A method for predicting surrounding rock grade during TBM excavation

Abstract

The present invention relates to the field of TBM technology, and in particular to a method for predicting surrounding rock grade during TBM excavation, comprising: obtaining TBM data of an excavation cycle of the TBM excavation process; inputting the TBM data of the excavation cycle into a trained LightGBM model to obtain a prediction result; pre-training a preset LightGBM model using a final balanced data set to obtain a trained LightGBM model; the final balanced data set is obtained by merging an original sample data set and a new sample data set; processing the original sample data set using a pre-set new sample generation strategy to obtain a new sample data set; the original sample data set includes multiple samples; each sample includes TBM data of an excavation cycle and a label of the surrounding rock grade corresponding to the TBM data of the excavation cycle.

Description

Technical Field

[0001] The present invention relates to the technical field of TBM, and in particular to a method for predicting surrounding rock grade during TBM excavation. Background Art

[0002] In the geological classification system widely used in China, rock mass is systematically divided into five grades, from Class I to Class V, based on engineering geological characteristics. This classification is of great significance in the construction process of tunnel boring machines (TBMs), because real-time and accurate assessment of rock mass grade can not only ensure the safety of the construction process, but also effectively improve construction efficiency. However, the geological conditions in TBM tunnel construction usually show complex and changeable characteristics, including differences in fracture geometry, changes in rock mass mechanical properties, and the complexity of water-rock interaction. The combined effect of these factors makes the real-time prediction of surrounding rock grade a very challenging task.

[0003] In recent years, with the continuous development of data-driven technology, the prediction of surrounding rock grade in TBM tunnel construction using machine learning algorithms has gradually become a research hotspot. These algorithms are able to handle complex nonlinear relationships and provide high prediction accuracy. However, existing prediction methods still have many shortcomings, especially in the imbalanced distribution of TBM data. Specifically, Class III rock mass dominates the TBM dataset, accounting for more than half of the total, while samples of minority categories such as Class V rock mass are seriously insufficient. This imbalance leads to poor prediction performance of traditional prediction models for minority categories such as Class V. In some studies, the prediction inaccuracy rate of Class V rock mass is as high as 69%.

[0004] Traditional prediction models usually aim to minimize the overall error. Although this method can perform well in the prediction of the majority class (such as Class III), it sacrifices the prediction performance of the minority class, resulting in unsatisfactory prediction results for small classes such as Class II and Class V. In addition, the processing methods for minority class samples in existing studies are usually simple and have limited effects, which are difficult to fully adapt to the complex and changeable geological environment during TBM tunnel construction, further limiting the practical application effect of the model. Summary of the invention

[0005] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a method for predicting surrounding rock grade during TBM excavation, which solves the technical problem of insufficient prediction performance caused by imbalance of surrounding rock grade data.

[0006] In order to achieve the above object, the main technical solutions adopted by the present invention include:

[0007] The embodiment of the present invention provides a method for predicting surrounding rock grade during TBM excavation, comprising:

[0008] S1, obtaining TBM data of one excavation cycle of the TBM excavation process;

[0009] S2, inputting the TBM data of the excavation cycle into the trained LightGBM model to obtain a prediction result;

[0010] The prediction results include: the surrounding rock grade corresponding to the TBM excavation process and the TBM data of the excavation cycle;

[0011] Among them, the preset LightGBM model is trained in advance using the final balanced data set to obtain the trained LightGBM model;

[0012] The final balanced data set is obtained by combining the original sample data set and the new sample data set;

[0013] Wherein, the original sample data set is processed using a preset new sample generation strategy to obtain a new sample data set;

[0014] The original sample data set includes multiple samples; each sample includes TBM data of a tunneling cycle and a label of the surrounding rock grade corresponding to the TBM data of the tunneling cycle; the TBM data of the tunneling cycle includes multiple features.

[0015] Preferably, the process of processing the original sample data set using a preset new sample generation strategy to obtain a new sample data set specifically includes:

[0016] A1. Copy the original sample data set to generate the same initial sample data set;

[0017] A2. According to the labels of the surrounding rock grades corresponding to the TBM data of each excavation cycle, the initial sample data set is divided into two disjoint first subsets and second subsets;

[0018] The first subset includes TBM data of tunneling cycles corresponding to the labels of surrounding rock grades of Class I, Class II, Class IV, and Class V;

[0019] The second subset includes TBM data of the excavation cycle whose corresponding surrounding rock grade label is a Class III surrounding rock grade label;

[0020] A3. For each sample in the first subset, obtain the local density of the sample;

[0021] A4. Based on the local density of the sample, obtain the number of nearest neighbors corresponding to the sample;

[0022] A5. Based on the number of nearest neighbors corresponding to the sample, obtain the nearest neighbor index of the sample;

[0023] A6. Based on the local density of each sample, obtain the weighted median corresponding to each feature in each nested subset of the first subset;

[0024] The nested subsets of the first subset include: a first nested subset, a second nested subset, a third nested subset, and a fourth nested subset;

