A model training method, a tunnel blasting block size prediction method and device

Abstract

The present application relates to the technical field of tunnel engineering blasting, and provides a model training method, a tunnel blasting block size prediction method and device to solve the problems of the existing tunnel blasting block size prediction method, such as incomplete consideration of influencing factors and low prediction accuracy. The method comprises obtaining a local sample set of tunnel blasting engineering historical sample data, and screening out a first parameter based on a local weighted Pearson correlation algorithm; performing scale scaling processing on the first parameter to obtain a first matrix; constructing a limit learning machine tunnel blasting block size prediction model optimized based on a butterfly optimization algorithm, and calculating a final output weight matrix combined with a smoothing soft truncation decoding mechanism to complete model training, and completing tunnel blasting block size prediction based on the model training method. The method of the present application is suitable for complex and variable tunnel blasting conditions, and the prediction result is more stable and reliable.

Description

Technical Field

[0001] This invention relates to the field of blasting construction technology, specifically to a model training method, a method and apparatus for predicting the size of blasted blocks in tunnels. Background Technology

[0002] Drill and blast method is currently the most widely used excavation method in tunnel engineering, and the size of the blasted fragments is one of the core indicators for evaluating the blasting effect. Excessively large blasted fragments will increase the workload of secondary crushing and affect muck removal efficiency; excessively small fragments will increase explosive consumption and damage to the surrounding rock. Accurately predicting the size of the blasted fragments is of great significance for optimizing blasting design, reducing construction costs, and ensuring project safety.

[0003] However, in tunnel blasting construction, the quality of the blasting design directly affects the blasting construction quality and cost control, as well as the effectiveness of the vibration reduction measures taken against the blast source. Reasonable selection of blasting parameters can proactively control the negative effects of blasting on existing structures (such as tunnel lining, adjacent existing tunnels, or underground pipelines), reduce over- and under-excavation and cumulative damage to the surrounding rock, and reduce the consumption of initial support and secondary lining materials and drilling and blasting construction costs. Therefore, seeking an efficient tunnel blasting block size prediction model has significant practical implications.

[0004] Currently, tunnel blasting block size prediction mainly employs empirical formulas and traditional machine learning methods. Empirical formulas (such as the Kuz-Ram model) consider only a few influencing factors, resulting in low prediction accuracy and poor adaptability. Traditional machine learning methods, such as backpropagation neural networks and support vector machines (SVM), suffer from slow convergence speed, susceptibility to local optima, and insufficient capacity for processing large datasets. Extreme Learning Machine (ELM), as a novel single-hidden-layer feedforward neural network, offers advantages such as fast training speed and strong generalization ability. However, because its input weights and hidden layer biases are randomly generated, the prediction results are unstable, and the accuracy is difficult to meet engineering requirements.

[0005] The Butterfly Optimization Algorithm (BOA) is a metaheuristic optimization algorithm that simulates the foraging behavior of butterflies in nature, which relies on pheromones to sense and transmit scents. The core idea of ​​this algorithm is to dynamically switch between global and local search modes based on the intensity of the scents released by individual butterflies, thereby guiding the swarm towards the optimal solution. Compared with other swarm intelligence optimization algorithms, BOA has advantages such as a simple mechanism, fewer parameters, and strong global search capabilities, and has been widely applied in various engineering fields such as function optimization, power dispatching, image processing, and neural network training. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a model training method, a method and device for predicting tunnel blasting block size based on the Pearson-BOA-ELM model, and employs a weighted Pearson correlation algorithm with homogeneous geological windows and forgetting decay to screen core parameters. This avoids the subjectivity of manual parameter selection, adapts to the gradual changes in geological conditions during tunnel excavation, weakens the interference of heterogeneous data across faults, and improves the matching degree between features and current working conditions.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] The first aspect of this invention provides a model training method, the method comprising:

[0009] A local sample set of historical sample data for tunnel blasting projects is obtained, and the first parameter is selected based on the local weighted Pearson correlation algorithm;

[0010] The first parameter is scaled according to the extreme value boundary algorithm to obtain the first matrix;

[0011] Based on the first matrix, an extreme learning machine tunnel blasting block size prediction model optimized by the butterfly optimization algorithm is constructed, and a global weight matrix containing the optimal input weights and the optimal hidden layer bias is determined.

[0012] The global weight matrix is ​​imported into the tunnel blasting block size prediction model of the extreme learning machine, and the final output weight matrix is ​​calculated by combining the smooth soft truncation decoding mechanism to complete the model training.

[0013] Preferably, the process of obtaining historical sample data for tunnel blasting projects includes:

[0014] Using the current excavation face as the time reference, historical blasting sample data corresponding to the length of the geological window of the same origin are extracted to form a local sample set;

[0015] Twelve initial feature sequences related to the average particle size of the explosive blocks were extracted from the local sample set;

[0016] For each of the twelve initial feature sequences, the weighted Pearson correlation is calculated using the weighted Pearson correlation algorithm;

[0017] The absolute values ​​of the weighted Pearson correlation coefficients are taken and then rearranged in descending order according to their numerical values. The top six initial feature sequences are selected to form a core feature array, which is referred to as the first parameter.

[0018] Preferably, the process of scaling the first parameter according to the extreme value boundary algorithm to obtain the first matrix includes:

[0019] Extract the maximum and minimum values ​​of the first parameter in the local sample set;

[0020] The normalized value is determined based on the maximum and minimum values ​​of the parameter.

[0021] All normalized values ​​are combined to form the first matrix.

[0022] Preferably, the process of constructing an extreme learning machine tunnel blasting block size prediction model based on the butterfly optimization algorithm includes:

[0023] Construct an extreme learning machine network framework consisting of an input layer, a hidden layer, and an output layer;

[0024] The parameters of the butterfly optimization algorithm, including the probability of the perceived modality, are set, and the spatial coordinates of each optimization butterfly are mapped one by one to the input weight matrix and hidden layer bias matrix of the extreme learning machine network framework for optimization.

