Tunnel surrounding rock classification method and system based on multi-dimensional data

Abstract

This application provides a method and system for classifying tunnel surrounding rock based on multidimensional data. The method includes: obtaining a candidate feature set, an initial tunnel surrounding rock level classifier, and an initial model hyperparameter range for tunnel surrounding rock samples; initializing a lemming population using random sampling based on the candidate feature set and the initial model hyperparameter range to obtain an initial lemming population; using an improved lemming optimization algorithm to perform a feature selection stage and a parameter fine optimization stage based on the lemming population, selecting the candidate feature set and the initial model hyperparameter range for tunnel surrounding rock samples to obtain an optimal surrounding rock feature subset and the corresponding optimal model hyperparameters; constructing a final surrounding rock level identification model based on the optimal surrounding rock feature subset and the optimal model hyperparameters, and using the final surrounding rock level identification model for tunnel surrounding rock classification. This method can adaptively balance the dual requirements of rapid preliminary judgment and accurate identification under complex working conditions, improving the practicality and reliability of the tunnel surrounding rock identification method.

Description

Technical Field

[0001] This application relates to the technical field of surrounding rock classification in underground engineering shield tunneling, specifically to a tunnel surrounding rock classification method and system based on multidimensional data. Background Technology

[0002] In the field of tunnel engineering, accurate and efficient identification of surrounding rock grades is crucial for construction safety and design optimization. Currently, data-driven intelligent classification methods have become the mainstream research direction. These methods typically extract multi-dimensional features from geological exploration data and use machine learning models to establish a mapping relationship between features and surrounding rock grades. To improve model performance, existing tunnel surrounding rock classification methods introduce metaheuristic optimization algorithms. This method models feature selection and parameter tuning as a complex combinatorial optimization problem, reducing reliance on human experience to some extent and automating the identification process.

[0003] In actual tunnel construction scenarios, on the one hand, on-site decision-making requires rock mass classification methods to quickly provide preliminary and reliable classification judgments to guide work process coordination and emergency response; on the other hand, facing complex and ever-changing geological conditions, rock mass classification methods must also possess sufficiently precise identification capabilities to ensure long-term safety. However, existing automated optimization methods for tunnel rock mass classification and identification models typically employ a single, fixed search pattern during the optimization process, failing to meet the dual requirements of balancing rapid preliminary classification judgments with accurate identification of complex working conditions. This results in models with weak generalization ability and insufficient reliability when facing complex geological conditions. Summary of the Invention

[0004] This application provides a tunnel surrounding rock classification method and system based on multidimensional data, which can adaptively balance the dual requirements of rapid preliminary judgment and accurate identification under complex working conditions, thereby improving the practicality and reliability of the tunnel surrounding rock identification method.

[0005] A first aspect of this application provides a tunnel surrounding rock classification method based on multidimensional data, the method comprising:

[0006] Obtain the candidate feature set, initial tunnel surrounding rock level classifier, and initial model hyperparameter range for tunnel surrounding rock samples;

[0007] The lemming population is initialized by random sampling based on the candidate feature set and the range of initial model hyperparameters to obtain the initial lemming population.

[0008] Based on the initial lemming population, an improved lemming optimization algorithm is used to perform a feature screening stage and a parameter fine optimization stage to screen the candidate feature set and the initial model hyperparameter range of the tunnel surrounding rock sample, thereby obtaining the optimal surrounding rock feature subset and the corresponding optimal model hyperparameters.

[0009] Based on the optimal subset of surrounding rock features and the optimal model hyperparameters, a final surrounding rock grade identification model is constructed, and the final surrounding rock grade identification model is used to classify the surrounding rock of tunnels.

[0010] In one possible implementation, the feature selection stage includes:

[0011] The energy factor of the initial lemming population is set to a fixed value greater than the first threshold, so that all lemming individuals in the initial lemming population only perform exploratory behavior to update their location.

[0012] The initial lemming population is updated and iterated through exploratory behavior and location. The feature selection scheme corresponding to the lemming individual with the highest fitness after each iteration is recorded as the current optimal feature selection scheme.

[0013] When the current optimal feature selection scheme remains unchanged in multiple consecutive iterations, it is determined that the stage switching condition is met, the feature selection stage ends, the current optimal feature selection scheme is output, and the parameter fine optimization stage begins.

[0014] In one possible implementation, determining that the stage switching condition is met when the current optimal feature selection scheme remains unchanged in multiple consecutive iterations includes:

[0015] Set up a queue to store the best historical feature selection schemes;

[0016] After each iteration, the current optimal feature selection scheme is added to the queue;

[0017] Determine whether the K most recent consecutive records in the queue are completely identical. If so, determine that the stage switching condition is met, where K is a preset positive integer.

[0018] In one possible implementation, the parameter fine-tuning stage includes:

[0019] Extract the feature subset and model hyperparameters corresponding to the current optimal feature selection scheme from the candidate feature set and the initial model hyperparameter range, and construct a temporary classifier;

[0020] Based on the temporary classifier, determine the classification confusion degree that reflects the temporary classifier;

[0021] Based on the classification confusion level, decisions are made regarding the continued exploration or development behavior of each lemming individual, the lemming's position is updated, and the optimal subset of surrounding rock features and the corresponding optimal model hyperparameters are obtained.

[0022] In one possible implementation, determining the classification perplexity reflecting the temporary classifier based on the temporary classifier includes:

[0023] Extract validation set data from the corresponding feature subset;

[0024] The temporary classifier is used to predict the surrounding rock grade for each sample in the validation set, and the probability value of each surrounding rock grade is obtained.

[0025] Calculate the entropy value for the probability value of each surrounding rock grade to obtain a list of entropy values ​​containing all validation set samples;

[0026] Based on the entropy value list, calculate the arithmetic mean of the entropy values ​​of all samples in the entropy value list;

[0027] The arithmetic mean is normalized to obtain the classification perplexity of the temporary classifier.

[0028] In one possible implementation, based on the classification confusion level, a decision is made regarding the continued exploration or development behavior of each lemming individual, the lemming's position is updated, and the optimal subset of surrounding rock features and the corresponding optimal model hyperparameters are obtained, including:

[0029] Based on the classification perplexity, calculate the adaptive energy decay coefficient for the current iteration progress;

[0030] Calculate the dynamic energy baseline value based on the adaptive energy decay coefficient and the current iteration progress;

[0031] Based on the dynamic energy baseline value, an independent final energy value is generated for each lemming individual in the population;

[0032] Based on whether the final energy value of each lemming individual is greater than the second threshold, control each lemming individual to perform exploratory or developmental behaviors to update its position;

[0033] When the maximum number of iterations is reached, the optimal subset of surrounding rock features and the corresponding optimal model hyperparameters are obtained.

[0034] In one possible implementation, calculating the adaptive energy decay coefficient for the current iteration progress based on the classification perplexity includes:

[0035] Obtain the preset lower bound of the attenuation coefficient, the preset upper bound of the attenuation coefficient, and the sensitivity coefficient;

[0036] Calculate the negative correlation adjustment factor for confusion based on the classification confusion level;

[0037] The adaptive energy attenuation coefficient is calculated using linear interpolation based on the preset lower bound of the attenuation coefficient, the preset upper bound of the attenuation coefficient, and the perplexity negative correlation adjustment factor.

