A tunnel soft surrounding rock intelligent identification method

Abstract

This invention relates to the field of intelligent construction technology and discloses an intelligent identification method for weak surrounding rock in tunnels. The method involves collecting multi-source construction data and manually labeling it to construct a labeled dataset; performing geological strength softening weighting, local block density estimation, and robustness standardization to achieve local robustness normalization; and recombining the data into a low-dimensional representation consistent with geological dynamics through dynamic sensitive feature screening, kernel mapping embedding, and statistical correction. Based on sample-level geological complexity, the method dynamically adjusts the activation steepness and jump fusion ratio of the network to achieve adaptive forward propagation. A dual-head identification model is constructed, combining ordered classification, continuous risk, complexity constraints, and regularization terms to form a unified loss function, which is updated using an Adam optimizer. Finally, the method outputs the surrounding rock level and failure probability. This invention combines geological strength softening factors with Huber weighted center estimation and utilizes the contribution of local block density modulation deviation to stably estimate feature parameters and effectively resist abnormal interference.

Description

Technical Field

[0001] This invention relates to the field of intelligent construction technology, specifically a method for intelligent identification of weak surrounding rock in tunnels. Background Technology

[0002] During tunnel construction, the rapid and accurate identification of surrounding rock geological conditions is a core element determining support design, optimization of construction parameters, and construction safety. However, tunnel engineering typically faces complex and variable geological environments, including fault fracture zones, lithological abrupt change zones, high-stress zones, and high-water-bearing zones. These geological anomalies often result in a wide range of surrounding rock conditions, from intact hard rock to extremely soft loose rock, posing significant challenges to on-site identification work.

[0003] 1. Existing technologies typically use global mean standard deviation normalization or fixed quantile normalization directly for multi-source surrounding rock data with different physical dimensions and statistical distributions. This makes it difficult to cope with bimodal, multimodal, and local outlier distributions caused by fault fracture zones and lithological abrupt change zones. It can easily lead to weak and anomalous samples dominating the overall statistical center or fail to reflect the local sample density differences under different geological conditions.

[0004] 2. When dealing with the multi-source heterogeneous characteristics of tunnel surrounding rock, existing technologies either directly input all the original features into the network, leading to redundancy and noise accumulation, or use linear dimensionality reduction methods such as principal component analysis, resulting in the loss of nonlinear coupling relationships and physical interpretability.

[0005] 3. Existing deep neural networks use the same network depth and activation function form for all surrounding rock samples, but the requirements for nonlinear mapping ability differ significantly between intact hard rock and extremely soft loose material. Simple samples are prone to overfitting while complex samples are not adequately represented.

[0006] 4. Existing methods either only output discrete surrounding rock levels without providing continuous risk quantification, or ignore the ordered properties of surrounding rock levels and treat them as ordinary classifications, and lack physical constraints on geological complexity coefficients. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides an intelligent method for identifying weak surrounding rock in tunnels, thereby resolving the problems mentioned in the background section.

[0008] To achieve the above objectives, this invention provides the following technical solution: an intelligent identification method for weak surrounding rock in tunnels, comprising the following steps: S1. Collect multi-source data related to the surrounding rock during tunnel construction, manually label the surrounding rock level and damage risk, and construct a labeled training dataset. S2. Perform geological strength index softening weighting, local block density scale estimation and robust standardization transformation on the training dataset, and robust local normalization processing on multi-source heterogeneous data of tunnel surrounding rock. S3. Through dynamic sensitive feature screening, kernel mapping embedding, and statistical correction output, the normalized features are reorganized into a low-dimensional representation with geodynamic consistency. S4. Based on the sample-level geological complexity coefficient, dynamically adjust the activation steepness and jump fusion ratio of each layer to achieve sample-level adaptive forward propagation; S5. Construct an intelligent rock identification model that includes an ordered classification head and a continuous probability output head. Combine ordered classification loss, continuous risk loss, complexity constraints and weight regularization to construct a unified loss function, and use the Adam optimizer for backpropagation update. S6. Utilize the intelligent rock identification model to output the rock level identification results and failure probability prediction values.

[0009] Furthermore, the multi-source data related to the surrounding rock in S1 include geological exploration data, construction response data, and field monitoring data.

[0010] Furthermore, S2 specifically refers to: S21. Huber weighted center estimation based on softening of geological strength index: First, construct the softening factor of geological strength index for each sample based on the rock quality index Q value and uniaxial compressive strength. Then, combine the factor with Huber robustness loss to make a stable estimate of the center position of each feature. S22. Adaptive scaling estimation based on local block density: First, estimate the local block density around each sample in the multidimensional feature space, and then use the local block density to adjust the contribution of absolute bias to obtain adaptive scaling parameters that better match the distribution of the surrounding rock block structure. S23. Robust local normalization output: After obtaining the robust center estimate and adaptive scale parameter for each feature, all original features are translated and scaled to obtain the standardized input for network training.

[0011] Furthermore, S3 specifically refers to: S31. Dynamic sensitive feature screening: Evaluate the statistical dependence strength between each input feature and the surrounding rock level label and failure probability label, and then select the feature with the highest comprehensive sensitivity to enter the embedding process. S32, Kernel Mapping Embedding: Low-dimensional embedding is constructed by superimposing kernel mapping and linear guidance. The Gaussian kernel captures global nonlinear similarity, while dynamic sensitive features are used as linear guidance terms, so that each embedding dimension has both physical orientation and retains complex coupling relationships. S33. Reorganize the statistical correction output of the features, and perform zero mean and unit variance correction on each embedding dimension to obtain a scale-consistent reorganized feature matrix.

[0012] Furthermore, S4 specifically refers to: S41. Geological complexity coefficient generation: This invention compresses the burial depth and groundwater flow rate into a geological complexity coefficient between 0 and 1, which is used to uniformly adjust the nonlinear intensity and jump fusion ratio of each layer. S42, Linear mapping and perturbation injection of geological stratification units: linear mapping is performed on the input features of each layer, and a random perturbation term with adjustable intensity is injected according to the complexity of the sample. S43. Geological complexity modulated activation: The steepness of the activation curve of each layer is dynamically adjusted according to the geological complexity coefficient. By introducing an adaptive activation function that modulates complexity, each sample has a nonlinear mapping intensity that matches its own geological complexity. S44, Gated Jump Fusion Output: The geological complexity coefficient is used to gated the fusion of two pathways, so that different samples can automatically select a more suitable effective depth. S45. The surrounding rock level is output in an ordered manner, using an ordered classification output method to generate the probability distribution of the surrounding rock level based on the final implicit feature vector. S46, Destruction probability output, for the first... For each sample, the final latent feature vector and weight vector are linearly mapped, and then compressed to between 0 and 1 using the Sigmoid function to obtain the predicted failure probability value of the sample. The larger the value, the higher the risk of surrounding rock instability or failure.