[0025] During the optimization process, the optimization ends after reaching the maximum number of iterations. The globally optimal butterfly individual with the minimum prediction mean square error is extracted, and its spatial coordinates are analyzed and output as a globally optimal weight matrix containing the optimal input weights and the optimal hidden layer bias.

[0026] Preferably, the optimization process also includes:

[0027] Dynamic random numbers with values ​​between 0 and 1 are generated using the standard uniform distribution function.

[0028] When the dynamic random number is less than the probability of the perceived modality, global wide-area exploration is performed, and the aroma concentration value of the butterfly optimization algorithm is used as the step size weight to perform a large-scale spatial coordinate jump update.

[0029] When the dynamic random number is greater than or equal to the perceived modal probability, a local fine-tuning operation is performed, and combined with the aroma concentration value, a small-range neighborhood coordinate fine-tuning update is performed.

[0030] The optimization ends after reaching the maximum number of iterations.

[0031] Preferably, the process of importing the global weight matrix into the extreme learning machine tunnel blasting block size prediction model and calculating the final output weight matrix using a smooth soft truncation decoding mechanism includes:

[0032] The global weight matrix output in step S3 is split into the input weight matrix and the hidden layer bias matrix;

[0033] The activation function of Sigmoid neurons is preset inside the hidden layer mapping network of the Extreme Learning Machine Network;

[0034] The hidden layer input matrix is ​​determined based on the hidden layer bias matrix, and the historical training mapping matrix is ​​obtained by combining the Sigmoid neuron activation function.

[0035] The singular value decomposition method is used to perform an orthogonal decoupling transformation on the historical training mapping matrix to decompose it into the left singular matrix, the right singular matrix and the singular values ​​of the hidden layer, and a singular truncation threshold is set.

[0036] A smooth soft truncation function based on the activation function of Sigmoid neurons is constructed to perform progressive decay filtering on singular terms less than a threshold, and the target generalized inverse matrix is ​​reconstructed based on the filtered singular values ​​combined with the left and right singular matrices.

[0037] The target generalized inverse matrix is ​​multiplied by the target expectation matrix formed by the average particle size sequence of actual bursts in the training sample set to obtain the final output weight matrix.

[0038] A second aspect of this invention provides a method for predicting tunnel blasting block size based on the Pearson-BOA-ELM model, the method comprising:

[0039] Obtain real-time sample data of the current excavation face and extract the second parameter of the real-time sample data;

[0040] The second parameter is scaled to obtain the second matrix;

[0041] The second matrix is ​​transformed and combined with the output weight matrix obtained based on the model training method described above to finally obtain the predicted tunnel blasting block size.

[0042] Preferably, the process of transforming the second matrix and combining it with the output weight matrix to finally output the predicted tunnel blasting block size includes:

[0043] After performing the same matrix multiplication and addition operations on the second matrix and the input weight matrix and hidden layer bias matrix obtained by the model training method described above, a nonlinear numerical transformation is performed through the Sigmoid activation function to generate the current inference mapping matrix.

[0044] A third aspect of the present invention provides a tunnel blasting block size prediction device based on the Pearson-BOA-ELM model. The device includes:

[0045] The acquisition module acquires real-time sample data of the current excavation face and extracts the second parameter of the real-time sample data;

[0046] The preprocessing module is used to scale the second parameter to obtain the second matrix;

[0047] The prediction module transforms the second matrix and combines it with the output weight matrix obtained based on the model training method described in any one of claims 1 to 6 to finally obtain the predicted tunnel blasting block size.

[0048] The specific beneficial effects are as follows:

[0049] 1. A weighted Pearson correlation algorithm with homogeneous geological windows and forgetting decay is used to screen core parameters. This avoids the subjectivity of manual parameter selection, adapts to the gradual changes in geological conditions during tunnel excavation, weakens the interference of heterogeneous data across faults, and improves the matching degree between features and current working conditions.

[0050] 2. The initial weights and biases of ELM are optimized by utilizing the global search capability of the butterfly optimization algorithm, which solves the problems of unstable results and easy getting trapped in local optima caused by the random initialization of traditional ELM.

[0051] 3. A dynamic scaling aroma intensity mechanism based on prediction error feedback is designed to automatically reduce the search step size and individual differences at the end of the optimization process, avoiding random noise in the model fitting the field. Compared with the conventional BOA-ELM model, it has better generalization performance and is suitable for complex and variable tunnel blasting conditions.

[0052] 4. A smooth soft-truncation singular value processing strategy is introduced in the output weight calculation stage to replace the traditional hard threshold truncation. This solves the output oscillation problem caused by the collinearity of the hidden layer matrix, has a stronger buffering capacity against field measurement disturbances, and the prediction results are more stable and reliable. Attached Figure Description

[0053] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a flowchart illustrating the model training method of the present invention;

[0055] Figure 2 This is a flowchart illustrating the tunnel blasting prediction method of the present invention.

[0056] Figure 3 A comparison chart showing the local weighted Pearson correlation descending order rearrangement of the initial feature sequences.

[0057] Figure 4 A comparison of the convergence curves of the normalized prediction mean square error for different algorithms.

[0058] Figure 5This is a comparison chart showing the actual average particle size of the exploded blocks with the predicted values ​​from various models. Detailed Implementation

[0059] The following will refer to the appendices in the embodiments of this application. Figures 1-5 The exemplary embodiments of this application are described in more detail below. Although exemplary embodiments of this application are shown in the accompanying drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of this application to those skilled in the art.