[0038] In one possible implementation, calculating the dynamic energy baseline value based on the adaptive energy decay coefficient and the current iteration progress includes:

[0039] Get the preset maximum number of iterations and energy scaling constant;

[0040] Calculate the product of the adaptive energy decay coefficient and the proportion of the current iteration number to the maximum iteration number;

[0041] Based on the product, calculate the radian value that reflects the remaining search space;

[0042] The dynamic energy reference value is obtained by mapping the radian value and the energy scaling constant through trigonometric functions.

[0043] In one possible implementation, controlling each lemming individual to perform exploratory or developmental behaviors to update its location includes:

[0044] If the final energy value of an individual lemming is greater than the second threshold, it is determined that the lemming has performed an exploration behavior, and one of the behaviors, long-distance migration and burrowing, is randomly selected to be performed in order to update the part of the lemming position vector that represents the feature selection scheme and the model hyperparameters.

[0045] If the final energy value of an individual lemming is not greater than the second threshold, it is determined that the lemming has performed development behavior, and one of the behaviors of foraging and avoiding predators is randomly selected to be performed in order to update the part of the lemming position vector that represents the model hyperparameters.

[0046] This example provides a tunnel surrounding rock classification method based on multidimensional data. First, it obtains the candidate feature set, initial classifier, and their hyperparameter range for the surrounding rock samples. Then, it initializes a lemming population based on this range. Next, it employs an improved lemming optimization algorithm, initially performing a feature selection phase to guide the population to quickly explore and lock in key feature combinations by maintaining a fixed high-energy state. Once the optimal feature selection scheme stabilizes, it switches to a parameter fine-tuning phase. In this phase, based on the prediction uncertainty of the current optimal temporary classifier on the validation set, it dynamically calculates the energy decay rate, thereby adaptively adjusting the population's global exploration and local exploitation behavior near the feature subset. The algorithm ultimately outputs the optimal feature subset and model hyperparameters, and constructs a surrounding rock grade identification model accordingly. Through a collaborative optimization mechanism that first quickly focuses on features and then finely tunes parameters based on performance feedback, the algorithm can effectively balance the dual requirements of rapid preliminary identification and high-precision reliable identification in tunnel construction. On the one hand, it can quickly converge to key discriminative features in the early stage of optimization, providing timely and directional guidance for on-site decision-making. On the other hand, in the later stage, it can automatically extend the exploration and strengthen the optimization for complex geological situations with high uncertainty in model classification, thereby significantly improving the classification accuracy and generalization robustness of the constructed identification model in difficult scenarios such as blurred feature boundaries.

[0047] A second aspect of this application provides a tunnel surrounding rock classification system based on multidimensional data, the system comprising:

[0048] The data acquisition module is used to acquire the candidate feature set of tunnel surrounding rock samples, the initial tunnel surrounding rock level classifier, and the initial model hyperparameter range;

[0049] The population initialization module is used to initialize the lemming population by random sampling based on the candidate feature set and the range of initial model hyperparameters, so as to obtain the initial lemming population.

[0050] The model optimization module is used to perform the feature screening stage and parameter fine optimization stage using the improved lemming optimization algorithm to screen the candidate feature set and the initial model hyperparameter range of the tunnel surrounding rock sample, and obtain the optimal surrounding rock feature subset and the corresponding optimal model hyperparameters.

[0051] The model building module is used to construct the final surrounding rock grade identification model based on the optimal surrounding rock feature subset and the optimal model hyperparameters, and to use the final surrounding rock grade identification model to classify the surrounding rock of the tunnel.

[0052] A third aspect of this application provides a terminal including a processor, an input device, an output device, and a memory, wherein the processor, input device, output device, and memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute the step instructions as described in the tunnel surrounding rock classification method based on multidimensional data in the first aspect of this application.

[0053] A fourth aspect of this application provides a computer-readable storage medium storing a computer program for electronic data interchange, wherein the computer program causes a computer to perform some or all of the steps described in the tunnel surrounding rock classification method based on multidimensional data in the first aspect of this application.

[0054] A fifth aspect of this application provides a computer program product, wherein the computer program product includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps described in the tunnel surrounding rock classification method based on multidimensional data in the first aspect of this application. The computer program product may be a software installation package. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This application provides a schematic diagram of the overall process for a tunnel surrounding rock classification method based on multidimensional data.

[0057] Figure 2 This application provides a schematic diagram of the overall structure of a tunnel surrounding rock classification system based on multidimensional data.

[0058] Figure 3 This application provides a schematic diagram of the structure of a terminal.

[0059] Figure label:

[0060] Data acquisition module-1, population initialization module-2, model optimization module-3, model construction module-4. Detailed Implementation

[0061] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0062] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0063] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0064] To better understand the tunnel surrounding rock classification method based on multidimensional data provided in this application, a brief introduction to the application scenario of this method is given below. Currently, data-driven intelligent classification methods have become a mainstream research direction. These methods typically extract multidimensional features from geological exploration data and use machine learning models to establish a mapping relationship between features and surrounding rock grades. To improve model performance, existing advanced methods generally introduce metaheuristic optimization algorithms, such as genetic algorithms and particle swarm optimization, to automatically complete two key tasks: first, to select the optimal feature subset from a large number of candidate features to reduce dimensionality and improve feature discriminative power; and second, to search for the optimal hyperparameter combination for the selected classification model to fully unleash the model's potential. This method reduces reliance on human experience to some extent by modeling feature selection and parameter tuning as a complex combinatorial optimization problem, thus automating the recognition process. However, these optimization methods based on general metaheuristic algorithms still have inherent defects when applied to the specific scenario of tunnel surrounding rock classification. First, the optimization process of surrounding rock grade identification has an inherent phased pattern. Feature selection determines the basic discrimination framework of the model, and importance usually takes precedence over fine-tuning of parameters. However, existing algorithms use a uniform search strategy to process features and parameters simultaneously, failing to explicitly distinguish and utilize this phase priority. This results in low efficiency of the search process in the global space and a tendency to converge prematurely near suboptimal feature combinations. Second, the search dynamics of the algorithm are open-loop. Its search strategy is usually pre-set based on fixed rules such as the number of iterations, failing to perceive and respond in real time to the actual classification performance of the model on the current solution. This leads to a lack of ability to adaptively adjust the search focus when facing ambiguous classification regions, thus limiting the upper limit of accuracy and generalization stability of the final classification model under complex geological conditions.

[0065] The tunnel surrounding rock classification method based on multidimensional data is applied to a tunnel surrounding rock classification system based on multidimensional data. Figure 1 A schematic diagram illustrating the overall process of a tunnel surrounding rock classification method based on multidimensional data is shown. Figure 1 As shown, it includes:

[0066] S1. Obtain the candidate feature set of the tunnel surrounding rock sample, the initial tunnel surrounding rock level classifier, and the initial model hyperparameter range.