[0025] The first nested subset includes: TBM data of the tunneling cycle whose corresponding surrounding rock grade label in the first subset is a label of the first type of surrounding rock grade;

[0026] The second nested subset includes: TBM data of the tunneling cycle whose corresponding surrounding rock grade label in the first subset is a label of Class II surrounding rock grade;

[0027] The third nested subset includes: TBM data of the tunneling cycle whose corresponding surrounding rock grade label in the first subset is a label of the Class IV surrounding rock grade;

[0028] The fourth nested subset includes: TBM data of the tunneling cycle whose corresponding surrounding rock grade label in the first subset is a label of the Vth surrounding rock grade;

[0029] A7. For the weighted median corresponding to each feature in each nested subset, obtain the mixed distance of each sample in the nested subset respectively;

[0030] A8. Obtain distribution weights according to the mixing distance of each sample in the nested subset, and normalize the distribution weights to obtain normalized distribution weight values;

[0031] A9. For each sample, based on the nearest neighbor index of the sample, construct a simplex containing the sample and its neighbors;

[0032] A10, generating multiple new samples for the simplex and normalized distribution weight value corresponding to each sample;

[0033] A11. Perform a perturbation on each feature of each new sample to obtain a new sample after the perturbation, and form a new sample data set from all the new samples after the perturbation.

[0034] Preferably,

[0035] The A3 specifically includes: for each sample in the first subset, using formula (1) to obtain the local density of the sample;

[0036] The formula (1) is:

[0037] ;

[0038] in, is the kernel function;

[0039] h is bandwidth;

[0040] n is the total number of samples in the first subset;

[0041] x i is the i-th sample of the local density to be calculated in the first subset;

[0042] x j is the jth sample in the first subset;

[0043] D i is the local density corresponding to the i-th sample in the first subset;

[0044] A4 specifically includes: based on the local density of the sample, using formula (2) to obtain the number of nearest neighbors corresponding to the sample;

[0045] The formula (2) is:

[0046] ;

[0047] in, is the number of nearest neighbors corresponding to the i-th sample in the first subset;

[0048] is the preset maximum number of nearest neighbors;

[0049] is the preset minimum number of nearest neighbors;

[0050] is the maximum local density among the local densities of all samples in the first subset;

[0051] The A5 specifically includes: based on the number of nearest neighbors corresponding to the sample, using a K nearest neighbor algorithm to determine the nearest neighbor index of the sample.

[0052] Preferably, A6 specifically includes: based on the local density of each sample, using formula (3) to obtain the density weight of the sample in the nested subset to which it belongs;

[0053] The formula (3) is:

[0054] w z= the local density of the zth sample in the nested subset / the sum of the local densities of all samples in the nested subset;

[0055] in, w z is the density weight of the zth sample in the nested subset;

[0056] According to the density weight of each sample in the nested subset, the weighted median corresponding to each feature in each nested subset is obtained using formula (4);

[0057] Wherein, the formula (4) is:

[0058] ;

[0059] Represents the weighted median corresponding to the wth feature of the sample in the nested subset

[0060] It means finding the objective function Get the minimum value of x;

[0061] in, ;

[0062] x is a free variable;

[0063] is the wth feature of the zth sample in the nested subset;

[0064] Z is the total number of samples in this nested subset;

[0065] A7 specifically includes: for each feature in each nested subset, the weighted median corresponding to each feature is used to obtain the mixed distance of each sample using formula (5); wherein the formula (5) is:

[0066] ;

[0067] in, a is a pre-set mixing parameter, and a∈[0,1];

[0068] is the mixing distance of the zth sample in the nested subset;

[0069] d is the total number of features corresponding to the sample;

[0070] A8 specifically includes: obtaining corresponding distribution weights using formula (6) according to the mixing distance of each sample in the nested subset;

[0071] The formula (6) is:

[0072] The distribution weight corresponding to the zth sample in the nested subset = the mixed distance of the zth sample in the nested subset / the sum of the mixed distances of all samples in the nested subset.

[0073] Preferably,

[0074] A9 specifically includes: for each sample, based on the nearest neighbor index of the sample, determining the neighbor corresponding to the nearest neighbor index of the sample, and constructing a simplex including the sample and its neighbors;

[0075] The A10 specifically includes: for each sample simplex, according to its normalized distribution weight value, and proportionally calculating the number of generated samples, and generating a corresponding number of new samples;

[0076] The label of the surrounding rock grade in the sample is the same as the labels of the surrounding rock grade of multiple new samples corresponding to the sample.

[0077] Preferably,

[0078] A11 specifically includes: using formula (7) to perturb each feature of each new sample to obtain a new sample after perturbation;

[0079] The formula (7) includes:

[0080] ;

[0081] is a value randomly drawn from a normal distribution with a mean of 0 and a standard deviation of 1;

[0082] is the wth feature of the zth sample in the nested subset Characteristics after perturbation;

[0083] ;

[0084] K is the number of nearest neighbors corresponding to the zth sample in the nested subset;

[0085] y kw is the wth feature of the kth neighbor corresponding to the zth sample in the nested subset;

[0086] .

