitting, and enhances the generalization ability of the model.

[0087] Training method: Construct multiple decision trees: Use different subsamples of the training data (obtained by random sampling) to construct multiple decision trees, and each tree is trained on a different subset of features.

[0088] Result averaging: The prediction results of all trees are averaged as the final output to reduce the high variance problem that may be caused by a single decision tree.

[0089] (4) Support Vector Machine (SVM) Regression Model

[0090] Model introduction: SVM regression predicts continuous variables by maximizing the interval and defining the support vector, and has strong adaptability to high-dimensional feature space. Support vector machines are suitable for small and medium-sized data sets, especially in high-dimensional space. They are very powerful and flexible, especially good at handling nonlinear separation problems, and can efficiently handle nonlinear relationships through kernel techniques.

[0091] Training method: Define the objective function: Use a loss function (such as ε-insensitive loss) to define the allowable error range and optimize the predicted value by maximizing the interval.

[0092] Kernel function selection: According to the linear separability of the data, select the appropriate kernel function (linear kernel, Gaussian kernel, or polynomial kernel) and map it to a high-dimensional space to improve model performance.

[0093] In the preliminary stage, you can start with simpler models (such as linear regression or decision trees) and evaluate their performance. If the simple model does not meet the prediction accuracy requirements, you can consider more complex models such as random forests, support vector machines, or neural networks. For such practical application problems, random forests are often recommended because they can achieve a good balance in many different types of regression problems, that is, while maintaining a high accuracy rate, the model's interpretability and computational complexity are still within an acceptable range.

[0094] Here we take the support vector regression (SVR) model as an example. Support vector regression (SVR) is a regression method based on the principle of support vector machine (SVM) and is used to solve regression problems. Similar to its application in classification tasks, the core idea of ​​support vector regression is to find a hyperplane (in high-dimensional space) that can fit all training data with the smallest error. Its working principle is that in support vector machine, the hyperplane is the decision boundary for classification. In support vector regression, the hyperplane becomes a prediction function that aims to predict continuous target values. SVR introduces a concept called ε-insensitive tube. This means that the prediction error within this band is ignored, that is, if the difference between the predicted value and the true value is less than or equal to ε, the predicted value is considered correct. The goal is to find a function that makes most of the data points fall within this band. The data points outside this band are called support vectors, which are crucial to determining the regression function. Similar to SVM, SVR also often uses kernel functions to deal with nonlinear problems. Through the kernel function, the data is mapped to a high-dimensional space, where it may be easier to find a suitable regression hyperplane.

[0095] Specific steps to use Support Vector Regression (SVR):

[0096] After the data determination module and data processing steps, the data types have been classified and labeled (pavement material, number of compaction passes, pavement moisture, etc.). Next, the SVR model is used to train the data and perform nonlinear mapping through kernel functions (such as linear kernel, RBF kernel).

[0097] Linear kernel: K(xi ,x)=xi·x

[0098] Polynomial kernel: K(x i ,x)=(x i ·x+c) d

[0099] RBF kernel: K(x i ,x)=exp(-γ||x i -x|| 2 )

[0100] Use the data in the training set to train the model. The training process of SVR is as follows:

[0101] Solve the following optimization problem:

[0102]

[0103] Under the constraints:

[0104]

[0105]

[0106] Where: ω: model weight vector b: model bias term C: regularization parameter ε: insensitive zone width y i is the target variable, i.e. the actual measured road surface natural frequency x i is the input feature vector.

[0107] Use the existing support vector machine algorithm library to solve this optimization problem and obtain ω and b.

[0108] Prediction process: When using the trained SVR model for prediction, the following formula is used:

[0109]

[0110] Among them, K(x i ,x): kernel function, calculates the new input data x and the training data x i The similarity between α i and α i * : Lagrange multiplier, from the solution of the optimization problem, b: bias term of the model.

