Anchoring fracture rock deformation displacement lightweight prediction method and equipment

Abstract

The invention discloses an anchoring crack rock deformation displacement lightweight prediction method and equipment, and belongs to the technical field of intelligent early warning of geotechnical engineering disasters. Constructing a prediction model by adopting a random configuration network method; partial hidden nodes with low contribution degree in the prediction model are eliminated by using a singular value decomposition method, the number of the hidden nodes in the prediction model is reduced, and the prediction model is lightened; excellent new hidden nodes are searched through a random hidden node increasing and replacing technology to replace original hidden nodes in the prediction model, a lightweight prediction model is reconstructed, and the prediction precision is improved; the lightweight prediction model is utilized to predict the deformation displacement of the anchoring crack rock; the method has simple steps, and can accurately predict the deformation displacement of the anchoring crack rock based on the position of the anchor rod and the load level of the anchoring crack rock, thereby ensuring the safety of geotechnical engineering.

Description

Technical Field

[0001] The present invention relates to a lightweight prediction method and device for the deformation displacement of anchored fractured rock, belonging to the technical field of intelligent early warning of geotechnical engineering disasters. Background Technique

[0002] Due to the influence of external loads (such as excavation disturbance, groundwater pressure) and complex geological conditions on the fractured rock mass after anchoring, its deformation displacement still has a high degree of uncertainty: if the displacement exceeds the critical threshold, it may cause anchoring failure and even chain collapse. Therefore, accurately predicting the deformation displacement of the anchored fractured rock mass under different load levels can provide key early warning information for engineering personnel, help optimize the layout of anchor bolts and the reinforcement intensity, and avoid overall instability caused by out-of-control local displacement. This prediction ability is not only the core requirement for ensuring engineering safety but also the key technical foundation for promoting the transformation of geotechnical engineering from "passive emergency rescue" to "active prevention and control".

[0003] The prior art with the publication number CN115392046A discloses a rock deformation prediction method based on the whole-process elastoplastic creep analysis, including the following steps: 1) constructing an expression of the plastic strain of a plastic element related to the time variable; 2) establishing a time-dependent damage function of the rock during the rock creep process based on the probability density function of rock damage; 3) substituting the time-dependent damage function into the expression in step 1) and considering the viscosity coefficient of the damaged plastic element to obtain a constitutive equation of the element stress for plastic deformation; 4) constructing a rock creep stress prediction model reflecting the change of mechanical properties with time t during the whole creep process; 5) using the rock creep stress prediction model to calculate the strain of the rock that undergoes creep deformation at the specified prediction time, and predicting the deformation of the rock; it can only realize the prediction of rock deformation, and cannot obtain information on the deformation situation after driving anchor bolts into the rock. At the same time, this method has high requirements for computing power, and the cost and efficiency are relatively low.

[0004] The prior art with the publication number CN104965027A discloses an analysis method for crack expansion of anchored rock mass based on image recognition and acoustic emission positioning, which belongs to the method of analyzing crack expansion of anchored rock mass. By loading different anchored specimens, on the one hand, the acoustic emission sensor receives the events occurring inside the specimen and the various signals generated, and the acoustic emission ring count of the specimen is processed to obtain the "activity coefficient-time" curve. The curve reflects the different activity periods during the specimen loading process, and the internal positioning is performed through the acoustic emission events, so as to characterize the internal damage and fracture of the specimen; on the other hand, the displacement during the loading process is photographed by a digital camera, and the displacement and strain of the specimen calculation area during the loading process are obtained by calculation, so that the crack expansion process during the specimen loading process can be intuitively judged. The acoustic emission characteristics and deformation characteristics of different anchored specimens under load are revealed, but it can only be predicted by statistical methods, which requires a high number of samples and cannot be predicted for some emergencies. Summary of the invention

[0005] Technical problem: In view of the shortcomings of the existing technology, a lightweight prediction method and device for the deformation and displacement of anchored fractured rocks is provided. By analyzing the compressive deformation data of anchored fractured rocks with noise or abnormal values, a higher prediction speed can be guaranteed for complex anchored fractured compressive deformation data, and it has the advantages of easy implementation and high-precision solution.