[0060] Example 1: As Figure 1 and Figure 2 A tunnel blasting prediction method based on an improved butterfly optimization algorithm includes the following steps:

[0061] S1. Obtain a local sample set of historical sample data for tunnel blasting projects, and select the first parameter based on the local weighted Pearson correlation algorithm;

[0062] In tunnel blasting, the resistance of rock to blasting is related to its physical and mechanical properties and geological conditions. Among the physical and mechanical properties of the rock, those relevant to blasting include density, porosity, degree of weathering, wave impedance, frost resistance, deformation characteristics, uniaxial compressive strength, uniaxial tensile strength, and shear strength. Furthermore, blasting parameters are determined in blasting design based on the physical and mechanical properties and geological structure of the rock. During blasting, when the borehole passes through weak rock layers or contains cavities, blasting energy may be released in other ways, reducing energy utilization and weakening the blasting effect. Therefore, when arranging borehole locations, the geological structures of the blasting area, such as joints, fissures, and weak interlayers, must be considered to prevent these weak interlayers from weakening the blasting effect.

[0063] In this embodiment, the current excavation face is used as the time reference, and historical blasting sample data corresponding to the length of the same geological window are extracted to form a local sample set.

[0064] Twelve initial feature sequences related to the average particle size of the blasting blocks were extracted from the local sample set. Each initial feature sequence is composed of the feature sample values ​​corresponding to each single blast in the local sample set. Specifically, these include density, uniaxial compressive strength, uniaxial tensile strength and wave impedance as rock physical and mechanical parameters, as well as minimum resistance line, tunnel excavation outline angle, explosive consumption, hole spacing, row spacing, borehole depth, charge coefficient and detonation delay time as blasting design parameters.

[0065] Simultaneously, the average particle size sequence of the actual blasted blocks corresponding to a single blast is extracted as an evaluation target;

[0066] For each of the twelve initial feature sequences, the weighted Pearson correlation algorithm is executed, specifically as follows:

[0067] Using the homologous geological window covered by the local sample set as the local interval, and using the preset forgetting decay base as the base and the excavation time sequence distance of the historical sample from the current working face as the exponent, a time decay weight is constructed.

[0068] By using time decay weighting, the feature sample values ​​within this local interval and the actual measured average particle size of the explosive block are weighted averaged to obtain the weighted feature mean and the weighted particle size mean.

[0069] Subsequently, the deviations between the feature sample values ​​and the weighted feature mean, as well as the deviations between the actual measured average particle size of the burst blocks and the weighted particle size mean, were calculated.

[0070] The time decay weight is multiplied by the product of the two aforementioned biases and by the square of each of the two aforementioned biases, respectively, to calculate the ratio of the product of the local weighted covariance and the respective weighted standard deviation, thereby generating the local weighted Pearson correlation of the corresponding initial feature sequence.

[0071] Specifically, the locally weighted Pearson correlation satisfies the following relationship:

[0072] ;

[0073] In the formula, For locally weighted Pearson correlation, This is the index for the current excavation face. This refers to a geological window of the same origin, which is a preset number of historical blasting cycles. For local sampling index, To characterize the forgetting decay basis of the time distance penalty, The feature sample value corresponding to the sampling index. This represents the weighted mean of features within the same geological window. The average particle size of the actual burst block corresponding to the sampling index. It represents the weighted average grain size within the same geological window.

[0074] in, This is the result of a power operation with the forgetting decay basis μ as the base and the temporal distance (tk) from the historical sample to the current time as the exponent. The larger the value, the closer the historical blasting sample corresponding to the local sampling index is to the spatiotemporal coordinates of the current excavation face. This makes the weight of feature sample values ​​that are closer to the current working face in the local weighted Pearson correlation calculation tend to increase. This makes the calculated correlation result more likely to reflect the transient physical dominant law of the current stratum lithology, and thus more likely to achieve the purpose of preventing feature hysteresis misjudgment caused by abrupt geological fault changes. The smaller the value, the further back in time and space the historical blasting samples are from the excavation. This reduces the contribution of old geological data in the local weighted Pearson correlation calculation, making the interference of cross-fault heterogeneous data on the current local correlation fitting weaker. As a result, it is more likely to achieve the goal of smoothing the geological transition period and extracting high-fidelity micro-correlation features.

[0075] For example, the origin geological window determines the time span by which the algorithm backtracks historical blasting samples when evaluating local correlation. The empirical value range is [10, 30]. In this embodiment, the origin geological window is set to 20. Implementers can set the origin geological window according to the severity of variations in tunnel geological faults. For example, when geological conditions are extremely unstable and lithological changes are frequent, the origin geological window can be appropriately reduced to improve the algorithm's response speed to new geological data; when the strata lithology is relatively homogeneous and stable, the origin geological window can be appropriately increased to enhance the robustness of statistical features.

[0076] For example, the forgetting attenuation base determines how quickly historical geological data diminishes in relevance assessment over time, with an empirical value range of [0.85, 0.99]. In this embodiment, the forgetting attenuation base is set to 0.95. Implementers can set the forgetting attenuation base according to the stratigraphic transition rate of the excavation face. For instance, when crossing a fault with uneven hardness, the forgetting attenuation base can be appropriately reduced to accelerate the elimination of old geological features; when the geological transition is relatively gentle, the forgetting attenuation base can be appropriately increased to retain more features from recent homologous samples.

[0077] The absolute values ​​of the weighted Pearson correlation coefficients are taken and rearranged in descending order according to their numerical values. The top six initial feature sequences are extracted, namely tensile strength, minimum resistance line, tunnel excavation outline angle, explosive consumption, hole spacing and row spacing, to form a core feature array. The core feature array is collectively referred to as the first parameter.