[0111] Step S23, model evaluation: Divide the processed data in step S22 into a training set and a test set, usually in a ratio of 7:3 or 8:2, to ensure the effectiveness of the model when tested on unseen data. The training set is used to train the model, while the test set is used to evaluate the prediction effect of the model to ensure the generalization ability of the model. For the prediction task of continuous variables, the evaluation indicators include mean square error, root mean square error, mean absolute error, and determination coefficient R 2 The mean square error is the average of the sum of the squares of the errors between the predicted value and the true value, which measures the degree of deviation of the predicted value. The root mean square error is the square root of the MSE, which provides a standardized measure of the prediction error. The mean absolute error is the average of the absolute errors between the predicted value and the true value, which is convenient for understanding the error size. The coefficient of determination R 2 The value measures the explanatory power of the model on the data, the coefficient of determination R 2 The closer the value is to 1, the better the prediction effect of the model. This step is also implemented in MATLAB by programming. During the implementation process, multiple models can be selected, and the final model is determined by comparing the prediction accuracy.

[0112] Step S24, model application: In actual work, the field data (different materials, different regions (temperatures), different moisture contents, different compaction times) are input into the trained model, and the predicted soil natural frequency is selected by the vibratory roller as the excitation frequency of the vibratory roller. At the same time, the actual soil natural frequency must be measured through the method of the data acquisition stage in the work to provide data for subsequent optimization feedback.

[0113] Step S3: Figure 3 As shown, the difference between the predicted soil natural frequency and the soil natural frequency measured in actual work is compared, and the prediction model in step S2 is adjusted and optimized. When the predicted soil natural frequency and the excitation frequency of the vibratory roller are the same, resonance will be achieved. At this time, the predicted soil natural frequency is the optimal excitation frequency of the vibratory roller based on machine learning; when the error between the predicted soil natural frequency and the excitation frequency of the vibratory roller exceeds 5%, the data is added to the training set, step S2 is repeated, and the soil natural frequency measured in actual work is reselected, and the model is retrained, so as to iterate and continuously improve. This step can evaluate the performance of the machine learning model, which helps to determine the accuracy and reliability of the model and provide directions for improvement.

[0114] It is very important to design an efficient feedback optimization step for machine learning prediction models. This step can help the model to continuously correct and improve itself, which is helpful for the accuracy of model prediction. This step is specifically after the model is established through test data. During actual construction, real-time prediction is performed according to the current road conditions. Input data and use the prediction model to predict. Compare the optimal frequency in actual construction with the database. If the gap is too large, update the database. This process requires the construction of a feedback mechanism. When a large difference is detected between the predicted frequency and the actual frequency, the data is re-added to the training, and the model is re-trained to iterate and continuously improve. When comparing data, the mean square error method can be used to judge the size of the error to determine whether the data needs to be re-added to the database training.

[0115] The formula for mean square error is:

[0116]

[0117] Where: f i The actual optimal frequency of the i-th sample; The model predicts the frequency of the i-th sample; m is the number of samples. The smaller the MSE value, the smaller the error between the model's predicted value and the actual value.

[0118] The present invention provides a vibratory roller optimal excitation frequency prediction method based on machine learning, and a vibratory roller optimal excitation frequency prediction system based on machine learning, including a data determination module, a database establishment module, and an optimization feedback module.

[0119] The data measurement module is used to collect raw data from the operation of the vibratory roller, and store the raw data after sorting and cleaning. The data measurement module plays an important role in the prediction model of the optimal excitation frequency selection of the vibratory roller based on machine learning. It is mainly responsible for collecting raw data from the operation of the vibratory roller. This module may involve many methods of collecting the natural frequency of the road surface, and the process must ensure that they operate normally and provide accurate data. It may include sensor calibration, data acquisition frequency setting, etc. Once the data is collected, the data measurement module may pre-process the raw data, including removing noise, filling missing values, filtering the data, etc., to ensure the quality and accuracy of the data. The collected data needs to be stored for subsequent use. The data measurement module can be responsible for storing the processed data in a database or saving it in other forms for subsequent machine learning modeling and optimization processes. The data measurement module must ensure that the collected data is accurate, which is crucial for training machine learning models and predicting the optimal excitation frequency. The data measurement module can collect multiple types of data, including vibration frequency, environmental conditions, etc., to provide more comprehensive information for the machine learning model, making its prediction more accurate and reliable.