[0087] Preferably,

[0088] The trained LightGBM model refers to the model obtained by tuning the improved whale optimization algorithm in the process of training the preset LightGBM model using the final balanced data set.

[0089] Preferably,

[0090] In the improved whale optimization algorithm, when the behavior switching parameter r1 is less than 0.5, the alternating attack phase and prey capture phase in the sailfish optimization algorithm are used to update the position of the individual representing the hyperparameters in the LightGBM model;

[0091] Among them, in the alternating attack phase of the sailfish optimization algorithm, formula (8) is used to update the individual position;

[0092] The formula (8) is:

[0093] ;

[0094] in, is the current global optimal solution;

[0095] is a randomly selected candidate location;

[0096] ;

[0097] rand(0,1) is a random number with a mean of 0 and a variance of 1;

[0098] is the position of the individual at the t+1th iteration;

[0099] X t is the position of the individual at the tth iteration;

[0100] ;

[0101] in, Represents the number of sailfish in the sailfish optimization algorithm; Represents the number of sardines in the sailfish optimization algorithm.

[0102] Preferably,

[0103] Among them, in the prey capture stage of the swordfish optimization algorithm, formula (9) is used to update the individual position;

[0104] The formula (9) is:

[0105] ;

[0106] Where r is a random factor, and r∈[0,1];

[0107] ;

[0108] A is the initial attack amplitude;

[0109] t is the current iteration number;

[0110] t max is the maximum number of iterations.

[0111] Preferably,

[0112] In the improved whale optimization algorithm, when the behavior switching parameter r1 is greater than or equal to 0.5 and when the preset random factor r2 is less than 0.4, the position of the individual representing the hyperparameter in the LightGBM model is updated using formula (10);

[0113] The formula (10) is:

[0114] ;

[0115] in, ;

[0116] b is the parameter controlling the spiral contraction amplitude;

[0117] l is a random number in [-1, 1];

[0118] In the improved whale optimization algorithm, when the behavior switching parameter r1 is greater than or equal to 0.5, and the random factor r2 is greater than or equal to 0.4 and less than 0.8, the position of the individual representing the hyperparameters in the LightGBM model is updated using formula (11);

[0119] The formula (11) is:

[0120] ;

[0121] In the improved whale optimization algorithm, when the behavior switching parameter r1 is greater than or equal to 0.5 and the random factor r2 is greater than or equal to 0.8, the position of the individual representing the hyperparameters in the LightGBM model is updated using formula (12);

[0122] The formula (12) is:

[0123] ;

[0124] is the position of the individual at the t-1th iteration;

[0125] ;

[0126] .

[0127] The beneficial effects of the present invention are as follows: a method for predicting surrounding rock grade during TBM excavation of the present invention achieves category balance of the final data set by generating a new sample data set and combining it with the original sample data set. After the number of minority category samples increases, the model can more accurately learn the features related to the minority category, thereby improving the overall prediction performance and overcoming the shortcoming of low prediction accuracy of traditional methods when dealing with minority categories (such as Class II and Class V surrounding rock grades).

[0128] In addition, the method for predicting surrounding rock grade during TBM excavation of the present invention, the LightGBM model has efficient training speed, good feature processing ability and adaptability to category imbalance, and can provide higher prediction accuracy under complex nonlinear feature conditions. By training the LightGBM model of the final balanced data set, the model deviation caused by uneven data distribution is reduced, and the adaptability to complex geological environments is enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0129] Figure 1 The present invention is a flow chart of a method for predicting surrounding rock grade during TBM excavation. DETAILED DESCRIPTION

[0130] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation modes in conjunction with the accompanying drawings.

[0131] In order to better understand the above technical solution, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a clearer and more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0132] See also Figure 1 This embodiment provides a method for predicting surrounding rock grade during TBM excavation, including:

[0133] S1. Acquire TBM data of one excavation cycle of the TBM excavation process;

[0134] S2, inputting the TBM data of the excavation cycle into the trained LightGBM model to obtain a prediction result;

[0135] The prediction results include: the surrounding rock grade corresponding to the TBM excavation process and the TBM data of the excavation cycle;

[0136] Among them, the preset LightGBM model is trained in advance using the final balanced data set to obtain the trained LightGBM model;

[0137] The final balanced data set is obtained by combining the original sample data set and the new sample data set;

[0138] Wherein, the original sample data set is processed using a preset new sample generation strategy to obtain a new sample data set;

[0139] The original sample data set includes multiple samples; each sample includes TBM data of a tunneling cycle and a label of the surrounding rock grade corresponding to the TBM data of the tunneling cycle; the TBM data of the tunneling cycle includes multiple features.