[0006] Technical solution: The present invention provides a lightweight prediction method for deformation and displacement of anchored fractured rock, comprising the following steps:

[0007] Collect anchored fractured rock compression test data, pre-process the data, and divide them into training set and test set;

[0008] The prediction model of deformation and displacement of anchored fractured rock with single hidden layer structure is generated by using random collocation network method.

[0009] The singular value decomposition method is used to eliminate redundant hidden nodes in the prediction model whose contribution does not meet the preset requirements, reduce the number of hidden nodes in the prediction model, and lighten the prediction model;

[0010] Design a random replacement hidden node method to find excellent new hidden nodes to replace the original hidden nodes in the lightweight prediction model, reconstruct the lightweight prediction model, and improve the prediction accuracy;

[0011] The prediction model is trained with a training set and a test set, and the predicted deformation displacement of the anchored fractured rock is obtained according to the anchor position and the anchored fractured rock load level parameters using the trained prediction model.

[0012] Further, the compressive strength test data of the anchored fractured rock includes: the position of the anchor bolt measured in the same specimen test, the load level of the anchored fractured rock, and the deformation displacement parameters of the anchored fractured rock. The preprocessing process is as follows: combine the anchor bolt position, the load level of the anchored fractured rock, and the deformation displacement parameters obtained from the experiment into an experimental data set, and randomly divide the experimental data set into two parts: a training set and a test set. The training set data is 70% randomly selected from the entire training data set, and the test set is 30% randomly selected from the entire training data set; add Gaussian noise to the data in the training set and the test set, with a noise level of 0.01, and normalize the input and output samples of the training set and the test set to [0, 1].

[0013] Further, the random configuration network method is used to generate a prediction model for the deformation displacement of the anchored fractured rock with a single hidden layer structure. The specific steps are as follows:

[0014] Let the maximum configurable number of hidden layer nodes of the prediction model be L max ;

[0015] Let the input of the training data set corresponding output The current single hidden layer of the SCN network configuration already has L - 1 hidden nodes, and L - 1 < L max , then the output f L-1 of the current prediction model obtained according to the input X is:

[0016]

[0017] where f0 = 0, indicating that the output of the prediction model is 0 when no hidden nodes are added in the initial state;

[0018] β j = [β j,1 , β j,2 ,.., β j,m T is the output weight of the jth hidden node; w j and b j are the input weight and bias of the jth hidden node respectively; g j () is the Sigmoid function, which is the activation function of the jth hidden node;

[0019] Let e L-1 represent the current network residual, then based on the expected output Y of the training data set, there is:

[0020] e L-1 = Y - f L-1 = [e L-1,1 , ··· e L-1,m T (2)

[0021] ​​The input weight w of the L-th hidden node of the prediction model j and the bias b j need to satisfy the following inequality supervision constraints:

[0022]

[0023] where q = 1, 2,..., m represents the dimension of the training set output; is the output of the hidden layer of the currently newly added hidden node, g() is the Sigmoid function; {μ L} is a sequence of non-negative real numbers satisfying 0 < μ L ≤ (1 - γ) and γ is the inequality constraint coefficient, satisfying 0 < γ < 1;

[0024] Based on the input weight w j and the bias b j obtained from the inequality supervision constraint of formula (3), the least squares method is applied to calculate the output weights of the entire hidden layer:

[0025]

[0026] where, H L = [h1, h2,..., h L , represents the output matrix of L hidden nodes in the hidden layer; represents the Moore-Penrose generalized inverse of H L ;

[0027] If the current network residual e L is greater than the preset tolerance error ε, then the formula (3) and formula (4) are repeatedly used to cycle and continue to configure and add a hidden node to the model, that is, the number of hidden nodes L = L + 1, and the initial value of L is 1; until e L is less than the preset tolerance error ε or the number of hidden nodes L reaches the maximum configurable number of hidden layer node setting values L max .