[0067] The candidate feature set is a collection of M geological and rock mechanics quantitative indicators, which are key parameters that can be obtained from tunnel engineering geological exploration and laboratory tests and can reflect the stability of the surrounding rock.

[0068] The initial tunnel surrounding rock level classifier is a pre-defined, optimized machine learning model framework. In a preferred embodiment of the invention, a support vector machine can be used as the initial classifier. The performance of this classifier is determined by its hyperparameters, and the range of the initial model hyperparameters represents the upper and lower limits of the allowed values ​​of these hyperparameters.

[0069] S2. Initialize the lemming population by random sampling based on the candidate feature set and the range of initial model hyperparameters to obtain the initial lemming population.

[0070] Specifically, based on the candidate feature set and the initial model hyperparameter range, a model containing [specific parameters] is constructed through random sampling. The initial lemming population of the candidate solutions Among them, population size The superscript is a preset positive integer. This represents the initial iteration number. Each lemming individual... ( () represents a complete model building scheme, whose location is determined by a A composite vector representation of dimension 1, where the dimension of the vector is 1. , The total number of candidate features. This represents the number of hyperparameters in the model to be optimized.

[0071] Furthermore, the structure and initialization generation method of the position vector are as follows:

[0072] The position vector of each lemming individual Divided into two parts according to dimensions:

[0073] The first paragraph is the beginning. Dimension (feature selection identifier segment), the vector of this segment Used to encode feature selection schemes. Each dimension ( For a given feature in the candidate feature set, its initial value is determined by the following formula:

[0074] (1)

[0075] In the formula, For in the interval A random number uniformly distributed on the vector. The features corresponding to the dimensions with a value of 1 in this vector segment will be selected to form the feature subset represented by this individual.

[0076] The second paragraph is after Dimension (model hyperparameter segment), the vector of this segment This is used to encode the specific values ​​of the hyperparameters of the model. Each dimension corresponds to a hyperparameter to be optimized. Let the th dimension be... ( The preset value range for each hyperparameter is: Then the initial value of its corresponding dimension in the position vector. The formula is shown below:

[0077] (2)

[0078] In the formula, For in the interval The formula uses a uniformly distributed random number to ensure that the initial value of each hyperparameter is generated randomly and uniformly within its preset feasible range.

[0079] In this example, the final result includes The lemming population of the initial solution .

[0080] S3. Based on the initial lemming population, the improved lemming optimization algorithm is used to perform the feature screening stage and parameter fine optimization stage to screen the candidate feature set and the initial model hyperparameter range of the tunnel surrounding rock sample, so as to obtain the optimal surrounding rock feature subset and the corresponding optimal model hyperparameters.

[0081] Step S3 includes a feature selection stage and a parameter fine-tuning stage.

[0082] In one possible implementation, the feature selection stage includes the following steps:

[0083] S3-1. Set the energy factor of the initial lemming population to a fixed value greater than the first threshold, so that all lemming individuals in the initial lemming population only perform exploratory behavior to update their location.

[0084] In this stage, a constant energy factor is set. ,and It is understandable that in each iteration, the energy of each lemming individual i... All equal to In the standard lemming algorithm, when the individual energy At this time, individual lemmings will perform exploratory behaviors (the exploratory behaviors of the lemming algorithm include long-distance migration or digging).

[0085] S3-2. Iterate through the exploratory behavior positions of the initial lemming population, and record the feature selection scheme corresponding to the lemming individual with the highest fitness after each iteration, which is then used as the current optimal feature selection scheme.

[0086] In each generation iteration, each lemming individual in the population P(t) is determined according to its energy. The individual performs exploratory behaviors to update its location. Specifically, for each individual, it randomly selects to perform either long-distance migration or burrowing behavior with a 50% probability, and updates its location vector according to the corresponding original lemming algorithm formula. After the position is updated, the lemming's position needs to be corrected by adjusting the position vector. For the dimension (feature selection part), if a dimension value is ≥0.5, it is set to 1; otherwise, it is set to 0. Dimension (hyperparameter part), constrain its value to a preset [ , Within the range.

[0087] S3-3. When the current optimal feature selection scheme remains unchanged in multiple consecutive iterations, it is determined that the stage switching condition is met, the feature selection stage ends, the current optimal feature selection scheme is output, and the parameter fine optimization stage begins.

[0088] S3-3 includes the following sub-steps:

[0089] S3-3-1. Set up a queue to store the best historical feature selection schemes.

[0090] S3-3-2. After each iteration, the current optimal feature selection scheme is added to the queue.

[0091] S3-3-3. Determine whether the K most recent consecutive records in the queue are completely identical. If so, determine that the stage switching condition is met, where K is a preset positive integer.

[0092] Specifically, in this example, we define the queue as Q and the stability judgment constant as K. After each iteration, the current optimal feature selection scheme is added to queue Q, and we check whether the last K consecutive records in queue Q are completely identical. If they are identical, the feature selection scheme is considered stable, and the stage switching condition is met. At this point, the output current optimal feature selection scheme is used as the optimal feature selection scheme for the first stage, and the process enters the parameter fine-tuning stage. If the condition is not met, then t = t + 1, and we return to step S3-1 to continue iterating until the switching condition is met or the preset maximum number of independent iterations for this stage is reached.

[0093] In one possible implementation, the parameter fine-tuning stage includes the following steps:

[0094] S3-4. Extract the feature subset and model hyperparameters corresponding to the current optimal feature selection scheme from the candidate feature set and the initial model hyperparameter range, and construct a temporary classifier.

[0095] Specifically, the feature subset corresponding to the current optimal feature selection scheme can be decoded to obtain a specific feature subset, and the data corresponding to these features can be extracted from the dataset. Simultaneously, the corresponding hyperparameter values ​​are decoded from the position vector of the current globally optimal lemming individual. Using this feature subset and hyperparameters, a new classifier model is trained on the training set as a temporary classifier.

[0096] S3-5. Determine the classification confusion degree reflecting the temporary classifier based on the temporary classifier.

[0097] Step S3-5 includes the following sub-steps:

[0098] S3-5-1. Extract the validation set data from the corresponding feature subset.

[0099] In each iteration of the parameter fine-tuning phase, a corresponding subset of features is extracted from the original dataset based on the feature selection scheme represented by the currently globally optimal lemming individual. Specifically, the original dataset has been divided into training, validation, and test sets during the preprocessing stage, with the validation set used to evaluate model performance in real time and generate feedback signals. In this step, only the feature columns corresponding to the current optimal feature subset are extracted from the validation set to form the validation set data matrix. ,in This represents the number of samples in the validation set. The number of features selected in the current feature subset.

[0100] S3-5-2. The temporary classifier is used to predict the surrounding rock grade for each sample in the validation set, and the probability value of each surrounding rock grade is obtained.