[0013] Furthermore, S5 specifically refers to: S51. Calculation of ordered classification loss of surrounding rock level: The ordered classification loss of surrounding rock level is constructed using the negative log-likelihood form. S52. Calculation of regression loss for probability of destruction: The regression loss for probability of destruction is constructed using the mean square error method. S53. Geological complexity constraint loss calculation: constrain the unreasonable overestimation of the relationship between the geological complexity coefficient and the actual surrounding rock level. By comparing samples one by one and penalizing the excess, the complexity coefficient is made to be basically consistent with the weakness of the surrounding rock. S54. Construction of total loss and parameter update: The multiple losses are summed according to their weights to form a unified total loss, and the Adam optimizer is used to update all trainable parameters.

[0014] Furthermore, S31 specifically refers to: S311. For each training sample, read the true label of the surrounding rock level and the true label of the failure probability. S312, For each feature ,Will and The average is calculated to obtain the comprehensive mutual information score. , The larger the value, the more sensitive the feature is to both the identification of the surrounding rock grade and the prediction of instability risk, and the more worthwhile it is to retain it in subsequent feature recombination. S313. Calculate the 23 features according to their comprehensive mutual information scores. Sort by largest to smallest, then select the first... These features form a dynamic sensitive feature index set. , The feature order remains fixed in subsequent steps and is used to construct the correspondence between linear guiding terms and embedding dimensions.

[0015] Furthermore, S32 specifically refers to: S321, using the normalized characteristic matrix Based on this, calculate the pairwise Euclidean distance between all training samples, and take the median of all pairwise Euclidean distances. Multiples as kernel width parameter ; S322, Set up the dynamic sensitive feature index The features in the data are ranked from highest to lowest based on their comprehensive mutual information scores. , No. The first embedding dimension corresponds to the first... One dynamic sensitive feature index This ensures that each embedded dimension has a clear physical orientation; S323, For each embedding dimension , with the first The normalized value sequence of each dynamic sensitive feature on all training samples is used as the target response, and the kernel mapping coefficient of this embedding dimension is solved by kernel ridge regression. S324. For each sample and embedding dimension First, calculate the weighted sum of kernel similarities between the sample and all training samples to obtain the kernel response.

[0016] Furthermore, S33 specifically refers to: S331, For each embedding dimension The mean and standard deviation of the original recombined values ​​of all training samples in this dimension are calculated. S332. Arrange all corrected embedding values ​​in dimensional order to form a reconstructed feature matrix. , dimension , As input data for the geological stratification adaptive deep neural network, and the first The first sample The values ​​of each embedding dimension are denoted as... , No. The recombined feature vector of each sample is denoted as . .

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0018] 1. This invention proposes a robust local normalization method, which combines the softening factor of the geological strength index with Huber weighted center estimation and utilizes the contribution of the absolute deviation of local block density modulation. This method can stably estimate the feature center and scale parameters, effectively resist the interference of anomalous samples, and preserve the natural fluctuation range of typical geological models.

[0019] 2. This invention proposes a nonlinear feature recombination method based on the consistency of geodynamics. First, it filters dynamic features that are sensitive to both the surrounding rock level and the failure probability through mutual information. Then, it uses kernel mapping to capture global nonlinear similarity and superimposes explicit linear guiding terms, so that each embedded dimension has both a clear physical orientation and can retain complex geodynamic relationships.

[0020] 3. This invention proposes a geological stratification adaptive deep neural network, which generates sample-level geological complexity coefficients based on burial depth and groundwater flow, dynamically adjusts the steepness of the activation function of each layer, and combines nonlinear responses with direct information from the previous layer according to complexity ratio through gated jump fusion, thereby achieving sample-level nonlinear intensity and effective depth adaptation.

[0021] 4. This invention jointly optimizes the ordered classification loss and regression loss, ensures the order relationship of levels by increasing the threshold, outputs the failure probability independently, and introduces geological complexity constraint loss to penalize the part of the complexity coefficient that exceeds the upper limit corresponding to the real level, so that the complexity mapping learned by the model conforms to the real evolution law of the weakness of the surrounding rock. Attached Figure Description

[0022] Figure 1 This is a flowchart of the present invention; Figure 2 This is an example diagram of Huber weighted center estimation; Figure 3 This is a diagram illustrating the dynamics-sensitive feature screening and mutual information score ranking. Detailed Implementation

[0023] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0024] like Figure 1 As shown, a method for intelligent identification of weak surrounding rock in tunnels is disclosed, specifically as follows: S1. Multi-source data collection and labeling of surrounding rock in tunnels: Collect multi-source data related to surrounding rock during tunnel construction, manually label the surrounding rock level and damage risk, and construct a labeled training dataset. During tunnel construction, it is necessary to collect three types of multi-source data simultaneously: geological survey, construction response, and on-site monitoring, in order to build a standard sample library for model training.

[0025] Geological survey data comes from the tunnel face and advanced geological forecasts, including longitudinal wave velocity, rock quality indicators, and uniaxial compressive strength obtained from core drilling tests, as well as geophysical data such as seismic wave reflection energy and resistivity. Construction response data is recorded in real time by sensors on drilling rigs, drill and blast operation recorders, and support construction equipment, covering key parameters such as drilling rate, single-cycle advance time, peak velocity of blasting vibration, anchor bolt axial force, and shotcrete stress. On-site monitoring data comes from automated monitoring systems deployed on the tunnel walls and arch, including arch settlement rate, peripheral convergence rate, microseismic event rate, face texture characteristics (such as fragmentation, joint density, average block size, roughness index, and texture entropy), burial depth, groundwater flow, rock quality index Q value, elastic wave impedance, and acoustic travel time.

[0026] Specifically, the tunnel surrounding rock input data includes geological exploration information, construction response information, and on-site monitoring information, totaling 23 features, arranged in a fixed order as follows: borehole longitudinal wave velocity (m / s), rock quality index (percentage), uniaxial compressive strength (MPa), drilling rate (m / min), single-cycle advance time (min / cycle), peak velocity of blasting vibration (mm / s), anchor bolt axial force (kN), shotcrete stress (MPa), seismic wave reflection energy (dimensionless), resistivity (ohm-meter), microseismic event rate (times / min), and crown settlement rate (…). The parameters include: millimeters per day, peripheral convergence rate (millimeters per day), fragmentation in the face texture features (dimensionless), joint density (strips per square meter), average block size (meter), roughness index (dimensionless), texture entropy (dimensionless), burial depth (meter), groundwater flow rate (liters per minute per 10 meters), rock quality index Q value (dimensionless), elastic wave impedance (kilograms per square meter per second), and sound wave travel time (microseconds per meter). Among these, the rock quality index Q value (or Barton Q value) is an engineering indicator that comprehensively reflects the quality of the rock mass; the higher the value, the better the rock mass quality.