[0028] Furthermore, the singular value decomposition method is used to remove some hidden nodes with low contribution degrees in the prediction model, reduce the number of hidden nodes in the prediction model, and establish a lightweight prediction model. The specific steps include:

[0029] Set the energy retention ratio threshold and compression ratio ratio in the singular value decomposition method, and calculate the target energy that the prediction model needs to retain and the number of hidden nodes that need to be removed;

[0030] Use formula (5) to perform the economy SVD (compact singular value decomposition) on the output matrix of the hidden layer:

[0031] H L = USV T (5)

[0032] Where N represents the number of samples and L represents the number of nodes; is the left singular matrix, r represents the rank of the output matrix H L Each column in the left singular matrix represents the main distribution pattern of each sample on the node output; is a diagonal matrix that contains the singular values σ L of the H i matrix, indicating the importance of different principal directions; is the transpose of the right singular matrix, and each column represents the contribution weight of the corresponding single node in all principal directions;

[0033] Calculate the sum of the squares of the singular values of the H L matrix using formula (6) to obtain the total energy E total :

[0034]

[0035] And calculate the target energy F according to the energy retention ratio threshold using formula (7) target :

[0036] E target = E total × threshold (7)

[0037] Use formula (8) to find the first k singular values that can satisfy the target energy F target :

[0038]

[0039] Where is the square of the i-th singular value in descending order in the diagonal matrix S, representing the energy contribution of the output matrix H L in the corresponding principal component direction. The higher the energy contribution, the more important the corresponding direction is to the network output. In formula (5), it is known that the total number of singular values in the diagonal matrix S is r. The sum of the squares of the first k singular values in formula (8), that is, the output matrix H L The sum of the energy contributions in the corresponding different principal component directions has satisfied the target energy E target , so the remaining singular values from k + 1 to r are removed as redundant; achieving the effect of retaining the main component directions and removing the secondary component directions, and can effectively remove noise interference without losing key information;

[0040] Based on the principal component direction of the first k singular values, calculate the importance score score of each hidden node in the prediction model j (j = 1, 2, …, K), that is, use score j to represent the comprehensive contribution degree of the importance of each node:

[0041]

[0042] In the formula, is the element in the i-th row and j-th column of the transpose of the right singular vector matrix, representing the projection weight of the j-th node in the i-th principal component direction;

[0043] Sort the importance scores of each node in the prediction model in descending order, and remove the L*ratio redundant hidden nodes with the lowest importance according to the compression ratio ratio, that is, delete the column vectors corresponding to the redundant nodes in the output matrix H L to obtain the reconstructed output matrix as H M :

[0044] M = L·(1 - ratio) (10)

[0045] In the formula, M represents the number of remaining column vectors retained in the original output matrix H L ; the reconstructed output matrix H M then represents the output matrix retaining the hidden nodes with high importance.

[0046] Furthermore, use the method of randomly adding and replacing hidden nodes to find excellent new hidden nodes to replace the original hidden nodes in the prediction model, and reconstruct the lightweight prediction model. The specific steps include:

[0047] Continue to use formulas (3) and (4) to iteratively add hidden nodes until the number of hidden nodes reaches the maximum value L max ;

[0048] If the network residual of the current prediction model does not reach within the preset tolerance error ε, use the random addition and replacement of hidden node technology to optimize the prediction model:

[0049] 1) Randomly remove a hidden node in the prediction model, and use formulas (3) and (4) to add a new hidden node to obtain the network residual of the new prediction model, denoted as

[0050] 2) Compare with : If is smaller, retain the newly added hidden node and update the prediction model; otherwise, restore the removed hidden node and remove the newly added hidden node;

[0051] 3) Repeat steps 1)-2) until it reaches within the preset tolerance error ε.

[0052] Furthermore, the network residual e of the deformation displacement prediction model for anchored fractured rocks L uses the root mean square error as the evaluation index for the training performance of the model. The calculation of the root mean square error is as shown in Equation (11):

[0053]

[0054] In the formula, n represents the number of training samples, represents the output sample value of the model training, and y i represents the initial sample value.

[0055] A computer device includes a processor and a memory. The processor is electrically connected to the memory. The memory is used to store instructions and data, and the processor is used to execute the lightweight prediction method for the deformation displacement of anchored fractured rocks.