[0078] like Figure 3The figure shows a comparison of the local weighted Pearson correlation descending order rearrangement of the initial feature sequences. As shown, conventional manual parameter selection strategies rely on subjective experience to extract initial features, which are greatly affected by heterogeneous cross-fault data. This not only increases the computational load during model training but also easily leads to feature redundancy and mapping bias when operating conditions vary. In contrast, this invention uses a weighted Pearson correlation algorithm to calculate local weighted Pearson correlations in the high-value region of the correlation distribution. The absolute values ​​of the local weighted Pearson correlations are then taken and rearranged in descending order to eliminate interference from heterogeneous cross-fault data, ensuring that the features match the current operating conditions and reducing training errors caused by low-correlation data. In the truncation region where the core feature array is defined, the top six initial feature sequences are strictly selected based on their ranking to form the core feature array as the first parameter, reducing computational complexity and the risk of model overfitting. In the clearing region of the remaining initial feature sequences, the influence of secondary factors is removed, and a fixed number of segments are used to present a clear core feature array. The screening results show that the present invention effectively takes into account the extraction requirements of key mechanical properties and core geometric parameters, suppresses full-dimensional input interference, and improves the parameter optimization efficiency of the extreme learning machine tunnel blasting block size prediction model while improving the matching degree between features and current working conditions.

[0079] S2. The first matrix is ​​obtained by performing dimensionless scaling on the first parameter according to the extreme value boundary algorithm.

[0080] It should be noted that the first parameter selected for inclusion in the database includes cross-dimensional measurements of stress and pressure attributes as well as spatial distance attributes. The heterogeneity of the underlying units causes the numerical features of each parameter to exhibit a wide range of abrupt changes. When an unnormalized data matrix is ​​injected into the hidden layer mapping network of an extreme learning machine, large-scale values ​​will forcibly dominate the activation state, inducing a mapping bias limitation where small-scale feature signals are deeply obscured. Therefore, this invention measures the relative weight potential of a single physical feature within the global limit boundary using normalized values, effectively removing the dimensional misalignment weights caused by differences in measurement units, and providing a stable, balanced, and dimensionless computational source with anti-saturation margin for the nonlinear activation of subsequent hidden layer nodes.

[0081] Specifically, the extreme value boundary algorithm is to extract the maximum and minimum values ​​of the first parameter in the local sample set, and to extract the basic difference by subtracting the minimum value from the measured value of a single first parameter.

[0082] The range span is derived by subtracting the minimum value from the maximum value of the parameter. The ratio of the basic difference to the range span is calculated and multiplied by the primary scaling constant. The product result is then superimposed with the secondary bias constant to obtain the normalized value.

[0083] The normalized values ​​of all transformations of the first parameter are combined into the first matrix.

[0084] Specifically, the normalized values ​​satisfy the following relation:

[0085] ;

[0086] In the formula, Normalized value For the current measured value of a single first parameter, For the parameter minimum value in a local sample set, For the local sample set, the parameter is the maximum value. The main scaling constant, This is the secondary bias constant.

[0087] in, The larger the value, the closer the current measured value of the single first parameter is to the extreme pressure limit of the local rock stratum objective environment, the larger the value of the normalized value formula term tends to be, the stronger the characteristic activation peak after compression mapping tends to be, and the more likely it is to highlight the current extreme blasting resistance characteristics. The smaller the value, the closer the current measured value of the single first parameter is to the lower bound of the numerical mapping allowed by the construction project. This makes the value of the normalized value formula term tend to decrease, and the converted excitation feedback tends to be in a deep suppression and blocking period. Thus, it tends to achieve the stable control purpose of suppressing the overreaction caused by the sudden change of the conventional edge parameter.

[0088] For example, the principal scaling constant determines the span of the dynamic distribution range of the feature data after dimensionless mapping, with an empirical value range of [0.85, 0.99]. In this embodiment, the principal scaling constant is set to 0.95. Implementers can set the principal scaling constant according to the linear and nonlinear region characteristics of the hidden layer activation function. For example, when it is necessary to avoid neurons from entering deep saturation dead zones to maintain gradient descent capability, the principal scaling constant can be appropriately reduced to compress feature peaks; when it is necessary to enhance feature contrast, the principal scaling constant can be appropriately increased.

[0089] For example, the secondary bias constant determines the safe buffer potential for shifting the normalized data as a whole towards the non-zero interval. Empirically, its value range is [0.01, 0.10]. In this embodiment, the secondary bias constant is set to 0.05. Implementers can set the secondary bias constant according to the impact of extreme minimum values ​​of the input source on the network's computational stability. For instance, when the input data contains a large number of minimum values ​​approaching zero and easily induces neuronal inactivation and dormancy, the secondary bias constant can be appropriately increased to raise the anti-saturation baseline; when the data itself is far from zero, the secondary bias constant can be appropriately decreased to reduce human mapping bias.

[0090] S3. Based on step S2, construct an extreme learning machine tunnel blasting block size prediction model optimized by the butterfly optimization algorithm, and determine a global weight matrix containing the optimal input weights and the optimal hidden layer bias.

[0091] It should be noted that, as a single-hidden-layer feedforward neural network, the traditional Extreme Learning Machine (ELM) randomly generates its input weights and hidden layer biases, leading to unstable prediction results. Furthermore, the multi-source noise at the blasting site causes the network mapping to form a wide-ranging, flat error basin. If the traditional butterfly optimization algorithm is simply introduced for optimization, when the error approaches the physical noise floor at the end of the optimization process, it can easily induce the network to over-fit the random white noise, causing the prediction model to lose its engineering generalization ability.

[0092] Therefore, this invention constructs a dynamic scaling basis that depends on the feedback of the prediction mean square error, provides high-intensity contrast stretching during the flat period of the error basin, adaptively attenuates the amplification intensity when the error approaches the noise floor, and uses an improved butterfly optimization algorithm to perform global collaborative optimization of the input parameters of the extreme learning machine, providing a core mapping operator with both accuracy and generalization robustness for block degree extrapolation.

[0093] Specifically, an extreme learning machine network framework consisting of an input layer, a hidden layer, and an output layer is constructed, and a training sample set consisting of the first matrix output in step S2 and the average particle size of the actual burst blocks is obtained.