[0027] The above 23 features are arranged in a fixed order to form the original input vector of each sample. The acquisition frequency is uniformly aligned to each construction cycle according to the actual change rate of each parameter, ensuring that each cycle corresponds to a complete set of surrounding rock condition records.

[0028] For each collected sample, an experienced geological engineer needs to perform a dual labeling process, combining on-site excavation findings, laboratory test results, and long-term monitoring data: First, a rock mass level label, which, according to relevant engineering standards, classifies the surrounding rock into six levels, from intact hard rock to extremely soft loose material, represented by numbers 1 to 6. The higher the value, the weaker and more fragmented the surrounding rock. Second, a failure probability label, which is a continuous value between 0 and 1, reflecting the risk level of instability or failure of the cross-section under current geological conditions. This label can be determined by a comprehensive assessment based on the measured rate of crown settlement and surrounding convergence, the frequency of microseismic events, and the state of groundwater seepage.

[0029] After the annotation is completed, the 23 original features, surrounding rock level labels, and failure probability labels of all samples are stored together in the training dataset.

[0030] To ensure the model's adaptability to complex geological conditions, the training data should cover scenarios with different burial depths, different groundwater conditions, different lithological abrupt change zones, and fault fracture zones as much as possible, with a total number of samples of no less than several thousand rings.

[0031] S2. Robust local normalization of multi-source heterogeneous data of tunnel surrounding rock: The training dataset is subjected to geological strength index softening weighting, local block density scale estimation and robust standardization transformation, and robust local normalization processing of multi-source heterogeneous data of tunnel surrounding rock. The heterogeneous data of tunnel surrounding rock from multiple sources show significant differences in physical dimensions, numerical ranges, and statistical distributions. Furthermore, fault fracture zones, lithological abrupt change zones, and local high water-bearing areas are prone to bimodal, multimodal, and local outlier phenomena. If mean standard deviation is directly used for normalization, abnormal fluctuations in weak surrounding rock sections may dominate the overall statistical center. If fixed quantile normalization is directly used, it is difficult to reflect the local sample density differences corresponding to different geological conditions.

[0032] This invention achieves stable alignment of multi-source heterogeneous features through softening weighting of geological strength indices, local block density scale estimation, and robust normalization transformation. The specific steps are as follows: S21. Huber-weighted center estimation based on softening of geological intensity indicators. The original characteristics of the surrounding rock contain typical samples, weak and abnormal samples, and local extreme samples. If the mean is used directly as the center location, it is easy to be biased by fault fracture sections or high water-bearing disturbance sections. If the median is used directly as the unique center estimate of all characteristics, it will weaken the response to the distribution of the main sample.

[0033] This invention first constructs a softening factor for the geological strength index of each sample based on the rock quality index Q value and uniaxial compressive strength. Then, it combines this factor with Huber robustness loss to perform a stable estimation of the center position of each feature. The specific steps are as follows: 1> Let the total number of training samples be... The original feature matrix is ​​denoted as , Used to carry all original surrounding rock input data, with dimensions of , among which, the The first sample The original values ​​of each feature are denoted as , This represents the sample index, with a value range of [value range missing]. arrive ; This represents the feature index, with a value range of 100. arrive .

[0034] From the original feature matrix Read the first The rock quality indicators Q-value and uniaxial compressive strength of each sample, among which, the first The rock quality index Q value of each sample is denoted as: This is the 21st characteristic, characterizing the integrity of the rock mass and the degree of joint development; The uniaxial compressive strength of each sample is denoted as . The third characteristic, measured in megapascals (MPa), characterizes the rock's ability to resist uniaxial compressive failure.

[0035] Furthermore, according to and Calculate the first The geological strength index of each sample is the softening factor. While preserving information about weak surrounding rock, the calculation method avoids weak and anomalous samples dominating the overall statistical center. ,in, This represents the function that takes the maximum value. Values ​​located at and Between these values, the smaller the value, the weaker and more fragmented the surrounding rock corresponding to the sample, and the more likely it is to experience local abnormal fluctuations. In this case, the contribution of the sample to the central estimation decreases accordingly. The larger the value, the more stable the surrounding rock corresponding to the sample, and the higher its contribution to the central estimation.

[0036] 2> For each feature Extract the original values ​​of all training samples for this feature. And calculate the interquartile range of this feature; No. The Huber threshold for each feature is denoted as . The interquartile range of this feature can be taken as 1 / 3. times, Used to distinguish between normal fluctuations and abnormal deviations; a larger value indicates greater tolerance for larger deviations.

[0037] Furthermore, for each feature Perform Huber weighted center estimation; No. Robust center estimates of the features are denoted as This is used to characterize the stable center location of the feature in the training samples and to provide a translation benchmark for subsequent scale estimation and normalization transformation.

[0038] In practical implementation, the first step can be to... The median of each feature is used as the initial center value. Then, the residual is calculated for each sample, which is the difference between the original value of the sample and the current center value. Then, when the absolute value of the residual is not greater than... When the absolute value of the residual is greater than a certain value, the sample is considered a sample with normal fluctuations. At that time, the sample was considered to have a large bias and was processed according to... The ratio of the residual absolute value to the effective weight is reduced; that is, for residual absolute values ​​greater than a threshold, the effective weight is reduced. Samples with excessive deviation are considered outliers. To reduce the impact of these samples on the center estimation, the effective weights of the samples are set to... ,in It is the absolute value of the residual, because Therefore, this is equivalent to proportionally reducing the weight of the sample; the larger the residual, the smaller the weight. For example, if... A certain sample residual Then its weight is reduced to ,like Then the weight is reduced to ; Then, the softening factor of the geological strength index. Multiply by Huber's effective weights to obtain the first... The sample at the th The comprehensive weights in the feature center estimation are then used to update the center values, i.e., the comprehensive weights. Repeat the above process until the change in the center value between two consecutive values ​​is less than a preset threshold (e.g., (or reach the maximum number of iterations (e.g., 20 times)).

[0039] In practice, Huber's effective weights are determined directly by comparing the residuals with a threshold, excluding the softening factor of the geological strength index. The specific rules are as follows: when In this case, the sample is considered a regular point in Huber's sense, with effective weights. ; when At that time, effective weight ; in , This is the center estimate for the current iteration.