[0056] Beneficial effects:

[0057] 1) Compared with the prediction model constructed by the original randomly configured network, the prediction model constructed by the randomly configured network improved based on two technical methods in this method has a very similar prediction error level and good prediction effect. 2) Compared with the prediction model constructed by the randomly configured network improved only by using the singular value decomposition or the random addition and replacement hidden node technical method, the prediction model constructed by the randomly configured network improved based on two technical methods reduces the redundant nodes in the hidden layer, improves the compactness of the model, and has a high prediction accuracy. 3) The prediction model of this method can be used to analyze the compressive deformation data of anchored fractured rocks with noise or outliers. For relatively complex compressive deformation data of anchored fractures, it can ensure a high prediction speed and has the advantages of being easy to implement and high-precision solution. Description of the Drawings

[0058] Figure 1 It is a schematic diagram of the prediction model framework constructed by the improved randomly configured network in the present invention

[0059] Figure 2 It is a graph of the training error RMSE result of the prediction model in the embodiment of the present invention

[0060] Figure 3 It is a graph of the training result of the prediction model in the embodiment of the present invention Detailed Embodiments

[0061] The following further elaborates on the embodiments of the present invention in conjunction with the drawings and tables.

[0062] As Figure 1As shown in the figure, the present invention discloses a lightweight prediction method for the deformation displacement of anchored fractured rocks, and the specific steps are as follows:

[0063] Collect the compressive experiment data of the anchored fractured rocks, combine the bolt position, the load level of the anchored fractured rocks and the compressive deformation parameters of the anchored fractured rocks into an experimental data set, and preprocess the experimental data set;

[0064] Use a random configuration network to generate a prediction model for the deformation displacement of the anchored fractured rocks with a single hidden layer structure, and use the training data set to train the prediction model;

[0065] Use the singular value decomposition method to eliminate the hidden nodes in the prediction model whose contribution degrees do not meet the preset requirements, reduce the number of hidden nodes in the prediction model, and construct a lightweight prediction model;

[0066] Combine the random addition and replacement of hidden nodes technology to add hidden nodes, find excellent new hidden nodes to replace the original hidden nodes in the prediction model, reconstruct the lightweight prediction model, and complete the training of the prediction model;

[0067] After inputting the bolt position and the load level parameter of the anchored fractured rocks into the trained prediction model, obtain the predicted deformation displacement information of the anchored fractured rocks.

[0068] The embodiments of the present invention preprocess the compressive experiment data of the anchored fractured rocks. The specific steps include:

[0069] This simulation experiment runs in the MATLAB R2021b environment. The CPU of the used PC is 2.5 GHz, and the memory is 32 GB RAM. The experimental data comes from the compressive experiment of the anchored fractured rocks. First, obtain the compressive deformation data of the anchored fractured rocks. The input features are the bolt position and the load level of the anchored fractured rocks, and the output feature is the deformation displacement of the anchored fractured rocks. Combine the bolt position, the load level of the anchored fractured rocks and the compressive deformation parameters of the anchored fractured rocks into an experimental data set;

[0070] Divide the data set. Randomly divide the sample data into two parts: a training set and a test set. The training set data is 70% randomly selected from the whole data set, and the test set is 30% randomly selected from the whole data set. There are 9505 groups of data in this data set, so the training set is 6653 groups of data among them, and the test set is 2852 groups of data among them. That is, by training the 6653 groups of data in the training set, to predict the 2852 groups of data in the test set;

[0071] In order to explore the prediction ability of this model under the interference of noise or outliers, it is necessary to add Gaussian noise to the data of the training set and the test set, with a noise level of 0.01, and normalize the input and output samples of the training set and the test set to [0, 1].