[0140] In this embodiment, during the TBM excavation process, an excavation cycle refers to the time period from the start of excavation to the completion of a complete excavation cycle. In this process, the TBM data records multiple key feature information reflecting the excavation status and rock mass characteristics, and these data are the basis for the prediction of the surrounding rock grade. The TBM data of the excavation cycle consists of multiple features, such as common features including: Cutterhead speed (RPM): the number of rotations of the cutterhead per minute, reflecting the excavation speed. Cutterhead torque (Torque): the rotational resistance of the cutterhead, which is significantly affected by the strength of the surrounding rock. Thrust force: the force acting on the cutterhead during the TBM advancement process, which is usually related to the hardness and degree of crushing of the surrounding rock. Penetration rate: the depth of excavation per minute, reflecting the machinability of the rock. The sections of the TBM excavation process are divided into ascending section, stable section, and descending section; ascending section data: refers to the stage where the cutterhead accelerates to cut the rock, usually including the initial loading characteristics. Stable section data: refers to the steady advancement stage when the cutterhead cuts the rock, which best reflects the rock properties. Descent data (if recorded): The deceleration phase before the end of tunnelling.

[0141] In this embodiment, LightGBM is a gradient boosting algorithm based on a decision tree. It has significant computational efficiency when processing large-scale data sets and high-dimensional data, and can quickly adapt to the large amount of real-time data generated during TBM excavation. In actual scenarios, there may be a class imbalance in the surrounding rock grade labels (for example, the difference in the distribution ratio of soft rock and hard rock). By adopting a new sample generation strategy and increasing the number of minority class samples, it helps the model avoid bias towards the majority class when predicting, thereby improving classification accuracy. The introduction of new sample data sets expands the diversity of training data, helps the model better adapt to complex and changeable excavation conditions, and improves the robustness of the model.

[0142] Specifically, the process of processing the original sample data set using a preset new sample generation strategy to obtain a new sample data set specifically includes:

[0143] A1. Copy the original sample data set to generate the same initial sample data set; copying the original data set as the initial sample data set ensures the retention of all original information and ensures that data is not lost in subsequent processing. By retaining the original data set, the correctness of data processing can be traced back and verified at any time in the subsequent process, avoiding the risks brought by direct modification of the original data.

[0144] A2. According to the labels of the surrounding rock grades corresponding to the TBM data of each excavation cycle, the initial sample data set is divided into two disjoint first subsets and second subsets;

[0145] The first subset includes TBM data of tunneling cycles corresponding to the labels of surrounding rock grades of Class I, Class II, Class IV, and Class V;

[0146] The second subset includes TBM data of the excavation cycle whose corresponding surrounding rock grade label is a Class III surrounding rock grade label;

[0147] In this embodiment, the samples are divided into different subsets according to the surrounding rock grade labels, so that different categories of data can be analyzed and processed separately more clearly, thereby ensuring the distinction between categories in the subsequent analysis process. By separating the samples with specific surrounding rock grade labels, the imbalance of categories can be effectively handled, especially for the optimization of some categories with less common surrounding rock grades.

[0148] A3. For each sample in the first subset, obtain the local density of the sample;

[0149] A4. Based on the local density of the sample, obtain the number of nearest neighbors corresponding to the sample; by determining the number of nearest neighbors based on the local density, it can help capture samples with similar characteristics and ensure full use of similar data during training. This step helps to understand the data structure around a sample, helps to maintain consistency and connection between samples in subsequent calculations, and improves the correlation between data.

[0150] A5. Based on the number of nearest neighbors corresponding to the sample, obtain the nearest neighbor index of the sample;

[0151] A6. Based on the local density of each sample, obtain the weighted median corresponding to each feature in each nested subset of the first subset;

[0152] The weighted median in this embodiment is more robust to outliers than the mean, and can more accurately describe the central tendency of the sample group and avoid being affected by outliers.

[0153] The nested subsets of the first subset include: a first nested subset, a second nested subset, a third nested subset, and a fourth nested subset;

[0154] The first nested subset includes: TBM data of the tunneling cycle whose corresponding surrounding rock grade label in the first subset is a label of the first type of surrounding rock grade;

[0155] The second nested subset includes: TBM data of the tunneling cycle whose corresponding surrounding rock grade label in the first subset is a label of Class II surrounding rock grade;

[0156] The third nested subset includes: TBM data of the tunneling cycle whose corresponding surrounding rock grade label in the first subset is a label of the Class IV surrounding rock grade;

[0157] The fourth nested subset includes: TBM data of the tunneling cycle whose corresponding surrounding rock grade label in the first subset is a label of the Vth surrounding rock grade;

[0158] A7. For the weighted median corresponding to each feature in each nested subset, obtain the mixed distance of each sample in the nested subset respectively;

[0159] A8. Obtain distribution weights according to the mixing distance of each sample in the nested subset, and normalize the distribution weights to obtain normalized distribution weight values;

[0160] In this embodiment, the normalized distribution weights ensure that during the sample generation process, samples with important features will have a greater impact on the new samples that are finally generated, thereby increasing the model's focus on key features. The normalization process allows the weights of all samples to be compared and adjusted on the same scale, avoiding the adverse effects of excessively large or small weight values ​​on the generated samples.

