Waste rock resilience modulus prediction method based on automatic machine learning

Abstract

The invention discloses a waste rock resilience modulus prediction method based on automatic machine learning, and belongs to the crossing field of mineral engineering, geotechnical engineering and machine learning, and the method comprises the following steps: collecting physical parameters and stress state data of waste rocks to construct a multi-source data set; preprocessing the multi-source data set to obtain input features; constructing an automatic machine learning model, and training the automatic machine learning model based on the input features to obtain a trained automatic machine learning model; and inputting the physical parameters and the stress state data of the waste rock into the trained automatic machine learning model to obtain the rebound modulus of the waste rock.

Description

Technical Field

[0001] The present invention belongs to the interdisciplinary technical field of mining engineering, geotechnical engineering and machine learning, and in particular relates to a method for predicting the rebound modulus of waste rock based on automatic machine learning. Background Art

[0002] Global mining activities generate approximately 6.5 billion tons of waste rock annually, with this amount increasing annually. If this waste rock is directly dumped on the surface, it can cause environmental problems such as mudslides, vegetation destruction, dust pollution, water contamination, and land occupation. However, with the rapid development of road and other infrastructure construction, the demand for aggregate and gravel has increased dramatically. Therefore, reusing waste rock for road construction has become an effective alternative, reducing environmental pollution while fully utilizing waste rock resources.

[0003] The rebound modulus of waste rock is a key indicator of its mechanical properties, reflecting its stress-strain behavior under traffic loads. Traditional methods for measuring the rebound modulus require repeated loading triaxial testing. Existing techniques require specialized equipment and are time-consuming and expensive (e.g., large-scale specimens with a diameter of 300 mm). Several empirical models have been proposed to quantify the rebound modulus of waste rock, such as the May and Witczak model and the Uzan model. However, these models primarily rely on factors such as volumetric stress, confining pressure, and differential stress, and fail to fully consider the influence of physical properties such as waste rock particle size distribution and shape factor on its rebound modulus, limiting their applicability and accuracy. Traditional neural network models with manually tuned parameters are subject to the risk of overfitting and have poor generalization capabilities for different particle size distributions.

[0004] There is an urgent need for an automated, high-precision prediction method that can integrate physical parameters and stress states, adapt to the heterogeneous characteristics of coarse-grained waste rock, reduce experimental costs and improve design efficiency. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention proposes a waste rock rebound modulus prediction method based on automatic machine learning to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above objectives, the present invention provides a method for predicting the rebound modulus of waste rock based on automatic machine learning, comprising:

[0007] Collect physical parameters and stress state data of waste rock to construct a multi-source data set;

[0008] Preprocessing the multi-source dataset to obtain input features;

[0009] Constructing an automatic machine learning model, and training the automatic machine learning model based on the input features to obtain a trained automatic machine learning model;

[0010] The physical parameters and stress state data of the waste rock are input into the trained automatic machine learning model to obtain the rebound modulus of the waste rock.

[0011] Optionally, a multi-source data set is obtained by simulating the effect of vehicle load on waste rock in the form of half-sine loading.

[0012] Optionally, performing Pearson correlation analysis on the multi-source dataset to obtain input features;

[0013] The input characteristics include: dry density, non-uniformity coefficient, limited particle size, effective particle size, gradation curve shape factor, confining pressure and deviatoric stress.

[0014] Optionally, the calculation expression of the gradation curve shape factor is:

[0015]

[0016] Where, P i is the weight percentage of waste rock passing the i-th sieve particle size; d i is the opening size of this special sieve; D max is the maximum particle size of waste rock particles; n is the shape factor of the gradient curve.

[0017] Optionally, the hidden layer structure of the automatic machine learning model is dynamically optimized through neural architecture search to determine the optimal number of hidden layers and neurons, and the activation function adopts ReLU.

[0018] Optionally, during training of the automatic machine learning model, an adaptive learning rate strategy is used to optimize the generalization ability of the model.

[0019] Optionally, the automatic machine learning model includes: an input layer, a normalization layer, a first hidden layer, a second hidden layer, and an output layer;

[0020] The normalization layer includes 7 neurons; the first hidden layer includes 32 neurons; the second hidden layer includes 1024 neurons; and the output layer includes one neuron.