[0040] 3> Output robust center estimates for each of the 23 features. These robust center estimates collectively constitute the center parameter set, which is used for subsequent local block density scaling and robust local normalization output.

[0041] It should be noted that although different surrounding rock samples are all training samples, weak fractured sections are usually accompanied by high fluctuations and strong disturbances. If their dominance over the global statistical center is not reduced, the normalization benchmark of all subsequent features will deviate from the normal geological background. By combining the softening factor of the geological strength index with Huber robustness estimation, we can obtain a more stable feature center position that is more consistent with the main geological state while retaining the information of weak surrounding rocks.

[0042] In one embodiment, such as Figure 2 The example of Huber-weighted center estimation is shown. The left figure shows the scatter plot relationship between sample residuals (original value minus initial median) and Huber effective weights, taking P-wave velocity characteristics as an example. When the absolute value of the residual exceeds the threshold δ, the weight decays. The right figure further multiplies the Huber weights by the softening factor of the geological strength index to obtain the comprehensive weight. The color depth represents the size of the softening factor of the geological strength index. This proves that the Huber robustness mechanism can automatically reduce the contribution of samples far from the center, and the comprehensive weight of weak samples (small softening factor of geological strength index) is lower, thus obtaining a more robust center estimation.

[0043] S22. Adaptive scaling estimation based on local block density Estimating the scale solely based on the global absolute bias can easily lead to treating low-density outlier samples and high-density typical samples equally, resulting in an overestimated scale parameter and thus compressing the separability between typical samples.

[0044] This invention first estimates the local block density around each sample in a multidimensional feature space, and then uses the local block density to adjust the contribution of the absolute bias, thereby obtaining adaptive scale parameters that better match the distribution of the surrounding rock block structure. The specific steps are as follows: 1> To avoid the dimensional differences of the 23 features directly affecting distance calculation, a temporary alignment feature is first constructed for density statistics only. Specifically, for each sample, the alignment feature is... Each feature is used to subtract the corresponding robust center estimate from the original value. Divide by the first The temporary alignment values ​​are obtained after the interquartile range of each feature; the 23 temporary alignment values ​​together form the first set of values. A temporary alignment vector for each sample, which is used for multidimensional distance and local density statistics.

[0045] Furthermore, the pairwise Euclidean distances are calculated between the temporary alignment vectors of all training samples, and the median of all pairwise Euclidean distances is taken as the reference distance. The local block density statistical radius is denoted as A reference distance can be taken. times, Used to control the neighborhood range, if the radius is too large, it is easy to mix different geological models in statistics, and if the radius is too small, it is easy to make the local density unstable.

[0046] 2> For each sample Centered on the temporary alignment vector of the sample, within a radius The number of neighborhood samples within the 23-dimensional hypersphere is statistically analyzed to obtain the local block density of that sample. , The neighborhood count characterizes the typicality of the feature pattern in which the sample is located. The larger the value, the closer the sample is to a high-density common geological state. The smaller the value, the more likely the sample is to be in a marginal state, a transitional state, or a local anomaly state. To ensure statistical stability, the neighborhood count includes the sample itself.

[0047] 3> Take the maximum value among all local block densities, denoted as . Then for each sample To construct the density modulation factor, one can use... The density modulation factor, calculated by this method, increases with the increase of local block density. This is used to amplify the bias contribution of high-density typical samples and suppress the bias contribution of low-density abnormal samples.

[0048] 4> For each feature Calculate the robust center estimate relative to all samples. The absolute deviation of each sample is calculated, and the absolute deviation of each sample is multiplied by the corresponding density modulation factor to obtain the density-modulated deviation sequence. Then, the median of the density-modulated bias sequence is taken to obtain the first... Adaptive scaling parameters for each feature , Used to depict the first Features surrounding robust center estimation It determines the stable fluctuation range and provides a scaling benchmark for subsequent normalization transformations.

[0049] In practical implementation, when the adaptive scaling parameter of a certain feature... Too small (e.g., less than) When ), The threshold is set to a preset lower limit to prevent excessive numerical amplification during subsequent normalization.

[0050] It should be noted that typical surrounding rock conditions usually form local high-density blocks in the characteristic space, while local anomalous states, transitional states, and occasional disturbance states are often located in low-density areas. By contributing the absolute deviation of local block density modulation, the scale estimation can better reflect the natural fluctuation range of typical geological models, thereby improving the ability of the normalization results to preserve the main geological structure.

[0051] S23, robust local normalization output After obtaining the robust center estimate and adaptive scaling parameters for each feature, all original features are translated and scaled to obtain standardized input for network training. This eliminates dimensional differences while preserving the local statistical structure among the surrounding rock samples. The specific steps are as follows: 1> For each sample Each feature , the original value Robust center estimation by subtracting the corresponding feature Then divide by the adaptive scaling parameter of the corresponding feature. With the minimum constant The sum of these values ​​yields the normalized value. Among them, the minimum constant Used to prevent the denominator from being 0, with the following example values: .

[0052] 2> Rearrange all normalized values ​​according to the original feature order to obtain the normalized feature matrix. , dimension , The The line represents the first The 23-dimensional normalized feature vector of each sample is the input data for the subsequent nonlinear feature recombination stage.

[0053] It should be noted that tunnel surrounding rock data may exhibit non-Gaussian distributions (bimodal, multimodal, outliers) due to fault fracture zones, abrupt lithological changes, and high water-bearing areas. Conventional methods may distort the normalization results. However, this invention preserves the main geological state through robust centers and uses density modulation to make the scale reflect the natural fluctuations of typical patterns, avoiding the compression and separability of outlier samples. This makes it more suitable for subsequent nonlinear feature recombination and network training.

[0054] It should also be noted that, except during the training phase, validation samples, test samples, and new tunnel cross-section samples will not be re-estimated. and Instead, it directly calls the data saved during the training phase. and Perform the same translation and scaling to ensure that the input scale is consistent between the training and inference phases.

[0055] In a method that facilitates engineering implementation, if a certain original feature is missing, the robust center of that feature on the training set can be used to estimate its value first. Missing values ​​are filled in and then normalized to prevent subsequent nonlinear feature recombination and network forward propagation from being interrupted by missing values.

[0056] S3. Nonlinear feature recombination based on geodynamic consistency: Through dynamic sensitive feature screening, kernel mapping embedding and statistical correction output, normalized features are recombined into a low-dimensional representation with geodynamic consistency. Significant nonlinear coupling and physical derivation relationships exist among the multi-source heterogeneous data characteristics of tunnel surrounding rock. For example, drilling rate, blasting vibration response, anchor bolt stress state, and surrounding rock quality indicators often reflect the same type of geological dynamic mechanism. If the 23 normalized features are directly input into the network, redundant information and weakly correlated noise are easily retained at the same time. If a linear dimensionality reduction method is directly adopted, it is difficult to maintain the physical interpretability of the surrounding rock evolution mechanism.