[0161] A9. For each sample, based on the nearest neighbor index of the sample, construct a simplex containing the sample and its neighbors;

[0162] Simplex construction can geometrically represent the relationship between each sample and its neighbors, so as to better understand the relationship between data points and improve the expressiveness of data structure. By constructing simplex, it can ensure that the information between neighbors is effectively integrated during the sample generation process, and improve the utilization of local information when generating samples.

[0163] A10, generating multiple new samples for the simplex and normalized distribution weight value corresponding to each sample;

[0164] By generating new samples, the training data set can be expanded, especially in areas where data is scarce, which helps avoid model overfitting and improves generalization capabilities. The generated new samples are closer to the original samples in terms of features, and the diversity of the data is enhanced through perturbations, allowing the model to learn more potential laws and improve its adaptability to complex tunneling environments.

[0165] A11. Perform a perturbation on each feature of each new sample to obtain a new sample after the perturbation, and form a new sample data set from all the new samples after the perturbation.

[0166] In the practical application of this embodiment, A3 specifically includes: for each sample in the first subset, using formula (1) to obtain the local density of the sample;

[0167] The formula (1) is:

[0168] ;

[0169] in, is the kernel function;

[0170] h is bandwidth;

[0171] n is the total number of samples in the first subset;

[0172] x i is the i-th sample of the local density to be calculated in the first subset;

[0173] x j is the jth sample in the first subset;

[0174] D i is the local density corresponding to the i-th sample in the first subset;

[0175] A4 specifically includes: based on the local density of the sample, using formula (2) to obtain the number of nearest neighbors corresponding to the sample;

[0176] The formula (2) is:

[0177] ;

[0178] in, is the number of nearest neighbors corresponding to the i-th sample in the first subset;

[0179] is the preset maximum number of nearest neighbors;

[0180] is the preset minimum number of nearest neighbors;

[0181] is the maximum local density among the local densities of all samples in the first subset;

[0182] A5 specifically includes: based on the number of nearest neighbors corresponding to the sample, using the K nearest neighbor algorithm to determine the nearest neighbor index of the sample. In this embodiment, once K nearest neighbor samples are selected, the positions of these samples in the original data set can be recorded, and these positions are the nearest neighbor indexes.

[0183] A6 specifically includes: based on the local density of each sample, using formula (3) to obtain the density weight of the sample in the nested subset to which it belongs;

[0184] The formula (3) is:

[0185] w z = the local density of the zth sample in the nested subset / the sum of the local densities of all samples in the nested subset;

[0186] in, w z is the density weight of the zth sample in the nested subset;

[0187] According to the density weight of each sample in the nested subset, the weighted median corresponding to each feature in each nested subset is obtained using formula (4);

[0188] Wherein, the formula (4) is:

[0189] ;

[0190] Represents the weighted median corresponding to the wth feature of the sample in the nested subset;

[0191] It means finding the objective function Get the minimum value of x;

[0192] in, ;

[0193] x is a free variable;

[0194] is the wth feature of the zth sample in the nested subset;

[0195] Z is the total number of samples in this nested subset;

[0196] A7 specifically includes: for each feature in each nested subset, the weighted median corresponding to each feature is used to obtain the mixed distance of each sample using formula (5); wherein the formula (5) is:

[0197] ;

[0198] in, a is a pre-set mixing parameter, and a∈[0,1];

[0199] is the mixing distance of the zth sample in the nested subset;

[0200] d is the total number of features corresponding to the sample;

[0201] A8 specifically includes: obtaining corresponding distribution weights using formula (6) according to the mixing distance of each sample in the nested subset;

[0202] The formula (6) is:

[0203] The distribution weight corresponding to the zth sample in the nested subset = the mixed distance of the zth sample in the nested subset / the sum of the mixed distances of all samples in the nested subset.

[0204] A9 specifically includes: for each sample, based on the nearest neighbor index of the sample, determining the neighbor corresponding to the nearest neighbor index of the sample, and constructing a simplex including the sample and its neighbors;

[0205] By calculating the distance between the target sample and its neighbors, the K nearest neighbor algorithm is used to obtain the nearest neighbor index of the target sample, and then the simplex composed of these neighbors is determined. For example, for each target sample, its nearest neighbors (i.e., samples in the K nearest neighbors) are determined, and a simplex is constructed based on these neighbors and the target sample. For example: for a two-dimensional data point, it may have 3 nearest neighbors, and a triangle is formed by these samples; for higher-dimensional data, the constructed simplex is a higher-dimensional geometric structure. In high-dimensional space, such a simplex helps the model understand the local structure of the sample and describes how these samples are close to each other in the feature space.

[0206] The A10 specifically includes: for each sample simplex, according to its normalized distribution weight value, and proportionally calculating the number of generated samples, and generating a corresponding number of new samples;

[0207] The label of the surrounding rock grade in this sample is the same as the label of the surrounding rock grade of multiple new samples corresponding to this sample. This means that the generated new sample is an "enhancement" of the corresponding sample, maintaining the consistency of the label and ensuring that the new sample is associated with the target task (surrounding rock grade prediction).