[0094] The initial population size and maximum number of iterations for butterfly optimization are set. The empirical range of the initial population size is 30 to 100, and it is set to 50 in this embodiment. The empirical range of the maximum number of iterations is 100 to 500, and it is set to 200 in this embodiment. An initial perception level is preset, and a perception mode probability for controlling the switching of search modes is preset. The value range of the perception mode probability is greater than 0 and less than 1, and it is set to 0.6 in this embodiment.

[0095] Define spatial mapping logic to map the spatial coordinates of each optimization butterfly individual to the input weight matrix and hidden layer bias matrix of the extreme learning machine network framework;

[0096] During the optimization iteration process, only the first matrix in the training sample set is injected into the extreme learning machine network, and the historical training mapping matrix is ​​generated by combining the input weights of the current individual with the hidden layer bias.

[0097] Based on the generalized inverse theory, combined with the historical training mapping matrix and the actual average particle size of the bursting blocks in the training sample set, a temporary output weight is calculated, and based on the temporary output weight, the training set prediction output under the current extreme learning machine network state is obtained. The training set prediction output is compared with the actual average particle size of the bursting blocks to generate the prediction mean square error.

[0098] Specifically, the prediction mean squared error satisfies the following relationship:

[0099] ;

[0100] In the formula, To predict the mean squared error, The total number of samples in the training sample set. This represents the actual average particle size of the bursting blocks. Predict the output for the training set.

[0101] Implementers can set the basic absorption term based on the breadth of the optimization space and the error perception coefficient based on the tolerance of the extreme learning machine network to overfitting risk. Based on the preset basic absorption term and error perception coefficient, and combined with the predicted mean square error, a dynamic scaling basis is obtained. Specifically, the dynamic scaling basis satisfies the following relationship:

[0102] ;

[0103] In the formula, For dynamic scaling base, Basic absorption term, The error perception coefficient. To predict the mean squared error.

[0104] To prevent the denominator from overflowing due to the prediction mean square error approaching zero in the later stages of optimization, a basin escape basis with minimal constant characteristics is introduced. The prediction mean square error is transformed into an inverse scalar representing individual fitness through reciprocal operations. The aroma concentration value of the current butterfly individual is then updated using a dynamic scaling basis and the initial perception value. Specifically, the aroma concentration value satisfies the following relationship:

[0105] ;

[0106] In the formula, This represents the aromatic concentration value. Initial perception level, To predict the mean squared error, As an escape base for the basin, It is a dynamically scaling base.

[0107] in, The larger the value, the more likely the current population is in a flat region with high error, indicating that the extreme learning machine network has failed to effectively capture geological patterns. Approaching 0, let Increase and approach the basic absorption term At this point, the characteristic differences of individuals with better fitness are amplified, guiding all individuals in the domain to quickly aggregate towards high-confidence regions, breaking through the optimization stagnation trap caused by multi-source noise; The smaller the value, the more it indicates that the algorithm has entered the final convergence stage, approaching the physical noise floor, reflecting that the Extreme Learning Machine network has basically fitted the geological features, making... Approaching 1, let The error decays rapidly and approaches zero. At this point, the amplification effect between small errors is smoothed out, the error jumps induced by random noise are eliminated, and the model is prevented from losing its generalization ability due to overfitting to ineffective background noise.

[0108] Furthermore, The larger the value, the more the Extreme Learning Machine Network needs to increase individual differences to accelerate convergence. Increase the size of the algorithm, giving it greater dominance in the search for optimal results and improving the search sensitivity of the swarm intelligence algorithm in complex geological environments; The smaller the value, the more likely the Extreme Learning Machine (ELM) network has reached its accuracy threshold and requires local fine-tuning. This reduces the stretching gain of fitness differences on the optimization process, thereby protecting the established effective mapping features and avoiding drastic oscillations in network weights caused by abnormal sample interference.

[0109] Dynamic random numbers with values ​​between 0 and 1 are generated using a standard uniform distribution function. Based on the comparison between these dynamic random numbers and the perceived modality probabilities, the spatial coordinate update strategy is dynamically switched.

[0110] When the dynamic random number is less than the probability of the perception mode, global wide-area exploration is performed. Specifically, the current optimization individual is guided to take the global best individual in the population as the center of gravity, and the aromatic concentration value calculated above is used as the step size weight to perform a large-scale spatial coordinate jump update.

[0111] When the dynamic random number is greater than or equal to the perceived modal probability, a local fine-tuning operation is performed. That is, the current optimization individual is guided to randomly select the coordinate difference between two other heterogeneous individuals within the current population as a reference benchmark, and combined with the aroma concentration value, a small-scale neighborhood coordinate fine-tuning update is performed.

[0112] The optimization ends after reaching the maximum number of iterations. The globally optimal butterfly individual with the minimum prediction mean square error is extracted, and its spatial coordinates are analyzed and output as a globally optimal weight matrix containing the optimal input weights and the optimal hidden layer bias.

[0113] For example, the basic absorbing term determines the upper limit of the maximum tensile intensity that the dynamic scaling basis can achieve in an extremely flat error basin, with an empirical value range of 0.1 to 0.9. In this embodiment, the basic absorbing term is set to 0.5. Implementers can set the basic absorbing term according to the breadth of the optimization space. For example, when the space is extremely large and easily gets trapped in local optima, the basic absorbing term can be appropriately increased to give the algorithm a stronger ability to escape traps; when the network parameter scale is small and smooth convergence is required, the basic absorbing term can be appropriately decreased to prevent excessive divergence.

[0114] For example, the error perception coefficient determines the sensitivity of the dynamic scaling basis to attenuation when the predicted mean square error approaches the noise floor. Empirically, its value ranges from 5 to 20; in this embodiment, the error perception coefficient is set to 10. Implementers can set the error perception coefficient according to the extreme learning machine network's tolerance for overfitting risk. For instance, when it is necessary to strictly prevent the extreme learning machine network from fitting random white noise that lacks physical meaning, the error perception coefficient can be appropriately increased to accelerate exponential decay at the end of the optimization phase; when it is necessary to deeply mine extremely weak physical mapping features, the error perception coefficient can be appropriately decreased to maintain scaling activity.