[0057] This invention reconstructs 23-dimensional normalized features into a low-dimensional representation with geodynamic consistency through dynamically sensitive feature screening, kernel mapping embedding, and statistical correction output. The specific steps are as follows: S31, Screening of Dynamically Sensitive Features The surrounding rock grade label and the failure probability label reflect the surrounding rock classification results and the level of instability risk, respectively. The input features that are truly dynamically sensitive usually have a statistical dependence on both of these types of outputs.

[0058] This invention evaluates the statistical dependency strength between each input feature and the surrounding rock level label and the failure probability label, and then selects the feature with the highest overall sensitivity to enter the embedding process. The specific steps are as follows: 1> For each training sample, read the true label of the surrounding rock level and the true label of the failure probability. Specifically, the first... The true label of the surrounding rock grade for each sample is denoted as: The value ranges from 1 to 6, corresponding to six levels of surrounding rock categories, from intact hard rock to extremely soft granular material. No. The true label of the probability of destruction for each sample is denoted as . The value ranges from 0 to 1, representing the continuous risk level of instability or failure of the surrounding rock.

[0059] For each feature Calculate the mutual information between the feature and the true label of the surrounding rock grade, and the mutual information between the feature and the true label of the failure probability. Specifically, the first... The mutual information between each feature and the true label of the surrounding rock grade is denoted as . This is used to measure the sensitivity of this feature to the determination of the surrounding rock grade; No. The mutual information between each feature and the true label of the probability of destruction is denoted as . This is used to measure the sensitivity of the feature to the characterization of instability risk.

[0060] In practical implementation, mutual information estimation is accomplished using kernel density estimation. Specifically, kernel density smoothing is first applied to the individual feature distribution, label distribution, and joint distribution respectively (the first step is to perform kernel density smoothing). The joint probability density function of each feature and its corresponding label is used to approximate the mutual information based on the smoothed probability density. The kernel density estimation bandwidth can be determined according to the Silverman rule to balance distribution smoothness and statistical stability. In addition, for discrete labels such as rock grade, the labels can be regarded as ordered discrete variables first, and then the discrete-continuous hybrid mutual information estimation method can be used to complete the statistical dependence measure.

[0061] 2> For each feature ,Will and The average is calculated to obtain the comprehensive mutual information score. , The larger the value, the more sensitive the feature is to both the identification of the surrounding rock grade and the prediction of instability risk, and the more worthwhile it is to retain it in subsequent feature recombination.

[0062] 3> Analyze the 23 features according to their comprehensive mutual information scores. Sort by largest to smallest, then select the first... These features form a dynamic sensitive feature index set. , The feature order in the model remains fixed in subsequent steps and is used to construct the correspondence between the linear guiding term and the embedding dimension; preferably, Take a positive integer less than 23, for example .

[0063] In one embodiment, such as Figure 3 As shown, dynamic sensitive features were screened, and mutual information scores were analyzed and ranked. The horizontal axis represents the comprehensive mutual information score (i.e., the average of the mutual information between the feature and the surrounding rock level label and the mutual information with the failure probability label, which is dimensionless), and the vertical axis represents the names of the 23 features. The length of the horizontal bar represents the comprehensive sensitivity of each feature to surrounding rock identification and risk prediction. The red dashed line marks the selection threshold of the top 10 features.

[0064] S32, kernel mapping embedding The geological dynamics of surrounding rocks are usually not simple linear relationships, so it is necessary to construct a global nonlinear similarity using all normalized features. At the same time, in order to ensure that the embedding results are still controlled by the dominant physical features, dynamic sensitive features need to be introduced as explicit linear guidance.

[0065] This invention constructs low-dimensional embeddings using a combination of kernel mapping and linear guidance. It captures global nonlinear similarity through a Gaussian kernel and utilizes dynamically sensitive features as linear guidance terms, ensuring that each embedding dimension has both physical orientation and retains complex coupling relationships. The specific steps are as follows: 1> Using the normalized characteristic matrix Based on this, calculate the pairwise Euclidean distance between all training samples, and take the median of all pairwise Euclidean distances. Multiples as kernel width parameter , This value controls the sensitivity of nonlinear similarity to changes in distance. A larger value indicates a slower decay of similarity, while a smaller value indicates a faster decay of similarity.

[0066] Furthermore, based on the kernel width parameter A Gaussian kernel matrix is ​​constructed between training samples to express the global nonlinear similarity between samples. The elements in the Gaussian kernel matrix are determined by the Euclidean distance between the normalized feature vectors of two samples. The closer the distance, the larger the kernel value, and the farther the distance, the smaller the kernel value.

[0067] In practical implementation, the normalized feature matrix is ​​used. ( Based on this, calculate the pairwise Euclidean distance between all training samples. Take 0.5 times the median of all distances as the kernel width parameter. Gaussian kernel matrix for A matrix of dimension, whose first dimension is... Line number Column elements ,in, Indicates the first The sample and the first The Euclidean distance between the normalized feature vectors of each sample. Indicates the first The normalized feature vector of each sample. Indicates the first The normalized feature vector of each sample. This represents the L2 norm.

[0068] In one embodiment, for example, if there are 3 samples, and the pairwise distances are d... 12 =2,d 13 =3,d23 =2.5, the median is 2.5. ,but .

[0069] 2> Index the dynamic sensitive features The features in the data are ranked from highest to lowest based on their comprehensive mutual information scores. , No. The first embedding dimension corresponds to the first... One dynamic sensitive feature index This allows each embedded dimension to have a clear physical orientation, rather than being generated entirely by black-box mapping.

[0070] 3> For each embedding dimension , with the first The normalized value sequence of each dynamic sensitive feature across all training samples is used as the target response. Kernel ridge regression is employed to solve for the kernel mapping coefficients of this embedding dimension. Specifically, the... The kernel mapping coefficient vector corresponding to each embedding dimension is denoted as . , dimension , of which The coefficients are denoted as , used to measure the first The training sample pair for the th training sample The contribution of each embedded dimension kernel response.

[0071] In practical implementation, a ridge regularization term can be added to the kernel matrix before solving the system of linear equations. The ridge regularization coefficients can be taken as follows: arrive A constant is used between these constants to improve solution stability; specifically, for each embedding dimension... ( arrive ), with the first Normalized value sequence of each dynamic sensitive feature on all training samples ( The kernel ridge regression solution is represented as a dimensional vector (the target response). ,in, Ridge regularity coefficients (e.g.) ), It is an identity matrix.