[0072] In an embodiment, a prediction model for the deformation displacement of anchored fissured rock with a single hidden layer structure is generated by using a random configuration network. The specific steps include:

[0073] Initialize the parameters required for random configuration network learning. To construct a lightweight prediction model, set the maximum number of hidden layer nodes L max to 80 and the tolerance error ε = 0.0354;

[0074] Let the input of the training data set be and the corresponding output be The single hidden layer of the current SCN network configuration already has L - 1 hidden nodes, and L - 1 < L max , then the output f L-1 of the current prediction model can be obtained according to the input X:

[0075]

[0076] where f0 = 0, indicating that the output of the prediction model is 0 when no hidden nodes are added in the initial state;

[0077] β j = [β j,1 , β j,2 ,.., β j,m T is the output weight of the j-th hidden node; w j and b j are the input weight and bias of the j-th hidden node respectively; g j () is the Sigmoid function, which is the activation function of the j-th hidden node;

[0078] Let e L-1 represent the current network residual, then based on the expected output Y of the training data set, we have:

[0079] e L-1 = Y - f L-1 = [e L-1,1 , ···, e L-1,m T (2)

[0080] The input weight w j and bias b j of the L-th hidden node of the prediction model need to satisfy the following inequality supervision constraints:

[0081]

[0082] In the formula, q = 1, 2,..., m represents the dimension of the training set output; is the output of the hidden layer of the currently added hidden node, and g() is the Sigmoid function; {μ​​L} is a sequence of non - negative real numbers, satisfying 0 < μ L ≤(1 - γ) and γ is the inequality constraint coefficient, satisfying 0 < γ < 1;

[0083] Based on the input weights w j and bias b j obtained from the inequality supervision constraint of formula (3), the output weights of the entire hidden layer are calculated using the least - squares method:

[0084]

[0085] In the formula, H L =[h1, h2, …, h L , representing the output matrix of L hidden nodes in the hidden layer; represents the Moore - Penrose generalized inverse of H L ;

[0086] If the current network residual e L is greater than the preset tolerance error ε = 0.0354, then the formula (3) and formula (4) are repeatedly used to cycle and add a hidden node to the model, that is, the number of hidden nodes L = L + 1, and the initial value of L is 1; until e L is less than the preset tolerance error ε = 0.0354 or the number of hidden nodes L reaches the maximum configurable number of hidden layer nodes L max = 80.

[0087] In the embodiment, the singular value decomposition method is used to remove some hidden nodes with low contribution degrees in the prediction model, reduce the number of hidden nodes in the prediction model, and establish a lightweight prediction model. The specific steps include:

[0088] Set the energy retention ratio threshold = 0.999 and the compression ratio ratio = 0.1 in the singular value decomposition method, which are used to calculate the target energy that the prediction model needs to retain and the number of hidden nodes that need to be removed;

[0089] Use formula (5) to perform the economy SVD (compact singular value decomposition) on the output matrix of the hidden layer:

[0090] H L =USV T (5)

[0091] In the formula, N represents the number of samples, and L represents the number of nodes; is the left singular matrix, r represents the rank of the output matrix H L , and each column in the left singular matrix represents the main distribution pattern of each sample on the node output; is a diagonal matrix that contains H L The singular values σ of the matrix i , indicating the importance of different principal directions; is the transpose of the right singular matrix, and each column represents the contribution weight of the corresponding single node in all principal directions;

[0092] Calculate H using formula (6) L The sum of the squares of the singular values of the matrix to obtain the total energy E total :

[0093]

[0094] And use formula (7) to calculate the target energy E according to the energy retention ratio threshold target :

[0095] E target = E total × threshold (7)

[0096] Use formula (8) to find the first k singular values that can satisfy the target energy E target :

[0097]

[0098] where is the square of the i-th singular value in descending order in the diagonal matrix S, representing the energy contribution of the output matrix H L in the corresponding principal component direction. The higher the energy contribution, the more important the corresponding direction is to the network output; in formula (5), it is known that the total number of singular values in the diagonal matrix S is r. The sum of the squares of the first k singular values in formula (8), that is, the output matrix H L The sum of the energy contributions in the corresponding different principal component directions has satisfied the target energy E target , so the remaining k + 1 to r singular values are regarded as redundant and removed; this step can effectively retain the main component directions and eliminate the secondary component directions, and can effectively remove noise interference without losing key information;

[0099] Based on the principal component directions of the first k singular values, calculate the importance score score of each hidden node in the prediction model j (j = 1, 2,..., L), that is, use score j to represent the comprehensive contribution degree of the importance of each node:

[0100]

[0101] where It is the element in the \(i\)-th row and \(j\)-th column of the transpose of the right singular vector matrix, representing the projection weight of the \(j\)-th node in the direction of the \(i\)-th principal component;