[0115] For example, the initial perception level and the basin escape basis determine the basic optimization step size and the zero overflow prevention baseline, respectively. The empirical value range of the initial perception level is 0.005 to 0.05, and the empirical value range of the basin escape basis is... to In this embodiment, the initial perception level is set to 0.01, and the basin escape baseline is set to... The implementers can make routine fine-tuning adjustments based on the variance of the underlying data.

[0116] S4. Import the global weight matrix into the tunnel blasting block size prediction model of the extreme learning machine, and calculate the final output weight matrix by combining the smooth soft truncation decoding mechanism.

[0117] It should be noted that the residual collinearity of explosive mechanics in the latent space of the Extreme Learning Machine (ELM) network leads to highly singular mapping features. Applying a conventional hard threshold to truncate small singular values ​​disrupts the continuity of the mathematical mapping space. When the input signal is subject to weak field perturbations, hard truncation causes drastic numerical jumps in the output weights between including and discarding features, inducing oscillations in the predicted output.

[0118] Therefore, this invention introduces a continuously differentiable smooth soft truncation function inside the solution of the generalized inverse operator, and performs a progressive smooth transition filtering of the singular value components that tend to zero, providing an intuitive and highly stable predictive scalar for the optimization of blasting construction design.

[0119] Specifically, in the final parameter calculation stage, in order to balance the optimization computing efficiency and the noise robustness of the final model, the global weight matrix output in step S3 is split into an input weight matrix and a hidden layer bias matrix, which is different from the standard generalized inverse operation used for fast fitness evaluation in step S3.

[0120] The activation function of Sigmoid neurons is preset inside the hidden layer mapping network of the Extreme Learning Machine Network;

[0121] Perform matrix multiplication on the normalized feature matrix extracted in step S2 and the input weight matrix, and superimpose the multiplication result on the hidden layer bias matrix to calculate the hidden layer input matrix; input each element of the hidden layer input matrix into the Sigmoid activation function to perform nonlinear numerical transformation to obtain the historical training mapping matrix.

[0122] The singular value decomposition method is used to perform an orthogonal decoupling transformation on the historical training mapping matrix, which decomposes the left singular matrix, the right singular matrix, and the hidden layer singular values ​​that reflect the intrinsic structural characteristics of the matrix, and a singular truncation threshold is set.

[0123] A smooth soft truncation function based on the Sigmoid neuron activation function is constructed. This function is called to compare the singular values ​​of the hidden layer with the set singular truncation threshold. Small singular terms smaller than the threshold are subjected to progressive decay filtering. The target generalized inverse matrix is ​​reconstructed based on the filtered singular values ​​and the left and right singular matrices.

[0124] The target generalized inverse matrix is ​​multiplied by the target expectation matrix formed by the average particle size sequence of actual bursts in the training sample set to obtain the final output weight matrix.

[0125] S5. Combine the second matrix with the input weight matrix and the hidden layer bias matrix to obtain the current inference mapping matrix; perform matrix multiplication between the current inference mapping matrix and the final output weight matrix to finally output the predicted tunnel blasting block size.

[0126] Entering the real-time inference stage, real-time blasting monitoring data with isomorphic properties to the training sample set is acquired, that is, real-time sample data of the current excavation face is acquired. The real-time sample data includes blasting construction parameters and geomechanical parameters. Feature terms with the same dimension as when constructing the historical training sample set are extracted, and the maximum and minimum values ​​of the parameters extracted in step S2 are strictly called to perform the same dimensionless scaling processing to form a second matrix, which is also called the real-time normalized feature matrix.

[0127] After performing the same matrix multiplication and addition operations on the real-time normalized feature matrix and the previously split input weight matrix and hidden layer bias matrix, a nonlinear numerical transformation is also performed through the Sigmoid activation function to generate the current inference mapping matrix.

[0128] Multiply the current inference mapping matrix with the previously obtained final output weight matrix by an execution matrix to obtain the tunnel blasting block size predicted at the current working face.

[0129] The specific calculation process of the aforementioned smooth truncation decoding and output prediction satisfies the following relationship:

[0130] ;

[0131] In the formula, For the first Diagonal terms of inverse features The first training mapping array for history A singular value in the hidden layer. For the sequence index of the singular values ​​of the hidden layer, The base of the natural logarithm, To smoothly transition the curvature, This is the singular truncation threshold.

[0132] ;

[0133] In the formula, The target generalized inverse matrix of the historical training mapping matrix, For the right singular matrix, Let be the left singular matrix. This represents performing a transpose operation on the left singular matrix. For the first to the second A diagonal matrix consisting of inverse diagonal terms and their main diagonal elements. This represents the total number of singular values ​​in the hidden layer.

[0134] ;

[0135] In the formula, The final output weight matrix, The average particle size of all actual bursts in the training sample set The target expectation matrix is ​​formed by arranging them in sequence.

[0136] ;

[0137] In the formula, The predicted blasting block size at the current tunnel face. This is the current inference mapping matrix.

[0138] in, The larger the value, the better. The more abundant the effective physical features and the less risk of collinear distortion, the better. The closer the value is to 1, the more... The more it tends to preserve the original Numerical values ​​make the model more sensitive to decoding effective signals, thereby improving the algorithm's accuracy in capturing hidden fault anomaly signals; The smaller the value, the better. The closer it gets, the more collinear distortion it causes. ,make The numerical values ​​exhibit a continuous and smooth decay and approach 0, which forcibly suppresses the amplification factor of tiny singular values, enabling the network to buffer against field measurement errors, thereby smoothing the final prediction output and preventing drastic jumps in the predicted values.