[0208] Each sample is given a weight based on its position, neighbor relationship, and local density, and the weight value represents the importance of the sample in generating new samples. The weights of all samples are adjusted proportionally so that their sum is 1. This ensures that when generating new samples, samples with larger weights have a greater impact on the generated samples, while samples with smaller weights have less impact.

[0209] According to the normalized weight value of each sample, the number of new samples generated by the sample is calculated. Samples with higher weights will generate more samples, and samples with lower weights will generate fewer samples. For example, a sample with a weight of 0.5 will generate more samples than a sample with a weight of 0.2. This ensures the diversity of the data and the balance of the samples.

[0210] In this embodiment, A11 specifically includes: using formula (7) to perturb each feature of each new sample to obtain a new sample after perturbation;

[0211] The formula (7) includes:

[0212] ;

[0213] is a value randomly drawn from a normal distribution with a mean of 0 and a standard deviation of 1;

[0214] is the wth feature of the zth sample in the nested subset Characteristics after perturbation;

[0215] ;

[0216] K is the number of nearest neighbors corresponding to the zth sample in the nested subset;

[0217] y kw is the wth feature of the kth neighbor corresponding to the zth sample in the nested subset;

[0218] .

[0219] In practical applications, the trained LightGBM model refers to a model obtained by tuning the preset LightGBM model using the final balanced data set through the improved whale optimization algorithm. In this embodiment, in addition to using the improved whale optimization algorithm for tuning, other existing algorithms can also be used to tune the preset LightGBM model, for example, genetic algorithm, simulated annealing, particle swarm optimization, etc., without specific limitation.

[0220] The original whale optimization algorithm is a metaheuristic optimization algorithm that simulates the foraging behavior of humpback whales and can achieve global optimization in the search space. When hunting, humpback whales capture targets by surrounding prey and creating bubble nets. These behaviors are modeled through mathematical formulas and are divided into random search phase, prey encirclement phase and spiral phase, thereby effectively switching dynamically between global search and local development. The dynamic switching of the above three mechanisms is controlled by the switching parameter r1: when the switching parameter r1 is less than 0.5, the individual chooses search or encirclement (encircling prey) behavior; when the switching parameter r1 is greater than or equal to 0.5, the individual performs spiral behavior. However, in this embodiment, the original whale optimization algorithm is improved to obtain an improved whale optimization algorithm:

[0221] In the improved whale optimization algorithm of this embodiment, when the behavior switching parameter r1 is less than 0.5, the alternating attack phase and prey capture phase in the sailfish optimization algorithm are used to update the position of the individual representing the hyperparameters in the LightGBM model;

[0222] Among them, in the alternating attack phase of the sailfish optimization algorithm, formula (8) is used to update the individual position;

[0223] The formula (8) is:

[0224] ;

[0225] in, is the current global optimal solution;

[0226] is a randomly selected candidate location;

[0227] ;

[0228] rand(0,1) is a random number with a mean of 0 and a variance of 1;

[0229] is the position of the individual at the t+1th iteration;

[0230] X t is the position of the individual at the tth iteration;

[0231] ;

[0232] in, Represents the number of sailfish in the sailfish optimization algorithm; Represents the number of sardines in the sailfish optimization algorithm.

[0233] Specifically, in the prey capture phase of the sailfish optimization algorithm, formula (9) is used to update the individual position;

[0234] The formula (9) is:

[0235] ;

[0236] Where r is a random factor, and r∈[0,1];

[0237] ;

[0238] A is the initial attack amplitude;

[0239] t is the current iteration number;

[0240] t max is the maximum number of iterations.

[0241] Preferably, in the improved whale optimization algorithm, when the behavior switching parameter r1 is greater than or equal to 0.5 and when the preset random factor r2 is less than 0.4, the position of the individual representing the hyperparameter in the LightGBM model is updated using formula (10);

[0242] The formula (10) is:

[0243] ;

[0244] in, ;

[0245] b is the parameter controlling the spiral contraction amplitude;

[0246] l is a random number in [-1, 1];

[0247] In the improved whale optimization algorithm, when the behavior switching parameter r1 is greater than or equal to 0.5, and the random factor r2 is greater than or equal to 0.4 and less than 0.8, the position of the individual representing the hyperparameters in the LightGBM model is updated using formula (11);

[0248] The formula (11) is:

[0249] ;

[0250] In the improved whale optimization algorithm, when the behavior switching parameter r1 is greater than or equal to 0.5 and the random factor r2 is greater than or equal to 0.8, the position of the individual representing the hyperparameters in the LightGBM model is updated using formula (12);

[0251] The formula (12) is:

[0252] ;

[0253] is the position of the individual at the t-1th iteration;

[0254] ;

[0255] .