[0101] Specifically, the temporary classifier constructed in steps S3-4 (e.g., a classification model based on support vector machines and employing softmax probability output) is used to evaluate the validation set data matrix. Make a prediction. For the first [number] [item] in the validation set... One sample ( The temporary classifier outputs a probability vector. ,in The total number of categories indicating the surrounding rock grade (usually 5, i.e., grades I, II, III, IV, and V). This indicates that the sample belongs to the first... The predicted probabilities of each category satisfy the following conditions: and .

[0102] In this example, we obtain the complete probability distribution vector for each sample. Rather than a single predicted category label, the probability vector aims to quantify the model's "classification uncertainty." The probability distribution contains information about the model's confidence level in classifying a sample. Therefore, when the model is very certain about classifying a sample, its probability distribution will concentrate on a single category; conversely, when the model is "confused" (e.g., when the model has difficulty distinguishing the boundary between Class III and Class IV rock types), the probability distribution will tend to be more even. Thus, the probability vector is the direct input for calculating subsequent uncertainty metrics and is crucial for capturing subtle changes in model performance.

[0103] S3-5-3. Calculate the entropy value for the probability value of each surrounding rock grade to obtain a list of entropy values ​​containing all validation set samples.

[0104] For each validation set sample Based on its probability vector Calculate the entropy value of its predicted probability distribution. Entropy is used to measure the uncertainty of a classifier in classifying a sample. The calculation formula is as follows:

[0105] (3)

[0106] In the formula, For each validation set sample The entropy value, preferably the natural logarithm in this embodiment, is obtained by traversing all validation set samples to obtain a list of entropy values. .

[0107] In this example, entropy is a classic metric for measuring the uncertainty of a random variable. Here, the model's prediction of a single sample is treated as a random event, and the entropy of this distribution is calculated. It can condense the uncertainty information contained in the probability vector into a scalar value. The larger the value, the more uniform the probability distribution, and the more uncertain and "confusing" the model is in classifying the sample. The smaller the entropy value, the more concentrated the probability distribution and the more certain the model classification. Therefore, the entropy value is a core intermediate indicator for quantifying the classification perplexity at the single-sample level.

[0108] S3-5-4. Based on the entropy value list, calculate the arithmetic mean of the entropy values ​​of all samples in the entropy value list.

[0109] Among them, the entropy list Sum all entropy values ​​and divide by the total number of samples in the validation set. The average entropy value is obtained. The calculation formula is as follows:

[0110] (4)

[0111] In the formula, This is the arithmetic mean of the entropy values ​​of all samples.

[0112] In this example, the perplexity of a single sample The results may be affected by a few outliers and are insufficient to represent the overall performance of the model. Calculate the arithmetic mean of the entropy values ​​of all validation set samples. The aim is to obtain an aggregated index that can robustly reflect the overall classification uncertainty level of the temporary classifier. It integrates the model's grasp of all samples (including those that are easy to classify and those that are difficult to classify), providing the algorithm with a stable feedback signal on the global performance of the current search point (i.e., the current feature subset and hyperparameter combination).

[0113] S3-5-5. Normalize the arithmetic mean to obtain the classification perplexity of the temporary classifier.

[0114] In this example, a temporary classifier is used to predict the probability distribution of each sample belonging to each surrounding rock grade on an independent validation set. Based on this, the classification perplexity is calculated. Classification confusion The calculation formula is as follows:

[0115] (5)

[0116] in, The classification perplexity ranges from [0,1]. A larger value indicates that the model is less certain about classifying the validation set. To validate the total number of samples in the set, This represents the total number of categories of surrounding rock grades. This represents the probability predicted by the temporary classifier that the i-th sample belongs to the c-th class.

[0117] In this example, normalization serves two key purposes. First, it standardizes the unit of measurement: entropy. The theoretical maximum value is (When the probability of all categories is equal) (Time). Divide by Can Strictly limited to the [0,1] interval, eliminating the number of categories. The varying impacts on the numerical range make the perplexity a standardized, interpretable proportional value. Secondly, it facilitates subsequent adjustment: normalizing the feedback signal makes it easier to approximate the preset attenuation coefficient boundary. Integrating and mapping parameters makes the parameter settings of the entire adaptive control loop more intuitive and the system behavior more stable. Therefore, classification perplexity... It is a normalized final feedback signal representing the overall uncertainty of the model. It directly and quantitatively reflects the performance of the current optimal model in the key task (surrounding rock grade classification) and is the source driving the dynamic adjustment of subsequent search strategies.

[0118] In summary, in this example, each step—from probability distribution calculation (S3-5-2) to single-sample entropy calculation (S3-5-3), then to average entropy calculation (S3-5-4), and finally to normalized perplexity calculation (S3-5-5)—is a process of refining the model's classification uncertainty from micro to macro, from primitive to standardized. These calculation steps ensure the accuracy of the feedback signal. It can sensitively capture the performance fluctuations of the model on difficult samples (such as those with blurred boundaries) and stably represent the overall state of the model, providing accurate and reliable quantitative input for the algorithm to achieve adaptive optimization of "perception-feedback-control".

[0119] S3-6. Based on the classification confusion level, make decisions on the continued exploration or development behavior of each lemming individual, update the lemming's position, and obtain the optimal subset of surrounding rock features and the corresponding optimal model hyperparameters.

[0120] Step S3-6 includes the following sub-steps:

[0121] S3-6-1. Calculate the adaptive energy decay coefficient for the current iteration progress based on the classification confusion degree.

[0122] Specifically, step S3-6-1 includes the following sub-steps:

[0123] S3-6-1-1, Obtain the preset lower bound of the attenuation coefficient, the preset upper bound of the attenuation coefficient, and the sensitivity coefficient.

[0124] Before performing the calculation, three preset fixed parameters are read from the algorithm configuration: the lower bound of the attenuation coefficient. Upper bound of attenuation coefficient and sensitivity coefficient .in, and It is a positive integer and satisfies For example, a set of feasible values , Sensitivity coefficient It is also a positive number, for example, it can be taken as... .

[0125] S3-6-1-2. Calculate the negative correlation adjustment factor of confusion degree based on the classification confusion degree.

[0126] The classification perplexity of the current iteration can be calculated using steps S3-5. Calculate an intermediate variable, called the perplexity negative correlation adjustment factor, whose value is... That is, first calculate and The difference is obtained. The value range is Internally, it is negatively correlated with confusion level; then this difference is analyzed. Exponentiation.

[0127] In this example, calculation This is the core of this step, completing the conversion and nonlinear mapping of the feedback signal; in the negative correlation conversion, the following steps are used: Instead This is because, in algorithm design, we hope that when the model has high perplexity ( When the energy decays (larger), it should be slower. (Should be small). Therefore, it is necessary to establish a connection with... The negatively correlated quantity is used as the basis for adjustment. Exactly satisfies: The larger, The smaller; in nonlinear mappings, for conduct Exponentiation is key to introducing nonlinearity. When When, the mapping is linear; when When the function is convex, its effect is: Smaller (more defined model) regions right The changes are relatively insensitive; while Larger regions (the model is confused about), right The rate of decay is highly sensitive to increases and decreases dramatically. This design makes the algorithm react more strongly and explicitly to "high perplexity" states, ensuring that when optimization encounters bottlenecks (such as getting stuck in suboptimal solutions or facing data with ambiguous class boundaries), the decay rate can be reduced quickly and sufficiently, providing a longer exploration window to escape local optima and find better solutions.