[0102] Sort the importance scores of each node in the prediction model in descending order, and remove the \(L\times ratio\) redundant hidden nodes with the lowest importance according to the compression ratio \(ratio\), that is, delete the column vectors corresponding to the redundant nodes in the output matrix \(H\) L to obtain the reconstructed output matrix \(H'\) M :

[0103] \(M = L\cdot(1 - ratio)\) (10)

[0104] In the formula, \(M\) represents the number of remaining column vectors retained in the original output matrix \(H\) L ; the reconstructed output matrix \(H'\) M represents the output matrix retaining the hidden nodes with high importance.

[0105] In the embodiment, the random addition and replacement of hidden nodes technology is used to find excellent hidden nodes to replace the original hidden nodes in the prediction model, and the lightweight prediction model is reconstructed. The specific steps include:

[0106] Continue to use formulas (3) and (4) to iteratively add hidden nodes until the number of hidden nodes reaches the maximum value \(L\) max = 80;

[0107] If the network residual of the current prediction model does not reach the preset tolerance error \(\varepsilon = 0.0354\), the random addition and replacement of hidden nodes technology is used to optimize the model:

[0108] 1) Randomly remove a hidden node in the prediction model, and use formulas (3) and (4) to add a new hidden node to obtain the network residual of the new prediction model, denoted as

[0109] 2) Compare with : If is smaller, retain the newly added hidden node and update the prediction model; otherwise, restore the removed hidden node and remove the newly added hidden node;

[0110] 3) Repeat steps 1)-2) until reaches the preset tolerance error \(\varepsilon = 0.0354\).

[0111] In the embodiment, the trained prediction model of the deformation displacement of the anchored fissured rock is used for prediction. The specific steps include: inputting the bolt position and the load level parameters of the anchored fissured rock into the trained prediction model to obtain the predicted deformation displacement information of the anchored fissured rock.

[0112] The effects of the present invention are illustrated by the following simulation experiments:

[0113] Experiment content:

[0114] 1) Substitute 6653 sets of training set data with the input features of bolt position and the load level of the anchored fractured rock, and the output feature being the deformation displacement of the anchored fractured rock, into the prediction model for the deformation displacement of the anchored fractured rock provided by the present invention for training, and then predict 2852 sets of test set data.

[0115] 2) Obtain the prediction results of this model, and compare them with the prediction models constructed based on the original SCN network for predicting the deformation displacement of the anchored fractured rock, as well as the prediction models constructed by improving the SCN network using the singular value decomposition and the random addition and replacement hidden node technology methods respectively.

[0116] Evaluation index:

[0117] The network residual e in the prediction model for the deformation displacement of the anchored fractured rock L Use the root mean square error as the evaluation index for the training performance of the model. The calculation of the root mean square error is as shown in Equation (11):

[0118]

[0119] Experimental results:

[0120] Each model algorithm runs independently 30 times on each data set. The simulation result is the mean value obtained from the 30 experimental results. The operation results are presented in a table. The experimental simulation results are shown in Table 1 below. Figure 2 and Figure 3 :

[0121] Table 1 Operation results of three algorithms

[0122]

[0123]

[0124] From the comprehensive comparison results of each model algorithm in Table 1, when the maximum number of hidden layer nodes L max is uniformly set to a fixed value: The prediction model based on Method 1 (random elimination) can have the maximum number of hidden layer nodes as L maxIt can reach within the tolerance error range, but the running time is relatively long; the prediction model based on Method 2 (singular value decomposition) has a relatively short running time, but the error level does not reach within the tolerance error range; while the prediction model constructed based on Method 3 proposed in this paper can not only reach within the tolerance error range, but also has a shorter running time than Method 1. In contrast, although the original prediction model can reach the tolerance error, it requires a larger number of hidden nodes. To sum up, the method proposed in this paper constructs a compact prediction model with satisfactory performance in a shorter time.