[0256] In this embodiment, the improved whale optimization algorithm introduces the alternating attack phase and the prey capture phase in the sailfish optimization algorithm, thereby increasing the behavioral diversity of the algorithm and helping to improve the global exploration capability during the search process. By introducing additional conditions (such as the value of the random factor r2), the algorithm can dynamically adjust the behavior according to the current state, thereby enhancing the adaptability and flexibility of the algorithm.

[0257] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0258] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0259] In the present invention, unless otherwise clearly specified and limited, when a first feature is “on” or “below” a second feature, it may be that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. Moreover, when a first feature is “above”, “above” or “above” a second feature, it may be that the first feature is directly above or obliquely above the second feature, or it may simply mean that the first feature is higher in level than the second feature. When a first feature is “below”, “below” or “below” a second feature, it may be that the first feature is directly below or obliquely below the second feature, or it may simply mean that the first feature is lower in level than the second feature.

[0260] In the description of this specification, the description of the terms "one embodiment", "some embodiments", "embodiment", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.

[0261] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may alter, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A method for predicting surrounding rock grade during TBM excavation, characterized in that: include: S1, obtaining TBM data of one excavation cycle of the TBM excavation process; S2, inputting the TBM data of the excavation cycle into the trained LightGBM model to obtain a prediction result; The prediction results include: the surrounding rock grade corresponding to the TBM excavation process and the TBM data of the excavation cycle; Among them, the preset LightGBM model is trained in advance using the final balanced data set to obtain the trained LightGBM model; The final balanced data set is obtained by combining the original sample data set and the new sample data set; Wherein, the original sample data set is processed using a preset new sample generation strategy to obtain a new sample data set; The original sample data set includes a plurality of samples; each sample includes TBM data of a tunneling cycle and a label of a surrounding rock grade corresponding to the TBM data of the tunneling cycle; the TBM data of the tunneling cycle includes a plurality of features; The adopting of a preset new sample generation strategy to process the original sample data set to obtain a new sample data set specifically includes: A1. Copy the original sample data set to generate the same initial sample data set; A2. According to the labels of the surrounding rock grades corresponding to the TBM data of each excavation cycle, the initial sample data set is divided into two disjoint first subsets and second subsets; The first subset includes TBM data of tunneling cycles corresponding to the labels of surrounding rock grades of Class I, Class II, Class IV, and Class V; The second subset includes TBM data of the excavation cycle whose corresponding surrounding rock grade label is a Class III surrounding rock grade label; A3. For each sample in the first subset, obtain the local density of the sample; A4. Based on the local density of the sample, obtain the number of nearest neighbors corresponding to the sample; A5. Based on the number of nearest neighbors corresponding to the sample, obtain the nearest neighbor index of the sample; A6. Based on the local density of each sample, obtain the weighted median corresponding to each feature in each nested subset of the first subset; The nested subsets of the first subset include: a first nested subset, a second nested subset, a third nested subset, and a fourth nested subset; The first nested subset includes: TBM data of the tunneling cycle whose corresponding surrounding rock grade label in the first subset is a label of the first type of surrounding rock grade; The second nested subset includes: TBM data of the tunneling cycle whose corresponding surrounding rock grade label in the first subset is a label of Class II surrounding rock grade; The third nested subset includes: TBM data of the tunneling cycle whose corresponding surrounding rock grade label in the first subset is a label of the Class IV surrounding rock grade; The fourth nested subset includes: TBM data of the tunneling cycle whose corresponding surrounding rock grade label in the first subset is a label of the Vth surrounding rock grade; A7. For the weighted median corresponding to each feature in each nested subset, obtain the mixed distance of each sample in the nested subset respectively; A8. Obtain distribution weights according to the mixing distance of each sample in the nested subset, and normalize the distribution weights to obtain normalized distribution weight values; A9. For each sample, based on the nearest neighbor index of the sample, construct a simplex containing the sample and its neighbors; A10, generating multiple new samples according to the simplex and normalized distribution weight value corresponding to each sample; A11. Perform a perturbation on each feature of each new sample to obtain a new sample after the perturbation, and form a new sample data set from all the new samples after the perturbation.

2. The method for predicting surrounding rock grade during TBM excavation according to claim 1, characterized in that: The A3 specifically includes: for each sample in the first subset, using formula (1) to obtain the local density of the sample; The formula (1) is: ； in, is the kernel function; h is bandwidth; n is the total number of samples in the first subset; x i is the i-th sample of the local density to be calculated in the first subset; x j is the jth sample in the first subset; D i is the local density corresponding to the i-th sample in the first subset; A4 specifically includes: based on the local density of the sample, using formula (2) to obtain the number of nearest neighbors corresponding to the sample; The formula (2) is: ； in, is the number of nearest neighbors corresponding to the i-th sample in the first subset; is the preset maximum number of nearest neighbors; is the preset minimum number of nearest neighbors; is the maximum local density among the local densities of all samples in the first subset; The A5 specifically includes: based on the number of nearest neighbors corresponding to the sample, using a K-nearest neighbor algorithm to determine the nearest neighbor index of the sample.