[0139] For example, the singularity truncation threshold determines the risk threshold for mathematical collinearity distortion caused by the hidden layer feature matrix, and the empirical range is [value missing]. to In this embodiment, the singularity truncation threshold is set to Implementers can set the singularity truncation threshold according to the severity of fluctuations in the condition number of the hidden layer array. For example, when the collinearity between the input blasting parameters is extremely strong, making it easy for matrix inversion numerical explosions to occur, the singularity truncation threshold can be appropriately increased to filter dangerous singular values ​​in advance. When it is necessary to retain more weak specific mapping features, the singularity truncation threshold can be appropriately decreased to avoid accidentally deleting useful underlying structural information.

[0140] For example, the smooth transition curvature determines the flexible buffer rate of damping attenuation when singular values ​​approach the truncation threshold. Empirically, the value ranges from 100 to 1000; in this embodiment, the smooth transition curvature is set to 500. Implementers can adjust the smooth transition curvature based on the frequency of weak singular value crossings caused by on-site measurement disturbances. For instance, when the measurement sensor is extremely unstable, causing singular values ​​to frequently jump around the singular truncation threshold, the smooth transition curvature can be appropriately reduced to lengthen the smooth transition band and alleviate output oscillations. When the input signal has a high signal-to-noise ratio and requires clean, hard feature truncation, the smooth transition curvature can be appropriately increased to accelerate the filtering of weak noise.

[0141] Example 2: A tunnel blasting block size prediction device based on the Pearson-BOA-ELM model, the device comprising:

[0142] The acquisition module acquires real-time sample data of the current excavation face and extracts the second parameter of the real-time sample data;

[0143] The preprocessing module is used to preprocess the second parameter to obtain the second matrix;

[0144] The prediction module is used to input the preprocessed second matrix into the Extreme Learning Machine tunnel blasting block size prediction model based on the butterfly optimization algorithm, and output the predicted tunnel blasting block size; wherein, the Extreme Learning Machine tunnel blasting block size prediction model based on the butterfly optimization algorithm is obtained based on the model training method described above.

[0145] The tunnel blasting block size prediction device based on the Pearson-BOA-ELM model in this application embodiment can be an electronic device or a component of an electronic device, such as an integrated circuit or a chip. The electronic device can be a terminal or other devices besides a terminal. For example, the electronic device can be a GPU BOX, mobile phone, tablet computer, laptop computer, PDA, in-vehicle electronic device, mobile internet device (MID), augmented reality (AR) / virtual reality (VR) device, robot, wearable device, ultra-mobile personal computer (UMPC), netbook, or personal digital assistant (PDA), etc. It can also be a server, network attached storage (NAS), personal computer (PC), television (TV), ATM, or self-service machine, etc. This application embodiment does not specifically limit the device.

[0146] The tunnel blasting block size prediction device based on the Pearson-BOA-ELM model in this application embodiment can be a device with an operating system. This operating system can be Android, Linux, Windows, or other possible operating systems; this application embodiment does not specifically limit it.

[0147] This application provides an electronic device, which includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the model training method of the foregoing embodiments, or the steps in a tunnel blasting block size prediction method based on the Pearson-BOA-ELM model.

[0148] This application also provides a computer-readable storage medium storing a computer program / instruction thereon, which, when executed by a processor, implements the model training method disclosed in this application, or a step in tunnel blasting block size prediction based on the Pearson-BOA-ELM model.

[0149] This application also provides a computer program product that, when run on an electronic device, enables the processor to implement the model training method disclosed in this application, or a step in tunnel blasting block size prediction based on the Pearson-BOA-ELM model.

[0150] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0151] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, electronic devices, and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0152] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0153] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0154] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0155] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0156] Simulation verification:

[0157] Figure 4 The figure shows a comparison of the convergence curves of the normalized prediction mean squared error (MRSE) for different algorithms. As shown, the traditional BP neural network model, using gradient descent, exhibits a slow decrease in MSE during the initial training phase and is susceptible to local optima in the later stages, leading to incomplete overall convergence. While the conventional butterfly-optimized extreme learning machine (EXM) model shows some convergence speed in the early training phase, the lack of a dynamic adjustment mechanism for the optimization step size makes it prone to misclassifying random noise as valid features in the later optimization phase, nearing the physical noise floor. This results in continuous oscillations in the MSE towards the end of convergence. In contrast, the model of this invention achieves rapid convergence in the initial training phase and performs local fine-tuning through dynamic adjustment of the search step size in the later optimization phase. This maintains a slight downward trend in the MSE towards the end of convergence, effectively avoiding premature convergence and achieving a lower global convergence error. Training results demonstrate that the model of this invention effectively suppresses noise interference in the later stages and improves the parameter optimization efficiency of the EXM tunnel blasting block size prediction model.

[0158] Figure 5The figure shows a comparison between the actual average particle size of the blasted blocks and the predicted values ​​of various models. As shown, the conventional butterfly-optimized extreme learning machine model uses a traditional hard threshold truncation, which is greatly affected by field measurement disturbances. Its predicted values ​​not only exhibit dense up-and-down oscillations during the stable operation period, but also show sluggish responses when encountering sudden changes in geological fault conditions, deviating significantly from the actual average particle size of the blasted blocks. In contrast, the model of this invention buffers measurement disturbances and filters high-frequency noise through a smooth soft truncation function during the stable operation period, keeping the predicted values ​​stable and consistent with the steady-state characteristics of tunnel blasting engineering. During the sudden change in conditions caused by geological faults, which cause the average particle size of the blasted blocks to surge, the model receives the sudden input features and immediately makes a positive response, with the predicted values ​​showing an upward trend jump. In the subsequent two to three blasting cycles, it gradually overcomes the historical data inertia of the same geological window, achieving stable tracking of the actual average particle size of the blasted blocks. During the data adjustment period after passing through the fault, the model adaptively adjusts to fit the actual average particle size of the blasted blocks. The tracking results show that the model of the present invention effectively takes into account both steady-state disturbance prevention and transient response requirements, suppresses cross-condition prediction jumps, and improves the stability of tunnel blasting block size prediction while providing a reliable basis for optimizing tunnel blasting construction parameters.