[0128] S3-6-1-3. Based on the preset lower bound of the attenuation coefficient, the preset upper bound of the attenuation coefficient, and the perplexity negative correlation adjustment factor, the adaptive energy attenuation coefficient is calculated using linear interpolation.

[0129] Among them, the negative correlation adjustment factor of the perplexity calculated in the previous step is used. As interpolation weights, within the preset attenuation coefficient boundary Linear interpolation is performed between them to calculate the adaptive energy decay coefficient of the current generation. Adaptive energy decay coefficient Classification confusion The calculation is performed using the formula shown below:

[0130] (6)

[0131] In the formula, Let be the adaptive energy decay coefficient for generation t; , The preset lower and upper limits of the attenuation coefficient, The preset sensitivity coefficient is used to adjust P. The degree of nonlinearity of the influence To classify confusion level.

[0132] In this example, this step is the one that generates the final control command, mapping the normalized feedback signal to specific control parameters through linear interpolation; for the linear interpolation mechanism, the formula is... It is a standard linear interpolation. The weights are... .when When the model is fully determined, the weight is 1. The algorithm employs the fastest energy decay rate, prompting the search to quickly shift from exploration to development, achieving efficient convergence. When the model is completely uncertain, the weights are 0. The algorithm employs the slowest energy decay rate to maintain a high energy state in the population for extensive global exploration. For intermediate states... , Then in Smooth changes within the interval.

[0133] Furthermore, the calculated It is the core bridging variable connecting model performance feedback and algorithm search behavior. It is not a fixed value, but rather changes with the model's real-time performance on the validation set (perplexity). It is dynamically adjusted. Its value directly determines the subsequent dynamic energy reference value. The decay rate is thus controlled, thereby macroscopically regulating the balance between "exploration" and "exploitation" in the entire population. In this way, the algorithm's search rhythm is no longer pre-set and unchanging, but becomes a closed-loop adaptive system that can intelligently adjust its search strategy based on the quality of the current solution. This is the key to improving the optimization effect and model generalization ability of this invention.

[0134] S3-6-2. Calculate the dynamic energy reference value based on the adaptive energy decay coefficient and the current iteration progress.

[0135] Step S3-6-2 includes the following sub-steps:

[0136] S3-6-2-1. Obtain the preset maximum number of iterations and energy scaling constant.

[0137] S3-6-2-2, Calculate the product of the adaptive energy decay coefficient and the ratio of the current iteration number to the maximum iteration number.

[0138] First, calculate the current iteration number. With maximum number of iterations The proportion, i.e. This value represents the percentage of progress the algorithm has made, ranging from [0,1]. Then, this percentage is compared with the adaptive energy decay coefficient calculated in step S3-6-1. Multiply to get the product .

[0139] In this example, This is a factor that grows linearly with time, reflecting the global progress of the algorithm from start to finish. Introducing it into the calculation is to introduce a time dimension into the energy decay process, ensuring that even the decay coefficient... As the model becomes confused, the energy baseline value will gradually decrease (albeit slowly) with each iteration, thus preventing the algorithm from remaining in the exploration phase indefinitely and ensuring eventual convergence.

[0140] For product This is the core metric for the dynamic decay rate. Among them, The model is adaptively adjusted based on its real-time performance, and This represents a fixed time progression. Multiplying the two means that the actual rate of energy decay is determined by both model feedback and iteration progress. When the model perplexity is high, When the value is relatively small, the product grows slowly, and the energy baseline decreases slowly; when the model is highly deterministic... The larger the value, the faster the product grows, and the faster the energy baseline decreases. This allows the algorithm to flexibly adjust its search pace based on the quality of the current solution without completely escaping the constraints of the time schedule.

[0141] S3-6-2-3. Calculate the radian value reflecting the remaining search space based on the product.

[0142] Here, the product obtained in the previous step can be... Substitute into the formula In the middle, we get a range in The value within the range is obtained; then, the arctangent function value of that value is calculated, thus obtaining... The output value (in radians) of the arctangent function ranges from... between.

[0143] In this example, the constructed Subtracting the product from 1 is used to construct a variable that decreases as the iteration proceeds. When Initially, this value is 1; with iteration, the value gradually decreases, and may even become negative. This decreasing characteristic can well map the characteristics of energy decay; then the arctangent function is applied. It has three functions. First, it will input from Mapping to a bounded interval First, it ensures the stability of the output value and avoids the occurrence of extreme values. Second, its nonlinear characteristics make the energy decay curve more flexible. It is not a simple linear decrease. It can decay slowly in the early stage, decay faster in the later stage, or adjust its shape according to the input. Third, the arctangent function is monotonically decreasing. Therefore, the smaller the input value, the smaller the output value. This is consistent with the expectation that the energy decreases as the search progresses.

[0144] S3-6-2-4. Based on the radian value and the energy scaling constant, the dynamic energy reference value is obtained through trigonometric function mapping.

[0145] In this example, the dynamic energy reference value The calculation formula is as follows:

[0146] (7)

[0147] In the formula, This is the dynamic energy baseline value for generation t. This is a preset energy scaling constant used to adjust the energy scale; It is the arctangent function; For adaptive energy decay coefficient, This represents the current iteration number. The maximum number of iterations is preset for the entire optimization process.

[0148] In this example, the dynamic energy reference value This is the final output obtained in this step, and one of the core outputs of the entire dynamic energy regulation mechanism. It represents the output at the current iteration. According to model feedback (through (and iteration progress) set an energy level benchmark for the entire lemming population. The value of directly determines the overall energy level of the population: the larger the value, the more the population tends to explore (because more individuals will have energy above the behavior switching threshold); the smaller the value, the more the population tends to exploit. Through the design of the formula in this step, It implements a non-linear, adaptive feedback decay trajectory, replacing the fixed decay curve in the original ALA (lemming algorithm). This allows the algorithm to adaptively adjust the global search pace according to the real-time difficulty of the optimization problem (manifested as model perplexity), thereby achieving the best balance between fast convergence and thorough exploration.

[0149] S3-6-3. Generate an independent final energy value for each lemming individual in the population based on the dynamic energy benchmark value.

[0150] Among them, the dynamic energy reference value is obtained Then, for each individual lemming in the population. Generate its independent final energy value The specific implementation is as follows: First, independently generate a range for each individual within the interval. Uniformly distributed random numbers The formula for calculating the final energy value of an individual lemming is as follows:

[0151] (8)

[0152] In the formula, This represents the final energy value of an individual lemming. This is the dynamic energy baseline value for generation t. These are uniformly distributed random numbers.