3. The method for predicting surrounding rock grade during TBM excavation according to claim 2, characterized in that: A6 specifically includes: based on the local density of each sample, using formula (3) to obtain the density weight of the sample in the nested subset to which it belongs; The formula (3) is: w z = the local density of the zth sample in the nested subset / the sum of the local densities of all samples in the nested subset; in, w z is the density weight of the zth sample in the nested subset; According to the density weight of each sample in the nested subset, the weighted median corresponding to each feature in each nested subset is obtained using formula (4); Wherein, the formula (4) is: ； Represents the weighted median corresponding to the wth feature of the sample in the nested subset; It means finding the objective function Get the minimum value of x; in, ; x is a free variable; is the wth feature of the zth sample in the nested subset; Z is the total number of samples in this nested subset; A7 specifically includes: for each feature in each nested subset, the weighted median corresponding to each feature is used to obtain the mixed distance of each sample using formula (5); wherein the formula (5) is: ； in, a is a pre-set mixing parameter, and a∈[0,1]; is the mixing distance of the zth sample in the nested subset; d is the total number of features corresponding to the sample; A8 specifically includes: obtaining corresponding distribution weights using formula (6) according to the mixing distance of each sample in the nested subset; The formula (6) is: The distribution weight corresponding to the zth sample in the nested subset = the mixed distance of the zth sample in the nested subset / the sum of the mixed distances of all samples in the nested subset.

4. The method for predicting surrounding rock grade during TBM excavation according to claim 3, characterized in that: A9 specifically includes: for each sample, based on the nearest neighbor index of the sample, determining the neighbor corresponding to the nearest neighbor index of the sample, and constructing a simplex including the sample and its neighbors; The A10 specifically includes: for each sample simplex, according to its normalized distribution weight value, and proportionally calculating the number of generated samples, and generating a corresponding number of new samples; The label of the surrounding rock grade in the sample is the same as the labels of the surrounding rock grade of multiple new samples corresponding to the sample.

5. The method for predicting surrounding rock grade during TBM excavation according to claim 4, characterized in that: A11 specifically includes: using formula (7) to perturb each feature of each new sample to obtain a new sample after perturbation; The formula (7) includes: ； is a value randomly drawn from a normal distribution with a mean of 0 and a standard deviation of 1; is the wth feature of the zth sample in the nested subset Characteristics after perturbation; ； K is the number of nearest neighbors corresponding to the zth sample in the nested subset; y kw is the wth feature of the kth neighbor corresponding to the zth sample in the nested subset; 。 6. The method for predicting surrounding rock grade during TBM excavation according to claim 5, characterized in that: The trained LightGBM model refers to the model obtained by tuning the improved whale optimization algorithm in the process of training the preset LightGBM model using the final balanced data set.

7. The method for predicting surrounding rock grade during TBM excavation according to claim 6, characterized in that: In the improved whale optimization algorithm, when the behavior switching parameter r1 is less than 0.5, the alternating attack phase and prey capture phase in the sailfish optimization algorithm are used to update the position of the individual representing the hyperparameters in the LightGBM model; Among them, in the alternating attack phase of the sailfish optimization algorithm, formula (8) is used to update the individual position; The formula (8) is: ； in, is the current global optimal solution; is a randomly selected candidate location; ； rand(0,1) is a random number with a mean of 0 and a variance of 1; is the position of the individual at the t+1th iteration; X t is the position of the individual at the tth iteration; ； in, Represents the number of sailfish in the sailfish optimization algorithm; Represents the number of sardines in the sailfish optimization algorithm.

8. The method for predicting surrounding rock grade during TBM excavation according to claim 7, characterized in that: in, In the prey capture phase of the sailfish optimization algorithm, formula (9) is used to update the individual position; The formula (9) is: ； Where r is a random factor, and r∈[0,1]; ； A is the initial attack amplitude; t is the current iteration number; t max is the maximum number of iterations.

9. The method for predicting surrounding rock grade during TBM excavation according to claim 8, characterized in that: In the improved whale optimization algorithm, when the behavior switching parameter r1 is greater than or equal to 0.5 and when the preset random factor r2 is less than 0.4, the position of the individual representing the hyperparameter in the LightGBM model is updated using formula (10); The formula (10) is: ； in, ; b is the parameter controlling the spiral contraction amplitude; l is a random number in [-1, 1]; In the improved whale optimization algorithm, when the behavior switching parameter r1 is greater than or equal to 0.5, and the random factor r2 is greater than or equal to 0.4 and less than 0.8, the position of the individual representing the hyperparameters in the LightGBM model is updated using formula (11); The formula (11) is: ； In the improved whale optimization algorithm, when the behavior switching parameter r1 is greater than or equal to 0.5 and the random factor r2 is greater than or equal to 0.8, the position of the individual representing the hyperparameters in the LightGBM model is updated using formula (12); The formula (12) is: ； is the position of the individual at the t-1th iteration; ； 。

Images