Claims

1. A lightweight prediction method for the deformation displacement of anchored fractured rock, characterized in that, The steps are as follows: Collect the compressive strength test data of anchored fractured rocks, preprocess the compressive strength test data of anchored fractured rocks, and divide it into a training set and a test set; Use the random configuration network method to generate a prediction model for the deformation displacement of anchored fractured rocks with a single hidden layer structure; Use the singular value decomposition method to remove redundant hidden nodes in the prediction model whose contribution degrees do not meet the preset requirements, reduce the number of hidden nodes in the prediction model, and lightweight the prediction model; Design a random addition and replacement hidden node method to find excellent new hidden nodes to replace the original hidden nodes in the lightweight prediction model, reconstruct the lightweight prediction model, and improve the prediction accuracy; Train the prediction model with the training set and the test set, and use the trained prediction model to obtain the predicted deformation displacement of the anchored fractured rocks according to the bolt position and the load level parameters of the anchored fractured rocks.

2. The lightweight prediction method for the deformation displacement of the anchored fractured rock according to claim 1, characterized in that The compressive strength test data of the anchored fractured rocks includes: the bolt position measured in the same specimen test, the load level of the anchored fractured rocks, and the deformation displacement parameters of the anchored fractured rocks. The preprocessing process is as follows: combine the bolt position, the load level of the anchored fractured rocks, and the deformation displacement parameters obtained from the experiment into an experimental data set, and randomly divide the experimental data set into two parts: a training set and a test set. The training set data is 70% randomly selected from the full training data set, and the test set is 30% randomly selected from the full training data set; add Gaussian noise to the data in the training set and the test set, with a noise level of 0.01, and normalize the input and output samples of the training set and the test set to [0, 1].

3. The lightweight prediction method for the deformation displacement of the anchored fractured rock according to claim 1, characterized in that: Use the random configuration network method to generate a prediction model for the deformation displacement of anchored fractured rocks with a single hidden layer structure. The specific steps include: Let the maximum configurable number of hidden layer nodes of the prediction model be L max ; Let the input of the training data set be \(X = \{x_1, x_2, \ldots, x\) K \}\), and the corresponding output be \(Y = \{y_1, y_2, \ldots, y\) K \}\). The current single hidden layer of the SCN network configuration already has \(L - 1\) hidden nodes, and \(L - 1 < L\) max , then the output \(f\) of the current prediction model L-1 is obtained according to the input \(X\): where f0 = 0, indicating that the output of the prediction model is 0 when no hidden nodes are added in the initial state; β j = [β j,1 , β j,2 , ···, β j,m T is the output weight of the j-th hidden node; w j and b j are respectively the input weight and bias of the j-th hidden node; g j () is the Sigmoid function, which is the activation function of the j-th hidden node;​ Let e L-1 represent the current network residual, then for the expected output Y based on the training data set, we have: e L-1 = Y - f L-1 = [e L-1,1 ,…e L-1,m T (2)​ The input weight w of the L-th hidden node of the prediction model j and the bias b j need to satisfy the following inequality supervision constraints: where \(q = 1, 2, \ldots, m\) represents the dimension of the training set output; is the output of the hidden layer of the currently newly added hidden node, \(g()\) is the Sigmoid function; \(\{\mu\) L}\) is a sequence of non - negative real numbers satisfying \(0 < \mu\) L \(\leq (1 - \gamma)\) and \(\gamma\) is the inequality constraint coefficient, satisfying \(0 < \gamma < 1\); The input weights w obtained based on the inequality supervision constraint of formula (3) j and the bias b j , use the least squares method to calculate the output weights of the entire hidden layer: where, H L = [h1, h2, …, h L , represents the output matrix of L hidden nodes in the hidden layer; represents the Moore - Penrose generalized inverse of H L ; If the current network residual e L is greater than the preset tolerance error ε, then the formulas (3) and (4) are repeatedly used to loop and continue to add a hidden node to the model, that is, the number of hidden nodes L = L + 1, and the initial value of L is 1; until e L is less than the preset tolerance error ε or the number of hidden nodes L reaches the maximum configurable number of hidden layer node setting value L max .