[0153] In this example, although the dynamic energy reference value A macro-level energy level is set for the entire population. However, if all individuals adopt the exact same energy value, the population's behavior will become homogeneous, losing diversity and easily falling into local optima. Therefore, random numbers are generated independently for each individual. It is a key mechanism for injecting individual random differences under unified control; log factor Uniformly distributed random numbers Transforming the variable into a random variable that follows an exponential distribution (with a mean of approximately 1) has two important functions: First, it ensures the energy value of the generated individuals. First, it is always positive; second, it ensures that the energy value of most individuals is at the baseline value. While there are fluctuations in the surrounding area, there are also individuals with significantly higher or lower energy levels, thus naturally forming behavioral diversity.

[0154] For the individual's final energy value It serves as the sole direct basis for each lemming's behavioral decisions. It inherits and amplifies the global search strategy (exploration / exploration tendency) inherent in the dynamic baseline, while simultaneously endowing each individual with a unique "personality" through random factors. This design of "unified macroscopic guidance and random microscopic variation" ensures that the population as a whole follows the optimal search rhythm determined by model performance feedback, while maintaining sufficient diversity at the individual level, enabling effective wide-area exploration while avoiding premature convergence.

[0155] S3-6-4. Based on whether the final energy value of each lemming individual is greater than the second threshold, control each lemming individual to perform exploration or development behavior to update its position.

[0156] S3-6-4-A. If the final energy value of an individual lemming is greater than the second threshold (usually 1 in the original lemming algorithm, and 1 in this example), then the lemming is determined to perform exploration behavior, and one of the behaviors of long-distance migration and burrowing is randomly selected to perform, so as to update the part of the lemming position vector that represents the feature selection scheme and the model hyperparameters.

[0157] Specifically, if The lemming is determined to perform exploratory behavior. Specifically, it is randomly selected to perform either long-distance migration or burrowing behavior with equal probability (e.g., 50%). When performing long-distance migration, the lemming's position vector (including feature selection and hyperparameter components) undergoes significant random perturbation or reset, aiming to escape the current region and explore the solution space far from the current location. When performing burrowing behavior, the lemming performs small-scale, high-frequency random perturbations near its current location, aiming to conduct detailed exploration of the neighborhood of the current solution.

[0158] S3-6-4-B. If the final energy value of an individual lemming is not greater than the second threshold, the lemming is determined to perform development behavior, and one of the behaviors of foraging and avoiding predators is randomly selected to perform, so as to update the part of the lemming position vector that represents the model hyperparameters.

[0159] Among them, if The lemming is then assessed for exhibiting developmental behavior. Specifically, it randomly selects to forage or evade predators with equal probability. Both behaviors constitute local fine-tuning, but their focuses differ slightly. They primarily update the portion of the position vector representing the model's hyperparameters, while keeping the feature selection portion unchanged. This is because, during the parameter fine-tuning phase, the optimal feature subset has been preliminarily identified, and the core task is to find the optimal hyperparameter combination based on this feature subset. Developmental behavior, through local fine-tuning of hyperparameters, fully exploits the model's performance potential under the current feature subset.

[0160] In this example, this step is the final execution link in the entire feedback control chain. Energy value The level of perplexity is directly determined by the preceding perplexity feedback and the dynamic baseline. When the overall perplexity of the model is high, Maintaining a high level leads to more individuals This increases the proportion of individuals in the population engaging in exploratory behavior, prompting the algorithm to strengthen its global search to find more discriminative solutions. Conversely, when the model is highly deterministic, more individuals engage in exploitative behavior, and the algorithm focuses on local optimization and convergence. This step transforms the soft feedback of model performance into real-time, automatic hard actions that adjust the population's exploratory-exploratory balance, directly demonstrating the algorithm's context-aware and adaptive capabilities.

[0161] S3-6-5. When the maximum number of iterations is reached, the optimal subset of surrounding rock features and the corresponding optimal model hyperparameters are obtained.

[0162] The process involves repeatedly executing steps S3-4 to S3-6-4 (note that S3-4 only constructs a temporary classifier once at the beginning of each generation using the current globally optimal individual). After each generation, t = t+1. At the end of each iteration, the population fitness is evaluated and the globally historical optimal solution is updated. When t reaches a preset threshold... The iteration terminates at this point. Finally, the position vector of the globally best lemming individual is decoded, and its previous position is determined. The dimensionality is converted into a binary feature selection scheme to obtain the optimal subset of surrounding rock features; then... The dimensions are mapped to specific hyperparameter values ​​to obtain the optimal model hyperparameters.

[0163] The iterative process consists of a closed loop in each complete cycle (generation): "evaluating the current optimal model performance (perplexity) -> calculating dynamic energy parameters -> individual behavior decisions and position updates -> fitness evaluation and optimal solution updates". Through multiple generations of iteration, the population continuously moves towards a better solution space region under the guidance of the adaptive strategy.

[0164] The result decoding involves decoding the position vector of the globally best lemming individual after the iteration terminates. This involves decoding the first... The dimension (feature selection identifier segment) is converted into a binary feature selection scheme, where the features corresponding to dimensions with a value of 1 constitute the optimal subset of surrounding rock features. The position vector of this subset is then processed. The values ​​of the dimension (model hyperparameter segment) are mapped back to the actual value range of each hyperparameter to obtain the optimal combination of model hyperparameters.

[0165] In summary, this example yields the optimal subset of surrounding rock features and the optimal model hyperparameters as the final output of the improved lemming optimization algorithm running under a dual-stage (feature selection and parameter fine-tuning) and adaptive feedback control mechanism. This represents the feature and parameter configurations found by the algorithm after comprehensively balancing rapid focusing with fine-tuning and global exploration with local development, resulting in the best expected classification performance on the validation set. This result forms the direct basis for subsequently constructing a high-precision, high-generalization-capability surrounding rock grade identification model.

[0166] S4. Based on the optimal subset of surrounding rock features and the optimal model hyperparameters, construct the final surrounding rock grade identification model, and use the final surrounding rock grade identification model to classify the surrounding rock of the tunnel.

[0167] After obtaining the optimal subset of surrounding rock features and the optimal model hyperparameters output by the improved lemming optimization algorithm, the final model construction and application stage begins. First, based on the optimal feature subset, corresponding feature data is selected from the complete multi-dimensional data of tunnel surrounding rock to form a refined dataset for model training and testing. Next, using this dataset, a selected initial classifier (such as a support vector machine) is configured and trained with the optimal model hyperparameters to obtain a fully calibrated, highly discriminative final surrounding rock grade identification model. This model embeds the complex mapping relationship between features and surrounding rock grades (I, II, III, IV, V) into its internal parameters. In practical applications, new sample data from the tunnel face or exploration section to be identified, after the same feature extraction and selection, is input into this final model, which quickly and automatically outputs the predicted surrounding rock grade results. In actual tunnel engineering, this model can provide immediate and reliable surrounding rock grade identification results for construction, significantly reducing reliance on human experience and assisting in construction decision-making, risk warning, and scheme optimization. Its core value lies in ensuring that, through the aforementioned adaptive optimization mechanism, the model can maintain high classification reliability and stability even when facing difficult scenarios with complex geological conditions and ambiguous category boundaries, thereby effectively improving the safety assurance capability and intelligence level of tunnel construction.