[0120] The database establishment module is used to select and train the machine learning algorithm in MATLAB to obtain the optimal excitation frequency prediction model of the vibratory roller based on machine learning and obtain the predicted soil natural frequency. Figure 2 As shown in the figure, the database establishment module based on machine learning mainly collects, organizes and trains data in two major directions: vibratory rollers and pavement. Vibratory rollers can be divided into many types according to different parameters, vibration frequency ranges and different operating conditions. In actual construction, they need to be selected according to different compaction materials, different road conditions and different compaction qualities. The pavement direction is more complicated. Since the natural frequency of soil is affected by many factors, and even by the cross-influence of many factors, more complete data is needed in the pavement direction to count the pavement conditions. Climate environments such as drought and humidity, temperature environments such as high temperature and low temperature, different pavement materials such as asphalt, concrete, soil, and different compaction times will affect their natural frequencies. When using machine learning to establish a database, the system will classify and label these factors that affect the natural frequency, and then train these data, cross-validate the data of different factors, and finally train a more complete database module.

[0121] The database establishment module plays a key role in the prediction model of the optimal excitation frequency selection of the vibratory roller based on machine learning. The main tasks of this module are data collection and organization, feature engineering, model training, and model evaluation. The prediction model can become more and more accurate through the continuous improvement of a large amount of data in this order. The machine learning database establishment module can automatically extract features from a large amount of data and train the model without manual processing and analysis of the data. By establishing a database through machine learning, large-scale data can be effectively processed and hidden patterns and relationships can be mined from it, which improves the accuracy and efficiency of the prediction model for the optimal excitation frequency selection of the vibratory roller. This module can be adjusted and optimized according to actual needs and changes in data, so that the prediction model can adapt to the operation of the vibratory roller under different conditions. By continuously accumulating new data and feedback information, the machine learning database establishment module can continuously improve the prediction model so that it can be continuously optimized and adapted in practical applications.

[0122] The optimization feedback module is used to compare the difference between the predicted soil natural frequency and the soil natural frequency measured in actual work, and to adjust and optimize the optimal excitation frequency prediction model of the vibratory roller based on machine learning. According to the collected feedback information, the optimization feedback module can adjust and optimize the machine learning model. This includes adjusting the parameters of the model, improving the feature engineering process, and even selecting different machine learning algorithms to improve the prediction performance of the model. Based on the feedback information collected in real time, the model parameters can be dynamically adjusted or the model can be updated to adapt to the changing operating environment and requirements, which enables the prediction model to remain efficient and accurate in actual applications. By continuously collecting and analyzing actual operation feedback information, the optimization feedback model can help the model to be continuously improved and optimized to adapt to changing working conditions and requirements. After the database in the database establishment module is generated, according to the provisions of the optimization feedback module, if the difference exceeds five percent, it will be re-entered into the database, and the soil natural frequencies under different compaction passes in Table 1 are input into the prediction model. Figure 4 It can be seen that the database will generate a valid area for each road surface. If the predicted value is within the valid area, it can be proved that the database is accurate. If it exceeds the valid area, it will be re-entered.

[0123] Table 1

[0124]

[0125] The real-time collection and processing of feedback information enables the model to be adjusted and optimized in a timely manner, ensuring that the accuracy and reliability of the prediction are maintained in actual operation. According to the operating characteristics and working environment of different vibratory rollers, the optimization feedback module can adjust the model in a targeted manner to make it more suitable for specific situations and improve the accuracy and practicality of the prediction. It can also adaptively adjust the parameters and structure of the model according to the feedback information collected at different stages to adapt to different working conditions and requirements.

Claims

1. A method for predicting the optimal excitation frequency of a vibratory roller based on machine learning, characterized in that: The steps include: Step S1, collecting raw data of soil under different conditions and working parameters of vibratory rollers under these conditions, investigating parameters of vibratory rollers of different models, processing and storing the data, and providing input data for prediction model training in step S2; Step S2, selecting and training a machine learning algorithm in MATLAB to obtain a prediction model for the optimal excitation frequency of a vibratory roller based on machine learning, thereby obtaining a predicted soil natural frequency; Step S3: compare the difference between the predicted soil natural frequency and the soil natural frequency measured in actual work, and adjust and optimize the prediction model in step S2.