4. The lightweight prediction method for the deformation displacement of the anchored fractured rock according to claim 3, characterized in that: Use the singular value decomposition method to remove some hidden nodes with low contribution degrees in the prediction model, reduce the number of hidden nodes in the prediction model, and establish a lightweight prediction model. The specific steps include: Set the energy retention ratio threshold and the compression ratio ratio in the singular value decomposition method, and calculate the target energy to be retained and the number of hidden nodes to be removed in the prediction model; Use formula (5) for the output matrix of the hidden layer to perform the Economy SVD (Compact Singular Value Decomposition): H L = USV T (5) Where N represents the number of samples and L represents the number of nodes; is the left singular matrix, and r represents the rank of the output matrix H L The rank of, and each column in the left singular matrix represents the main distribution pattern of each sample on the node output; is a diagonal matrix that contains the singular values σ L of the H i matrix, indicating the importance of different principal directions; is the transpose of the right singular matrix, and each column represents the contribution weight of the corresponding single node in all principal directions; Calculate H using formula (6) L The sum of the squares of the singular values of the matrix to obtain the total energy E total : And calculate the target energy E according to the energy retention ratio threshold using formula (7). target : E target = E total × threshold(7) Use formula (8) to find the top k singular values that can satisfy the target energy E target : In the formula, is the square of the i-th singular value after descending order sorting in the diagonal matrix S, representing the energy contribution of the output matrix H L in the corresponding principal component direction. The higher the energy contribution, the more important the corresponding direction is to the network output. In the formula (5), it is known that the total number of singular values in the diagonal matrix S is r. The sum of the squares of the first k singular values in the formula (8), that is, the output matrix H L the sum of the energy contributions in the corresponding different principal component directions has satisfied the target energy E target , so the remaining singular values from k + 1 to r are removed as redundant; achieving the effect of retaining the main component directions and eliminating the secondary component directions, which can effectively remove noise interference without losing key information; Based on the principal component directions of the top k singular values, calculate the importance score score of each hidden node in the prediction model j (j = 1, 2, …, L), that is, use score j to represent the comprehensive contribution degree of the importance of each node: In the formula, is the element in the \(i\)-th row and \(j\)-th column of the transpose of the right singular vector matrix, representing the projection weight of the \(j\)-th node in the \(i\)-th principal component direction; Sort the importance scores of each node in the prediction model in descending order, and remove the L * ratio redundant hidden nodes with the lowest importance according to the compression ratio ratio, that is, delete the column vectors corresponding to the redundant nodes in the output matrix H L to obtain the reconstructed output matrix as H M : M = L·(1 - ratio) (10) Where M represents the number of remaining column vectors retained in the original output matrix H L ; the reconstructed output matrix H M represents the output matrix retaining the hidden nodes with high importance 5. The lightweight prediction method for the deformation displacement of anchored fractured rock according to claim 4, characterized in that: Use the random addition and replacement hidden node method to find excellent new hidden nodes to replace the original hidden nodes in the prediction model, and reconstruct the lightweight prediction model. The specific steps include: Continue to iteratively add hidden nodes using Equation (3) and Equation (4) until the number of hidden nodes reaches the maximum value L max ; If the network residual of the current prediction model does not reach within the preset tolerance error ε, the random addition and replacement hidden node technology is used to optimize the prediction model: 1) Randomly remove a hidden node from the prediction model, and add a new hidden node using formulas (3) and (4) to obtain the network residual of the new prediction model, denoted as 2) Comparison and to determine their magnitudes: If is smaller, retain the newly added hidden nodes and update the prediction model; otherwise, restore the removed hidden nodes and remove the newly added hidden nodes; 3) Repeat steps 1)-2) until within the preset tolerance error ε.

6. The lightweight prediction method for the deformation displacement of anchored fractured rock according to claim 5, characterized in that: Network residual e of the deformation displacement prediction model for anchored fractured rock L The root mean square error is used as the evaluation index for the training performance of the model. The calculation of the root mean square error is shown in Equation (11): Where n represents the number of training samples, represents the output sample value of model training, and y i represents the initial sample value.

7. A computer device, characterized in that, It includes a processor and a memory. The processor is electrically connected to the memory. The memory is used to store instructions and data, and the processor is used to execute the lightweight prediction method for the deformation displacement of the anchored fractured rocks according to any one of claims 1-6.

Images