[0021] Compared with the prior art, the present invention has the following advantages and technical effects:

[0022] The proposed method for predicting the rebound modulus of waste rock based on automated machine learning comprehensively considers the physical properties and stress state of the waste rock, utilizing automated machine learning technology to construct a high-precision prediction model. This method overcomes the limitations of traditional empirical models, significantly improving prediction accuracy and generalization capabilities, and is adaptable to waste rock with varying particle size distributions and stress conditions. Furthermore, without relying on expensive experimental equipment and complex experimental procedures, this method rapidly outputs prediction results, significantly reducing engineering costs and time. This provides an efficient and reliable scientific basis for engineering applications such as road base design, demonstrating significant engineering application value and widespread adoption. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0024] Figure 1 This is a flowchart of the technical solution of an embodiment of the present invention

[0025] Figure 2 The confining pressure and deviatoric stress distribution of the triaxial cyclic loading and unloading test of an embodiment of the present invention;

[0026] Figure 3 This is a Pearson correlation heat map of basic physical parameters of waste rock materials according to an embodiment of the present invention;

[0027] Figure 4 The AutoML prediction model structure of the embodiment of the present invention;

[0028] Figure 5 The evolution of the loss MAE of the AutoML model of the embodiment of the present invention with the number of iterations;

[0029] Figure 6 Predictive performance of the AutoML model of an embodiment of the present invention on training and test datasets. DETAILED DESCRIPTION

[0030] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0031] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0032] Example 1

[0033] The present invention aims to provide a method for predicting the rebound modulus of waste rock based on automatic machine learning (AutoML) to solve the problems of insufficient consideration of physical properties, low prediction accuracy and high experimental cost in the existing technology.

[0034] like Figure 1 As shown, this embodiment provides a method for predicting the elastic modulus of waste rock based on automated machine learning, comprising the following steps: collecting physical parameters and stress state data of the waste rock to construct a multi-source dataset; preprocessing the multi-source dataset to obtain input features; constructing an automated machine learning model, and training the automated machine learning model based on the input features to obtain a trained automated machine learning model; and inputting the physical parameters and stress state data of the waste rock into the trained automated machine learning model to obtain the elastic modulus of the waste rock. The specific implementation process is as follows:

[0035] (1) Multi-source data collection

[0036] Based on the repeated load triaxial test of large-size (300mm×600mm) and regular-size (150mm×300mm) specimens, the stresses at different stress levels (confining pressure σ3, deviatoric stress σ d ) and physical parameters (dry density ρ d , unevenness coefficient C u , limited particle size D 60 , effective particle size D 10 , gradation curve shape factor n). The maximum particle size of waste rock (D max ) are set to 25mm, 40mm and 60mm respectively to reduce the impact of particle size effect on the test results. Mine roads are an important scene for waste rock utilization. Unlike ordinary roads, mining trucks have slow speeds and high loads. Therefore, based on the characteristics of overweight mining trucks, low-frequency (1Hz) high-amplitude (≥1000kPa) semi-sine loading is used to simulate the effect of vehicle load on waste rock. By testing the rebound modulus of waste rock under various stress paths, the triaxial test confining pressure σ3 and deviatoric stress σ d The ranges of variation are 20~1000kPa and 20~2000kPa respectively, and the deviatoric stress σ d The ratio of the pressure to the confining pressure σ3 is as low as 0.7 and as high as 10, e.g. Figure 2 The detailed information of its physical parameters is shown in Table 1.

[0037] Table 1

[0038]

[0039] Calculation principle of gradation curve shape factor n:

[0040] The Talbot equation (Equation 1) was used to quantitatively characterize the gradation of waste rock, with the gradation shape factor varying between 0.3 and 0.7. The higher the shape factor n, the steeper the slope of the PSD curve, the more uniform the particle size distribution, and the worse the gradation.

[0041]

[0042] Where P i is the weight percentage of waste rock passing the i-th sieve particle size; d i is the opening size of this special sieve; D max is the maximum particle size of waste rock particles; n is the shape factor of the gradient curve.

[0043] (2) Data preprocessing

[0044] Feature selection: Select features that have a significant impact on the rebound modulus.

[0045] In order to reduce the dimensionality of the data, the input feature parameters of the machine learning model are selected, and the dry density ρ d , maximum particle size D max , gradation parameters CU, CC, D 60 、D 30 、D 10 , gradation coefficient n, GC, sand content PC, powder content FC, confining pressure σ3 and deviatoric stress σ d Perform a Pearson correlation analysis. The Pearson coefficient r is a statistical measure used to measure the strength and direction of the linear relationship between two variables x and y (Equation 2). It is widely used in correlation analysis, where r = 1 indicates a perfect positive correlation between the two variables, r = -1 indicates a perfect negative correlation, and r = 0 indicates no correlation at all.