2. The method for predicting the optimal excitation frequency of a vibratory roller based on machine learning according to claim 1, characterized in that: The specific steps of step S1 are as follows: Step S11, in a laboratory or field operation environment, systematically collect original data of soil under different conditions and corresponding working parameters of the vibratory roller under these conditions, and investigate rollers with different functions, different working parameters, and different usage scenarios, so as to match different models of vibratory rollers in advance for subsequent different parameters or scenarios. The collected original data include the original soil natural frequency of the vibratory roller under different regions, climatic conditions, and road surface types, and related parameters that affect the original soil natural frequency. The influencing factors include road surface type, temperature, water content, environmental conditions, soil density, and compaction degree under different compaction times; Step S12: uniformly format the collected raw data to ensure that each record contains complete data information and the quantity of each type of data is equal; Step S13: Check each record and fill or delete missing data items to ensure data integrity; Step S14: Store the sorted and cleaned data set in a structured database to provide input data for subsequent model training.

3. The method for predicting the optimal excitation frequency of a vibratory roller based on machine learning according to claim 2 is characterized in that: The specific steps of step S2 are as follows: Step S21, feature engineering: extracting and selecting features from the original data to extract parameters that affect the original soil natural frequency; Step S22, model selection and training: according to the data characteristics and prediction task requirements, select a suitable machine learning algorithm to establish a nonlinear relationship between each parameter and the original soil natural frequency, and obtain the predicted soil natural frequency for different road surface types, different environmental conditions, and different soil density combinations; Step S23, model evaluation: divide the data in step S22 into a training set and a test set, and evaluate the model accuracy through the test set; Step S24, model application: In actual work, the field data is input into the trained model, and the predicted soil natural frequency is selected by the vibratory roller as the excitation frequency of the vibratory roller. At the same time, the actual soil natural frequency must be measured through the method of the data acquisition stage in the work to provide data for the optimization feedback of step S3.

4. The method for predicting the optimal excitation frequency of a vibratory roller based on machine learning according to claim 3 is characterized in that: In step S21, the parameters that affect the original soil natural frequency are extracted from the original data. For numerical features, standardization or normalization methods are used to scale them so that each feature has the same scale in the model, thereby enhancing the convergence of the model. Correlation analysis or principal component analysis methods are used to screen out features that have a greater impact on the vibration frequency, thereby reducing data dimensions and computational complexity.

5. The method for predicting the optimal excitation frequency of a vibratory roller based on machine learning according to claim 3 is characterized in that: The prediction task in step S22 is a continuous variable, which is numerical data that changes continuously with the changes in the excitation frequency and the excitation force. The selectable machine learning algorithms include linear regression, decision tree regression, random forest regression, support vector machine regression or neural network.

6. The method for predicting the optimal excitation frequency of a vibratory roller based on machine learning according to claim 5 is characterized in that: In step S23, for the prediction task of continuous variables, the evaluation indicators include mean square error, root mean square error, mean absolute error, determination coefficient R 2 The mean square error is the average of the sum of the squares of the errors between the predicted value and the true value, which measures the degree of deviation of the predicted value. The root mean square error is the square root of the MSE, which provides a standardized measure of the prediction error. The mean absolute error is the average of the absolute errors between the predicted value and the true value, which is convenient for understanding the error size; the determination coefficient R 2 The value measures the explanatory power of the model on the data, the coefficient of determination R 2 The closer the value is to 1, the better the model's prediction performance.

7. The method for predicting the optimal excitation frequency of a vibratory roller based on machine learning according to claim 5, characterized in that: In step S3, when the predicted soil natural frequency and the excitation frequency of the vibratory roller are the same and reach resonance, the predicted soil natural frequency at this time is the optimal excitation frequency of the vibratory roller based on machine learning; when the error between the predicted soil natural frequency and the excitation frequency of the vibratory roller exceeds 5%, the data is added to the training set, step S2 is repeated, and the soil natural frequency measured in actual work is reselected, and the model is retrained, so as to iterate and improve continuously.

8. A prediction system for the optimal excitation frequency of a vibratory roller based on machine learning, used in any prediction method of any one of items 1 to 7, characterized in that: It includes a data measurement module, a database establishment module, and an optimization feedback module; the data measurement module is used to collect the original data of the soil under different conditions and the working parameters of the vibratory roller under these conditions, investigate the parameters of vibratory rollers of different models, and process and store the data; the database establishment module is used to select and train the machine learning algorithm in MATLAB, obtain the optimal excitation frequency prediction model of the vibratory roller based on machine learning, and obtain the predicted soil natural frequency; the optimization feedback module is used to compare the difference between the predicted soil natural frequency and the soil natural frequency measured in actual work, and adjust and optimize the optimal excitation frequency prediction model of the vibratory roller based on machine learning.

Images