[0159] The above provides a detailed description of a model training method, a tunnel blasting block size prediction method and apparatus based on the Pearson-BOA-ELM model, and the implementation methods of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A model training method, characterized in that, The method includes: A local sample set of historical sample data for tunnel blasting projects is obtained, and the first parameter is selected based on the local weighted Pearson correlation algorithm; The first parameter is scaled according to the extreme value boundary algorithm to obtain the first matrix; Based on the first matrix, an extreme learning machine tunnel blasting block size prediction model optimized by the butterfly optimization algorithm is constructed, and a global weight matrix containing the optimal input weights and the optimal hidden layer bias is determined. The global weight matrix is ​​imported into the tunnel blasting block size prediction model of the extreme learning machine, and the final output weight matrix is ​​calculated by combining the smooth soft truncation decoding mechanism to complete the model training.

2. The method according to claim 1, characterized in that, The process of obtaining historical sample data for tunnel blasting projects includes: Using the current excavation face as the time reference, historical blasting sample data corresponding to the length of the geological window of the same origin are extracted to form a local sample set; Twelve initial feature sequences related to the average particle size of the bursts were extracted from a local sample set; For each of the twelve initial feature sequences, the weighted Pearson correlation is calculated using the weighted Pearson correlation algorithm; The absolute values ​​of the weighted Pearson correlation coefficients are taken and then rearranged in descending order according to their numerical values. The top six initial feature sequences are selected to form a core feature array, which is referred to as the first parameter.

3. The method according to claim 2, characterized in that, The process of scaling the first parameter according to the extreme value boundary algorithm to obtain the first matrix includes: Extract the maximum and minimum values ​​of the first parameter in the local sample set; The normalized value is determined based on the maximum and minimum values ​​of the parameter. All normalized values ​​are combined to form the first matrix.

4. The method according to claim 3, characterized in that, The process of constructing an extreme learning machine tunnel blasting block size prediction model based on the butterfly optimization algorithm includes: Construct an extreme learning machine network framework consisting of an input layer, a hidden layer, and an output layer; The parameters of the butterfly optimization algorithm, including the probability of the perceived modality, are set, and the spatial coordinates of each optimization butterfly are mapped one by one to the input weight matrix and hidden layer bias matrix of the extreme learning machine network framework for optimization. During the optimization process, the optimization ends after reaching the maximum number of iterations. The globally optimal butterfly individual with the minimum prediction mean square error is extracted, and its spatial coordinates are analyzed and output as a globally optimal weight matrix containing the optimal input weights and the optimal hidden layer bias.

5. The method according to claim 4, characterized in that, The optimization process also includes: Dynamic random numbers with values ​​between 0 and 1 are generated using the standard uniform distribution function. When the dynamic random number is less than the probability of the perceived modality, global wide-area exploration is performed, and the aroma concentration value of the butterfly optimization algorithm is used as the step size weight to perform a large-scale spatial coordinate jump update. When the dynamic random number is greater than or equal to the perceived modal probability, a local fine-tuning operation is performed, and combined with the aroma concentration value, a small-range neighborhood coordinate fine-tuning update is performed. The optimization ends after reaching the maximum number of iterations.

6. The method according to claim 5, characterized in that, The process of importing the global weight matrix into the extreme learning machine tunnel blasting block size prediction model and calculating the final output weight matrix using a smooth soft-truncation decoding mechanism includes: The global weight matrix output in step S3 is split into the input weight matrix and the hidden layer bias matrix; The activation function of Sigmoid neurons is preset inside the hidden layer mapping network of the Extreme Learning Machine Network; The hidden layer input matrix is ​​determined based on the hidden layer bias matrix, and the historical training mapping matrix is ​​obtained by combining the Sigmoid neuron activation function. The singular value decomposition method is used to perform an orthogonal decoupling transformation on the historical training mapping matrix to decompose it into the left singular matrix, the right singular matrix and the singular values ​​of the hidden layer, and a singular truncation threshold is set. A smooth soft truncation function based on the activation function of Sigmoid neurons is constructed to perform progressive decay filtering on singular terms less than a threshold, and the target generalized inverse matrix is ​​reconstructed based on the filtered singular values ​​combined with the left and right singular matrices. The target generalized inverse matrix is ​​multiplied by the target expectation matrix formed by the average particle size sequence of actual bursts in the training sample set to obtain the final output weight matrix.

7. A method for predicting the size of blasted blocks in tunnels, characterized in that, The method includes: Obtain real-time sample data of the current excavation face and extract the second parameter of the real-time sample data; The second parameter is scaled to obtain the second matrix; The second matrix is ​​transformed and combined with the output weight matrix obtained based on the model training method described in any one of claims 1 to 6 to finally obtain the predicted tunnel blasting block size.

8. The prediction method according to claim 7, characterized in that, The process of transforming the second matrix and combining it with the output weight matrix to finally output the predicted tunnel blasting block size includes: After performing the same matrix multiplication and addition operations on the second matrix and the input weight matrix and hidden layer bias matrix obtained by the model training method according to any one of claims 1 to 6, a nonlinear numerical transformation is performed through the Sigmoid activation function to obtain the current inference mapping matrix. The current inference mapping matrix is ​​then multiplied by the previously obtained final output weight matrix to obtain the predicted tunnel blasting block size.

9. A tunnel blasting block size prediction device, characterized in that, The device includes: The acquisition module acquires real-time sample data of the current excavation face and extracts the second parameter of the real-time sample data; The preprocessing module is used to scale the second parameter to obtain the second matrix; The prediction module transforms the second matrix and combines it with the output weight matrix obtained based on the model training method described in any one of claims 1 to 6 to finally obtain the predicted tunnel blasting block size.

Images