[0168] This example provides a tunnel surrounding rock classification method based on multidimensional data. First, it obtains the candidate feature set, initial classifier, and their hyperparameter range for the surrounding rock samples. Then, it initializes a lemming population based on this range. Next, it employs an improved lemming optimization algorithm, initially performing a feature selection phase to guide the population to quickly explore and lock in key feature combinations by maintaining a fixed high-energy state. Once the optimal feature selection scheme stabilizes, it switches to a parameter fine-tuning phase. In this phase, based on the prediction uncertainty of the current optimal temporary classifier on the validation set, it dynamically calculates the energy decay rate, thereby adaptively adjusting the population's global exploration and local exploitation behavior near the feature subset. The algorithm ultimately outputs the optimal feature subset and model hyperparameters, and constructs a surrounding rock grade identification model accordingly. Through a collaborative optimization mechanism that first quickly focuses on features and then finely tunes parameters based on performance feedback, the algorithm can effectively balance the dual requirements of rapid preliminary identification and high-precision reliable identification in tunnel construction. On the one hand, it can quickly converge to key discriminative features in the early stage of optimization, providing timely and directional guidance for on-site decision-making. On the other hand, in the later stage, it can automatically extend the exploration and strengthen the optimization for complex geological situations with high uncertainty in model classification, thereby significantly improving the classification accuracy and generalization robustness of the constructed identification model in difficult scenarios such as blurred feature boundaries.

[0169] For those consistent with the above, please refer to Figure 2 , Figure 2 This application provides a schematic diagram of the structure of a tunnel surrounding rock classification system based on multidimensional data. For example... Figure 2 As shown, the system includes:

[0170] Data acquisition module 1 is used to acquire the candidate feature set of tunnel surrounding rock samples, the initial tunnel surrounding rock level classifier, and the initial model hyperparameter range;

[0171] Population initialization module 2 is used to initialize the lemming population by random sampling based on the candidate feature set and the range of initial model hyperparameters to obtain the initial lemming population;

[0172] Model optimization module 3 is used to perform the feature screening stage and parameter fine optimization stage using the improved lemming optimization algorithm to screen the candidate feature set and the initial model hyperparameter range of the tunnel surrounding rock sample, and obtain the optimal surrounding rock feature subset and the corresponding optimal model hyperparameters.

[0173] Model building module 4 is used to build a final surrounding rock grade identification model based on the optimal surrounding rock feature subset and the optimal model hyperparameters, and to use the final surrounding rock grade identification model to classify the surrounding rock of the tunnel.

[0174] For examples consistent with the above embodiments, please refer to... Figure 3 , Figure 3A schematic diagram of a terminal structure provided in an embodiment of this application is shown in the figure. It includes a processor, an input device, an output device, and a memory. The processor, input device, output device, and memory are interconnected. The memory is used to store a computer program, which includes program instructions. The processor is configured to call the program instructions. The program includes instructions for performing the following steps.

[0175] Obtain the candidate feature set, initial tunnel surrounding rock level classifier, and initial model hyperparameter range for tunnel surrounding rock samples;

[0176] The lemming population is initialized by random sampling based on the candidate feature set and the range of initial model hyperparameters to obtain the initial lemming population.

[0177] Based on the initial lemming population, an improved lemming optimization algorithm is used to perform a feature screening stage and a parameter fine optimization stage to screen the candidate feature set and the initial model hyperparameter range of the tunnel surrounding rock sample, thereby obtaining the optimal surrounding rock feature subset and the corresponding optimal model hyperparameters.

[0178] Based on the optimal subset of surrounding rock features and the optimal model hyperparameters, a final surrounding rock grade identification model is constructed, and the final surrounding rock grade identification model is used to classify the surrounding rock of tunnels.

[0179] This example provides a tunnel surrounding rock classification method based on multidimensional data. First, it obtains the candidate feature set, initial classifier, and hyperparameter range of the surrounding rock samples, and initializes a lemming population based on this range. Then, it employs an improved lemming optimization algorithm, first performing a feature selection phase to guide the population to quickly explore and lock in key feature combinations by fixing a high-energy state. Once the optimal feature selection scheme stabilizes, it switches to a parameter fine-tuning phase. In this phase, based on the prediction uncertainty of the current optimal temporary classifier on the validation set, it dynamically calculates the energy decay rate, thereby adaptively adjusting the population's global exploration and local development behavior near the feature subset. The algorithm ultimately outputs the optimal feature subset and model hyperparameters, and constructs a surrounding rock grade identification model accordingly. Through a collaborative optimization mechanism that first quickly focuses on features and then finely tunes parameters based on performance feedback, the algorithm can effectively balance the dual requirements of rapid preliminary identification and high-precision reliable identification in tunnel construction. On the one hand, it can quickly converge to key discriminative features in the early stage of optimization, providing timely and directional guidance for on-site decision-making. On the other hand, in the later stage, it can automatically extend the exploration and strengthen the optimization for complex geological situations with high uncertainty in model classification, thereby significantly improving the classification accuracy and generalization robustness of the constructed identification model in difficult scenarios such as blurred feature boundaries.

[0180] The above mainly describes the solutions of the embodiments of this application from the perspective of the method execution process. It is understood that, in order to achieve the above functions, the terminal includes the corresponding hardware structure and / or software modules for executing each function. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments provided herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0181] This application embodiment can divide the terminal into functional units according to the above method example. For example, each function can be divided into a separate functional unit, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this application embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0182] This application also provides a computer storage medium storing a computer program for electronic data interchange, which causes a computer to perform some or all of the steps of any of the tunnel surrounding rock classification methods based on multidimensional data as described in the above method embodiments.

[0183] This application also provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program that causes a computer to perform some or all of the steps of any of the tunnel surrounding rock classification methods based on multidimensional data as described in the above method embodiments.

[0184] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0185] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0186] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical or other forms.

[0187] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0188] Furthermore, the functional units in the various embodiments of the application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software program module.

[0189] If the integrated unit is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0190] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage device, which may include: a flash drive, a read-only memory, a random access memory, a magnetic disk, or an optical disk, etc.