[0072] 4> For each sample and embedding dimension First, calculate the weighted sum of kernel similarities between the sample and all training samples to obtain the kernel response. ,in, Indicates the first The kernel mapping coefficient vector corresponding to each embedding dimension The first in The component is used to measure the first component. The training sample pair for the th training sample The contribution strength of the kernel response in each embedded dimension; then the sample is applied to the corresponding dynamic sensitive feature. Normalized values ​​on Combined with guiding coefficient , as a linear leading term Adding a nuclear response, forming the first The original recombination value of each embedding dimension , represented as Among them, the guiding coefficient Used to control the strength of the dominant feature's control over this embedding dimension. A constant value between 0.1 and 0.5 is acceptable.

[0073] It should be noted that, except during the training phase, for new tunnel cross-section samples, the kernel mapping coefficients are not recalculated; instead, the normalized features and kernel width parameters of the training reference samples saved during the training phase are used. kernel mapping coefficients for each embedding dimension First, calculate the kernel similarity between the new sample and all training reference samples. Then, superimpose the linear guiding term in the same way as in the training phase to obtain the original recombinant values ​​of each embedding dimension of the new sample.

[0074] It should also be noted that the true dynamic model of the surrounding rock is usually a combination of "nonlinear overall similarity + control by a few dominant physical features". Relying solely on kernel mapping can easily lose physical interpretability, while relying solely on a few explicit features can hardly reflect complex nonlinear coupling. By superimposing the two, we can preserve complex geodynamic relationships while maintaining the engineering interpretability of the embedding results.

[0075] S33, Recombination Feature Statistical Correction Output Although different embedding dimensions all come from the superposition of kernel mapping and linear guidance, their numerical range and fluctuation amplitude are usually different. If they are directly input into subsequent networks, some embedding dimensions may dominate the gradient due to their large scale.

[0076] This invention applies zero-mean and unit-variance correction to each embedding dimension to obtain a scale-consistent reconstructed feature matrix. The specific steps are as follows: 1> For each embedding dimension Calculate the mean and standard deviation of the original recombined values ​​of all training samples in this dimension. Specifically, the first... The mean of each embedding dimension is denoted as . , used to complete center alignment; No. The standard deviation of each embedded dimension is denoted as . It is used to achieve scale uniformity.

[0077] Furthermore, for each sample and embedding dimension Subtract the original recombination value of the embedding dimension Divide by With the minimum constant The sum of these values ​​yields the statistically corrected embedding value. ; Among them, when If the value is too small, a lower limit can be set, for example... To avoid numerical instability.

[0078] 2> Arrange all corrected embedding values ​​in dimensional order to form a reconstructed feature matrix. , dimension , As input data for the geological stratification adaptive deep neural network, and the first The first sample The values ​​of each embedding dimension are denoted as... , No. The recombined feature vector of each sample is denoted as . .

[0079] It should be noted that, except during the training phase, validation samples, test samples, and new tunnel cross-section samples are not re-statistically analyzed. and Instead, it directly uses the data saved during the training phase. and Apply the same statistical corrections.

[0080] S4. Forward propagation calculation of geological stratification adaptive deep neural network: Based on sample-level geological complexity coefficient, the activation steepness and jump fusion ratio of each layer are dynamically adjusted to achieve sample-level adaptive forward propagation. The identification of surrounding rock spans a wide range of geological conditions, from intact hard rock to extremely soft granular material. Different samples have different requirements for network depth and nonlinear strength. Hard rock samples can usually be identified by shallow relationships, while soft surrounding rock samples are often affected by high ground stress, high water content, plastic rheology and large deformation effects at the same time, requiring stronger nonlinear mapping capabilities. If a fixed depth and fixed activation intensity are used for all samples, it is easy to have both overfitting of simple samples and insufficient representation of complex samples.

[0081] This invention achieves sample-level adaptive forward propagation by dynamically adjusting the activation steepness and jump fusion ratio of each layer through a sample-level geological complexity coefficient. The specific steps are as follows: S41, Generation of Geological Complexity Coefficient Burial depth and groundwater flow are two key factors affecting the complexity of surrounding rock. Greater burial depth typically corresponds to a stronger geostress environment, while greater groundwater flow usually corresponds to more pronounced seepage, softening, and disturbance effects. This invention compresses these two factors into a geological complexity coefficient between 0 and 1, used to uniformly adjust the nonlinear intensity and jump fusion ratio of each layer. The specific steps are as follows: From the normalized feature matrix The 19th and 20th features were extracted and used as the normalized values ​​for burial depth and groundwater flow, respectively. The normalized value of the burial depth for each sample is denoted as... It characterizes the strength of the geostress environment and burial conditions; No. The normalized groundwater flow rate of each sample is denoted as . It characterizes the extent to which groundwater activity affects softening, seepage, and disturbance.

[0082] Furthermore, let the burial depth weighting parameter be... The groundwater flow weighting parameter is The complexity bias parameter is ,and , and All are trainable scalars shared across the entire network, used to map burial depth and groundwater flow rate to a unified complexity metric; For each sample ,Will and Perform a linear combination, then compress it to between 0 and 1 using the sigmoid activation function to obtain the geological complexity coefficient. , represented as ; in, This represents the Sigmoid activation function; The closer the value is to 1, the more complex the surrounding rock environment of the sample; the closer the value is to 0, the simpler the surrounding rock environment of the sample.

[0083] In practical implementation, to improve training stability, the following can be done: Make a slight truncation, for example, limiting it to between 0.05 and 0.95, to avoid gradient propagation being too weak due to extreme gating ratios in the early stages of training.

[0084] It should be noted that the burial depth and groundwater flow These are two key conditions reflecting the complexity of the surrounding rock. The greater the burial depth, the higher the level of in-situ stress, and the more prone the surrounding rock is to plastic deformation and rockburst. Conversely, the greater the groundwater flow, the more easily the surrounding rock softens, seeps, and becomes unstable. During training, the burial depth parameter is considered. and groundwater flow weight parameters It will adjust to a positive value through learning (or remain positive and initialized), thus linear combination The larger the value, the greater the burial depth and / or groundwater flow, and the more complex the geological environment. The Sigmoid function monotonically compresses this linear combination to a smaller value. Interval, therefore The closer to 1, the higher the complexity; the closer to 0, the lower the complexity.