[0046]

[0047] When |r| ≥ 0.8, the two variables are generally considered to be strongly correlated; when 0.2 < |r| < 0.8, there is an acceptable correlation; when |r| ≤ 0.2, the correlation between the two variables is weak. Therefore, when |r| ≥ 0.8, the two variables are considered to be highly correlated, and one of the variables can be removed for dimensionality reduction. Figure 3 The results of Pearson correlation analysis showed that D max and ρ d 、D max With D 60 、D 60 With D 30 、C C with C U 、C C With FC, C UThe absolute values of the Pearson coefficients |r| with FC, n and SC are all greater than 0.8, indicating that there is a strong correlation between these parameter variables and dimensionality reduction can be performed. Considering the relationship between these variables, D max 、C C 、FC、D 30 and SC are deleted and not used as input feature parameters of the machine learning model. Therefore, ρ d 、C U 、D 60 、D 10 , n, σ3, σ d Seven parameters were selected as input features for the automated machine learning model to predict the modulus of resilience of coarse-grained gangue.

[0048] (3) AutoML model construction

[0049] Use automatic machine learning (AutoML) technology to build prediction models, such as Figure 4 As shown, the specific structure is as follows:

[0050] Input layer: 7-dimensional feature vector (ρ d 、C u 、D 60 、D 10 , n, σ3, σ d ).

[0051] Normalization layer: Normalizes the input features. The maximum and minimum normalization method is used to normalize the dataset, limiting the data to [0, 1] (Equation 3), which can eliminate the influence of the magnitude of the input variables.

[0052]

[0053] Automated architecture search: When building a deep learning model, some hyperparameters that determine the model architecture must be determined, such as: the number of hidden layers? How many neurons or filters should each layer contain? Should the activation function use ReLU or other functions? What dropout ratio should be used? Automatic machine learning algorithms can free researchers from the heavy task of parameter adjustment. This embodiment only needs to specify a decision space that allows the computer to automatically search. The automatic machine learning algorithm can search the architecture space. After the model is evaluated, the neural structure is searched and the neural structure is deformed using Bayesian optimization to generate a new model. Finally, the best neural network architecture model is selected by evaluating the accuracy of the model. The hyperparameter optimization process usually includes:

[0054] a. Select a set of hyperparameters (automatically selected).

[0055] b. Build the corresponding model.

[0056] c. Fit the model on the training data and measure its final performance on the validation data.

[0057] d. Select the next set of hyperparameters to try (automatically selected).

[0058] e. Repeat the above process.

[0059] f. Finally, measure the performance of the model on the test data.

[0060] The key to this process is an algorithm that, given many sets of hyperparameters, uses the historical validation performance to select the next set of hyperparameters to evaluate. A variety of techniques are available, including Bayesian optimization, genetic algorithms, and simple random search. In this example, Neural Architecture Search (NAS) was used to dynamically optimize hyperparameters such as the hidden layer structure and activation function. The optimal architecture ultimately determined by AutoML was: two hidden layers, with 32 and 1024 neurons, respectively, and the ReLU activation function.

[0061] Regularization design: A random dropout mechanism is introduced in the second hidden layer to suppress overfitting.

[0062] Dropout is the application of dropout to a layer in a neural network. This means that during training, some of the output features of that layer are randomly discarded (set to 0). Suppose that during training, the return value of a layer for a given input sample should be the vector [0.2, 0.5, 1.3, 0.8, 1.1]. After using dropout, several random elements of this vector will become 0, such as [0, 0.5, 1.3, 0, 1.1]. The dropout rate is the proportion of features set to 0, typically ranging from 0.2 to 0.5. During testing, no units are discarded, but the output value of the layer needs to be reduced by the dropout rate because more units are activated than during training, which requires balancing. In automated machine learning algorithms, the introduction of dropout can reduce the effect of overfitting.

[0063] Output layer: A linear layer that outputs the predicted value of the modulus of resilience of the waste rock.

[0064] (4) Model training and evaluation

[0065] Adopting the adaptive learning rate strategy, the model is trained using the training data set and trained for 1500 iterations. Figure 5 As shown. The prediction performance of the model is evaluated by the test data set to ensure that the model has good generalization ability, as shown in Figure 6 shown.

[0066] The learning rate is one of the most core hyperparameters in machine learning, determining the magnitude of model adjustments during each parameter update. Specifically, the learning rate controls the step size of the model's movement in the opposite direction of the gradient (i.e., the direction of decreasing loss), acting as a "step regulator" during the optimization process. Excessively large learning rates facilitate rapid initial convergence but can lead to divergence near the optimal solution. Excessively small learning rates can steadily approach the optimal solution, but converge very slowly and can easily fall into local optimality. An adaptive learning rate strategy can continuously optimize the learning rate during model training, mitigating any reduction in model accuracy caused by the learning rate.