[0191] The embodiments of this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A tunnel surrounding rock classification method based on multi-dimensional data, characterized in that, include: The candidate feature set, initial tunnel surrounding rock level classifier, and initial model hyperparameter range are obtained from the tunnel surrounding rock samples. The candidate feature set is a collection of M geological and rock mechanics quantitative indicators, which are key parameters that can be obtained from tunnel engineering geological exploration and laboratory tests and can reflect the stability of the surrounding rock. The lemming population is initialized by random sampling based on the candidate feature set and the range of initial model hyperparameters to obtain the initial lemming population. Based on the initial lemming population, an improved lemming optimization algorithm is used to perform a feature screening stage and a parameter fine optimization stage to screen the candidate feature set and the initial model hyperparameter range of the tunnel surrounding rock sample, thereby obtaining the optimal surrounding rock feature subset and the corresponding optimal model hyperparameters. Based on the optimal subset of surrounding rock features and the optimal model hyperparameters, a final surrounding rock grade identification model is constructed, and the final surrounding rock grade identification model is used to classify the surrounding rock of the tunnel. The feature selection stage includes: The energy factor of the initial lemming population is set to a fixed value greater than the first threshold, so that all lemming individuals in the initial lemming population only perform exploratory behavior to update their location. The initial lemming population is updated and iterated through exploratory behavior and location. The feature selection scheme corresponding to the lemming individual with the highest fitness after each iteration is recorded as the current optimal feature selection scheme. When the current optimal feature selection scheme remains unchanged in multiple consecutive iterations, it is determined that the stage switching condition is met, the feature selection stage ends, the current optimal feature selection scheme is output, and the parameter fine optimization stage begins. The parameter fine-tuning stage includes: Extract the feature subset and model hyperparameters corresponding to the current optimal feature selection scheme from the candidate feature set and the initial model hyperparameter range, and construct a temporary classifier; Based on the temporary classifier, determine the classification confusion degree that reflects the temporary classifier; Based on the classification confusion level, decisions are made regarding the continued exploration or development behavior of each lemming individual, the lemming's position is updated, and the optimal subset of surrounding rock features and the corresponding optimal model hyperparameters are obtained.

2. The tunnel surrounding rock classification method based on multi-dimensional data according to claim 1, characterized in that, When the current optimal feature selection scheme remains unchanged in multiple consecutive iterations, it is determined that the stage switching condition is met, including: Set up a queue to store the best historical feature selection schemes; After each iteration, the current optimal feature selection scheme is added to the queue; Determine whether the K most recent consecutive records in the queue are completely identical. If so, determine that the stage switching condition is met, where K is a preset positive integer. 3.The tunnel surrounding rock classification method based on multi-dimensional data according to claim 1, characterized in that, The step of determining the classification confusion degree reflecting the temporary classifier based on the temporary classifier includes: Extract validation set data from the corresponding feature subset; The temporary classifier is used to predict the surrounding rock grade for each sample in the validation set, and the probability value of each surrounding rock grade is obtained. Calculate the entropy value for the probability value of each surrounding rock grade to obtain a list of entropy values ​​containing all validation set samples; Based on the entropy value list, calculate the arithmetic mean of the entropy values ​​of all samples in the entropy value list; The arithmetic mean is normalized to obtain the classification perplexity of the temporary classifier.

4. The method of claim 1, wherein, Based on the classification confusion level, decisions are made regarding the continued exploration or development behavior of each lemming individual, the lemming's position is updated, and the optimal subset of surrounding rock features and the corresponding optimal model hyperparameters are obtained, including: Based on the classification perplexity, calculate the adaptive energy decay coefficient for the current iteration progress; Calculate the dynamic energy baseline value based on the adaptive energy decay coefficient and the current iteration progress; Based on the dynamic energy baseline value, an independent final energy value is generated for each lemming individual in the population; Based on whether the final energy value of each lemming individual is greater than the second threshold, control each lemming individual to perform exploratory or developmental behaviors to update its position; When the maximum number of iterations is reached, the optimal subset of surrounding rock features and the corresponding optimal model hyperparameters are obtained.

5. The tunnel surrounding rock classification method based on multi-dimensional data according to claim 4, characterized in that, The step of calculating the adaptive energy decay coefficient for the current iteration progress based on the classification perplexity includes: Obtain the preset lower bound of the attenuation coefficient, the preset upper bound of the attenuation coefficient, and the sensitivity coefficient; Calculate the negative correlation adjustment factor for confusion based on the classification confusion level; The adaptive energy attenuation coefficient is calculated using linear interpolation based on the preset lower bound of the attenuation coefficient, the preset upper bound of the attenuation coefficient, and the perplexity negative correlation adjustment factor.

6. The tunnel surrounding rock classification method based on multi-dimensional data according to claim 4, characterized in that, The step of calculating the dynamic energy baseline value based on the adaptive energy decay coefficient and the current iteration progress includes: Get the preset maximum number of iterations and energy scaling constant; Calculate the product of the adaptive energy decay coefficient and the proportion of the current iteration number to the maximum iteration number; Based on the product, calculate the radian value that reflects the remaining search space; The dynamic energy reference value is obtained by mapping the radian value and the energy scaling constant through trigonometric functions.

7. The tunnel surrounding rock classification method based on multidimensional data according to claim 4, characterized in that, The control of individual lemmings to perform exploratory or developmental behaviors to update their location includes: If the final energy value of an individual lemming is greater than the second threshold, it is determined that the lemming has performed an exploration behavior, and one of the behaviors, long-distance migration and burrowing, is randomly selected to be performed in order to update the part of the lemming position vector that represents the feature selection scheme and the model hyperparameters. If the final energy value of an individual lemming is not greater than the second threshold, it is determined that the lemming has performed development behavior, and one of the behaviors of foraging and avoiding predators is randomly selected to be performed in order to update the part of the lemming position vector that represents the model hyperparameters.

8. A tunnel surrounding rock classification system based on multidimensional data, characterized in that, include: The data acquisition module is used to acquire candidate feature sets, initial tunnel surrounding rock level classifiers, and initial model hyperparameter ranges for tunnel surrounding rock samples. The candidate feature set is a collection of M geological and rock mechanics quantitative indicators, which are key parameters that can be obtained from tunnel engineering geological exploration and laboratory tests and can reflect the stability of the surrounding rock. The population initialization module is used to initialize the lemming population by random sampling based on the candidate feature set and the range of initial model hyperparameters, so as to obtain the initial lemming population. The model optimization module is used to perform the feature screening stage and parameter fine optimization stage using the improved lemming optimization algorithm to screen the candidate feature set and the initial model hyperparameter range of the tunnel surrounding rock sample, and obtain the optimal surrounding rock feature subset and the corresponding optimal model hyperparameters. The model building module is used to construct the final surrounding rock grade identification model based on the optimal surrounding rock feature subset and the optimal model hyperparameters, and to use the final surrounding rock grade identification model to classify the surrounding rock of the tunnel. The feature selection stage includes: The energy factor of the initial lemming population is set to a fixed value greater than the first threshold, so that all lemming individuals in the initial lemming population only perform exploratory behavior to update their location. The initial lemming population is updated and iterated through exploratory behavior and location. The feature selection scheme corresponding to the lemming individual with the highest fitness after each iteration is recorded as the current optimal feature selection scheme. When the current optimal feature selection scheme remains unchanged in multiple consecutive iterations, it is determined that the stage switching condition is met, the feature selection stage ends, the current optimal feature selection scheme is output, and the parameter fine optimization stage begins. The parameter fine-tuning stage includes: Extract the feature subset and model hyperparameters corresponding to the current optimal feature selection scheme from the candidate feature set and the initial model hyperparameter range, and construct a temporary classifier; Based on the temporary classifier, determine the classification confusion degree that reflects the temporary classifier; Based on the classification confusion level, decisions are made regarding the continued exploration or development behavior of each lemming individual, the lemming's position is updated, and the optimal subset of surrounding rock features and the corresponding optimal model hyperparameters are obtained.

Images