[0085] S42, Linear Mapping of Geological Stratification Units and Disturbance Injection To enable the network to learn stable hierarchical mappings while adapting to local fluctuations in weak surrounding rock scenarios, this invention performs linear mapping on the input features of each layer and injects a random perturbation term with adjustable strength according to the sample complexity. The specific steps are as follows: 1> Define geological stratification units by Composed of sequentially connected layers ( Each layer contains three core operations: linear mapping, perturbation injection, and complexity-modulated activation, and finally, the output is fused through gated jump. The dimension of the hidden vector in each layer is 1. This is consistent with the dimensions of recombination features.

[0086] No. The sample enters the... The feature vectors before the layer are denoted as ,and Equivalent to the first The first sample The layer outputs a feature vector, which is then processed by the first layer. The output feature vector after layer processing is denoted as ,when hour, That is, the first The recombined feature vector of each sample, i.e. .

[0087] The first The first sample Layer output feature vector Enter the first Geological stratification unit, the first The layer weight matrix is ​​denoted as , dimension ;No. The layer bias vector is denoted as , dimension ; and These are all trainable parameters used to complete the linear mapping of this layer.

[0088] In practical implementation, the first step is to calculate the matrix multiplication and bias superposition. The sample at the th The linear response vector of the layer , Dimensions It is the direct input for the current layer activation calculation and gating fusion.

[0089] 2> To simulate uncertain disturbances in weak surrounding rock scenarios, a geological disturbance vector is generated for each sample and each layer. , Dimensions Each component can be sampled from a Gaussian distribution with a mean of 0 and a standard deviation of 0.01; Then, , Superimposed on the linear response vector Above, that is This enables complex samples to have a stronger perturbation modeling capability, while simple samples are only affected by weak perturbations.

[0090] It should be noted that, except during the training phase, no random perturbations are injected during the inference phase; that is, the perturbations are directly set... It is a zero vector to ensure the stability and repeatability of the output of the new tunnel cross-section samples.

[0091] S43, Geological Complexity Modulation Activation Different samples have different levels of complexity and different requirements for the steepness of the activation function. Complex samples require stronger nonlinearity to express the complex relationship under the combined effects of high stress, high water content and plastic deformation, while simple samples are more suitable for preserving a relatively stable linear approximation.

[0092] This invention dynamically adjusts the steepness of the activation curves of each layer based on the geological complexity coefficient. By introducing an adaptive activation function modulated by complexity, each sample has a nonlinear mapping intensity that matches its own geological complexity. The specific steps are as follows: 1> Let the first The basic steepness parameter of the layer is The complexity gain parameter is , and All are the first The layers share a trainable scalar, with initial values ​​of 0.5 and 0.2 respectively. in, Used to control the underlying nonlinear strength of the activation function Used to amplify the modulating effect of the geological complexity coefficient on the activation steepness.

[0093] 2> Regarding the first Layer linear response vector Each component in the algorithm undergoes element-wise adaptive activation, calculated as follows: ; in, Represents any single component in a linear response vector; Representing components The output after complexity modulation activation; This represents the hyperbolic tangent function, used to compress large inputs to a finite interval; When the geological complexity coefficient When the value is large, the activation curve is steeper, and the network has a stronger nonlinear expressive ability; when the geological complexity coefficient is large... When the activation curve is smaller, the activation curve is flatter, and the network is closer to a linear mapping.

[0094] Furthermore, the activation results of all components are combined in their original order to obtain the first... The sample at the th The nonlinear response vector of the layer has geological complexity adaptability, that is, under the same network parameters, different samples obtain activation outputs of different shapes, which represent the feature expression of the current layer after nonlinear transformation. The features of complex samples are strongly nonlinearly distorted to characterize complex mechanisms such as high stress and high water content, while simple samples retain a relatively stable feature manifold.

[0095] S44, Gated Jump Fusion Output Simply enhancing the nonlinear transformation is not enough to achieve effective depth selection at the sample level; adaptive allocation between the nonlinear response of the current layer and the directly connected path of the previous layer is also required.

[0096] This invention utilizes a geological complexity coefficient to perform gated fusion of two pathways, enabling different samples to automatically select a more suitable effective depth. The specific steps are as follows: 1> For the first The nth sample, the nth The nonlinear response vector of the first layer is proportionally fused with the input feature vector of the previous layer to obtain the first layer. Layer output feature vector The proportion of nonlinear transformation paths is: The proportion of directly connected paths is ; When the geological complexity coefficient When the coefficient is large, the current layer relies more on the nonlinear transformation result; when the geological complexity coefficient is large... When the current layer is smaller, it relies more on the direct connection information of the previous layer.

[0097] 2> The first Layer output feature vector As the first The input of each layer is traversed sequentially. The geological stratification unit is ultimately obtained as the first layer. The final latent feature vector of each sample after processing through all layers , Comprehensive characterization is required for identifying the load-bearing surrounding rock level and predicting the probability of failure.

[0098] It should be noted that the optimal network depth varies depending on the complexity of the surrounding rock samples. By using gated skip fusion, simple samples can retain more shallow representations to avoid overfitting, while complex samples can use more deep nonlinear representations to enhance the ability to characterize the complex mechanisms of weak surrounding rocks.

[0099] S45, orderly output of surrounding rock grade The surrounding rock grade is a discrete label with a natural order relationship. If ordinary multi-class output is used directly, the ordered attribute of "adjacent grades are closer and cross-grade errors are more serious" is easily ignored.

[0100] This invention employs an ordered classification output method to generate a probability distribution of surrounding rock grade based on the final hidden feature vector. The specific steps are as follows: 1> Let the weight vector of the main scoring for the surrounding rock grade be... , dimension The main rating bias is ,and and All of these are trainable parameters used to map the final latent feature vector to a single surrounding rock grade score.

[0101] For the For each sample, first, based on the final latent feature vector... Calculate the main score for the surrounding rock grade ,Right now , This represents the continuous position of the sample along the direction of the weakness of the surrounding rock; a larger value generally indicates a weaker surrounding rock. for The transpose of .

[0102] 2> Set 5 ordered boundary thresholds , , , , These correspond to five boundary positions within a Class 6 surrounding rock. To ensure the stability of the boundary order, an incremental parameterization method can be used to generate the thresholds in the actual implementation. For example, a set of unconstrained parameters can be learned first, and then a positive value accumulation function can be used to ensure stability. .

[0103] In practical implementation, to ensure that the boundary threshold is satisfied We cannot directly learn unconstrained real numbers; instead, we should use an incremental parameterization method. Specifically, we first learn a set of unconstrained auxiliary parameters. , , , , Then, the threshold is constructed by accumulating the exponential function (or Softplus), that is: , , , , ; Due to the exponential function A constant positive value ensures that the threshold strictly increases. For example, if , , , , ,but , , , , It satisfies an increasing relationship.