[0067] The dataset was randomly split into a training set (70%) and a test set (30%). The machine learning model was first trained on the training set to search for the optimal architecture and hyperparameters. The generalization ability of the constructed model was then verified on the test set. To reduce overfitting during machine learning model training, the original training dataset was further randomly subdivided into a 70% training set and a 30% validation set, used to train the model's architecture and hyperparameters and verify its predictive accuracy, respectively.

[0068] Evaluation metrics for AutoML models include:

[0069] Coefficient of determination (R 2 ): ≥0.95

[0070] Mean absolute error (MAE): ≤30MPa

[0071] (5) Prediction Application

[0072] Input the physical parameters of waste rock and the expected stress state, and output the predicted value of the rebound modulus to guide the design of the road base.

[0073] Furthermore, as a specific implementation of this embodiment, model training and verification are performed through the following example 1.

[0074] Data preparation: The rebound modulus was measured through triaxial compression test, and a total of 200 sets of experimental data were collected, including ρ d =1.8~2.2g / cm 3 、C u =1.5~4.0, σ3=20~1000kPa and other parameters.

[0075] Model configuration:

[0076] Input layer normalization: Z-score normalization is used to process each feature.

[0077] Hidden layer structure: The optimal number of layers (such as 2 layers) and the number of neurons are dynamically determined through NAS.

[0078] Search space settings:

[0079] Table 2

[0080]

[0081] Training process: Set the batch size to 64, the initial learning rate (η) to 0.001, and use early stopping to prevent overfitting.

[0082] Learning rate strategy: CosineAnnealing (T0=15): η max =0.001,η min =0.00005

[0083] Early stopping condition: the validation loss decreases by <1% for 15 consecutive epochs.

[0084] Example 2: Engineering Application

[0085] Input the ρ of a mine waste rock d =2.0g / cm 3 、C u =3.2, n=0.5, σ3=80kPa, σ d =200kPa, the predicted rebound modulus is 320MPa, and the error with the measured value is less than 5%, which meets the design requirements of the road base.

[0086] The present invention has the following technical advantages:

[0087] (1) Improved prediction accuracy: Compared with the traditional empirical model, the AutoML model of the present invention shows higher accuracy (R 2 ≥0.95, MAE<30MPa).

[0088] (2) Enhanced generalization ability: The model can comprehensively consider the physical properties of waste rock (C u =1.5~4.0) and stress state (σ3=20~1000kPa,σ d =20~2000kPa).

[0089] (3) Efficient and convenient: Without relying on expensive experimental equipment, the rebound modulus of waste rock can be quickly predicted, reducing engineering costs.

[0090] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for predicting the rebound modulus of waste rock based on automatic machine learning, characterized in that: The following steps are involved: Collect physical parameters and stress state data of waste rock to construct a multi-source data set; Preprocessing the multi-source dataset to obtain input features; Constructing an automatic machine learning model, and training the automatic machine learning model based on the input features to obtain a trained automatic machine learning model; The physical parameters and stress state data of the waste rock are input into the trained automatic machine learning model to obtain the rebound modulus of the waste rock.

2. The method for predicting the rebound modulus of waste rock based on automatic machine learning according to claim 1, characterized in that: The multi-source data set was obtained by simulating the effect of vehicle load on waste rock using half-sine loading.

3. The method for predicting the rebound modulus of waste rock based on automatic machine learning according to claim 1, characterized in that: Performing Pearson correlation analysis on the multi-source data set to obtain input features; The input characteristics include: dry density, non-uniformity coefficient, limited particle size, effective particle size, gradation curve shape factor, confining pressure and deviatoric stress.

4. The method for predicting the rebound modulus of waste rock based on automatic machine learning according to claim 3, characterized in that: The calculation expression of the gradation curve shape factor is: Where, P i is the weight percentage of waste rock passing the i-th sieve particle size; d i is the opening size of this special sieve; D max is the maximum particle size of waste rock particles; n is the shape factor of the gradient curve.

5. The method for predicting the rebound modulus of waste rock based on automatic machine learning according to claim 1, characterized in that: The hidden layer structure of the automatic machine learning model is dynamically optimized through neural architecture search to determine the optimal number of hidden layers and neurons, and the activation function adopts ReLU.

6. The method for predicting the rebound modulus of waste rock based on automatic machine learning according to claim 1, characterized in that: When training the automatic machine learning model, an adaptive learning rate strategy is used to optimize the model's generalization ability.

7. The method for predicting the rebound modulus of waste rock based on automatic machine learning according to claim 1, characterized in that: The automatic machine learning model includes: an input layer, a normalization layer, a first hidden layer, a second hidden layer and an output layer; The normalization layer includes 7 neurons; the first hidden layer includes 32 neurons; the second hidden layer includes 1024 neurons; and the output layer includes one neuron.

Images