[0104] 3> For each boundary Calculate the first The cumulative log odds of a sample at this boundary ,Right now Then, the cumulative probability is obtained through the Sigmoid function. ,Right now ,in, Indicates the first The surrounding rock grade of each sample is no higher than that of the first. The probability of level.

[0105] In practical implementation, the predicted probability of Class 1 surrounding rock is directly taken as... The predicted probability for surrounding rock grades 2 to 5 is the difference between two adjacent cumulative probabilities; the predicted probability for surrounding rock grade 6 is... Based on this, we can obtain the predicted probabilities of six surrounding rock grades, and the sum of the six probabilities is 1.

[0106] 4> Output the predicted probabilities of the six surrounding rock levels in order from level 1 to level 6, forming the... The probability distribution of surrounding rock grade for each sample, with the grade corresponding to the highest probability being used as the final surrounding rock grade identification result.

[0107] S46, Destruction Probability Output The surrounding rock classification can only provide discrete results and cannot directly reflect the continuous risk level. Therefore, it is necessary to set up an independent failure probability output head to perform regression prediction on the risk of surrounding rock instability or failure. The specific steps are as follows: Let the weight vector of the destruction probability output head be... , dimension The bias term is ,and and All of these are trainable parameters used to map the final latent feature vector to a continuous risk space.

[0108] For the Each sample will result in the final latent feature vector. With weight vector A linear mapping is performed, and then the sample is compressed to between 0 and 1 using the Sigmoid function to obtain the predicted probability of destruction. , The higher the value, the higher the risk of instability or damage to the surrounding rock.

[0109] S5. Joint loss function calculation and backpropagation optimization: Construct a rock-surrounding intelligent identification model that includes an ordered classification head and a continuous probability output head. Combine ordered classification loss, continuous risk loss, complexity constraints and weight regularization to construct a unified loss function, and use the Adam optimizer for backpropagation update.

[0110] The training samples contain both surrounding rock level labels and failure probability labels. Relying solely on a single classification loss or a single regression loss is insufficient to fully constrain the model's learning direction. Without geological consistency constraints, the geological complexity coefficient may also become overly dependent on local inputs and deviate from the true evolution law of the surrounding rock.

[0111] This invention integrates ordered classification objectives, continuous risk objectives, complexity constraint objectives, and weight regularization terms into a unified optimization framework, and employs the Adam optimizer to implement backpropagation updates. The specific steps are as follows: S51. Calculation of Losses Based on Orderly Classification of Surrounding Rock Levels The surrounding rock grades have a natural order relationship. During training, it is necessary to directly constrain the predicted probability corresponding to the true grade. This invention uses the negative log-likelihood form to construct the ordered classification loss of the surrounding rock grades. The specific steps are as follows: 1> For each sample Read the true surrounding rock grade from the probability distribution of surrounding rock grades The corresponding prediction probability is used to characterize the confidence level of the model in the actual surrounding rock grade.

[0112] Then, the natural logarithm of the predicted probabilities of the true levels for all training samples is taken, averaged, and then the result is negative to obtain the ordered classification loss for the surrounding rock level. Losses classified by surrounding rock level The smaller the value, the more accurate the identification of the surrounding rock grade; When the model assigns a low probability to the true level, the ordered classification loss of the surrounding rock level... This will significantly increase the penalty for misclassification.

[0113] S52, Calculation of regression loss based on probability of destruction The probability of destruction is a continuous risk label, and it is necessary to constrain the numerical deviation between the predicted value and the actual value. This invention uses the mean squared error form to construct the regression loss for the probability of destruction. The specific steps are as follows: For each sample Calculate the predicted probability of damage With the true label of probability of destruction The difference between the predicted and actual risks represents the degree to which the predicted risk deviates from the actual risk.

[0114] Then, the average of the squared differences across all samples is used to obtain the regression loss for the probability of destruction. Probability of destruction regression loss The smaller the value, the more accurate the continuous risk prediction.

[0115] S53, Calculation of Geological Complexity Constraint Loss The geological complexity coefficient is used to adjust the nonlinear strength and effective depth of the network. Its trend should be basically consistent with the weakness of the surrounding rock. If some low-level surrounding rock samples are given too high a complexity coefficient, the network will use too strong nonlinearity for simple samples, affecting the stability of the model.

[0116] This invention imposes constraints on the unreasonable overestimation of the relationship between the geological complexity coefficient and the actual surrounding rock grade. By comparing samples one by one and penalizing the excess, the complexity coefficient is made to be basically consistent with the weakness of the surrounding rock. The specific steps are as follows: For each sample According to the actual surrounding rock grade Generate a reference value for the upper limit of complexity; specifically, you can use... The calculation method yields a reference value for the upper limit of complexity. This reference value ranges from 0 to 1, with a larger value for higher rock quality levels. It represents the upper limit of complexity allowed for that level. For example, (Most complete hard rock) time , (Extremely soft loose material) .

[0117] Then, the first Geological complexity coefficient of each sample Compared with the upper limit of complexity reference value, when No penalty is imposed when the value is not higher than the reference value. When the value is higher than the reference value, only the excess portion is subject to a squared penalty.

[0118] Then, the average of the out-of-limit penalties for all samples is used to obtain the geological complexity constraint loss. ,Right now Geological complexity constraint loss The smaller the value, the more accurately the geological complexity mapping matches the actual weakness of the surrounding rock.

[0119] S54. Total Loss Construction and Parameter Update To simultaneously consider surrounding rock level identification, continuous risk prediction, geological consistency, and model generalization ability, this invention sums multiple losses according to weights to form a unified total loss, and updates all trainable parameters using the Adam optimizer. The specific steps are as follows: 1> Classify the losses of surrounding rock in an orderly manner by grade Probability of destruction regression loss Geological complexity constraint loss The total loss is obtained by combining the weighted regularization terms (according to the preset weights below). ; The weight of the regression loss for the probability of destruction is denoted as... This is used to balance the importance of the surrounding rock grade identification task and the continuous risk prediction task, with an example value of 0.5. The weight of the geological complexity constraint loss is denoted as This is used to control the strength of geological consistency constraints, with an example value of 0.01. The weight of the weight regularization term is denoted as This is used to suppress the risk of overfitting caused by excessively large network parameters; an example value is 0.001.

[0120] In one implementation, the weight regularization term refers to applying an L2 norm penalty to all trainable weight matrices (excluding biases) in the network. The parameters requiring regularization include the weight matrices of each geological stratification unit. ( The weight vector of the output head for the surrounding rock level Weight vector of the probability output head of destruction Then the L2 regularization term The calculation method is expressed as .

[0121] 2> The Adam optimizer is used to perform backpropagation updates on all trainable parameters. The initial learning rate of the Adam optimizer can be 0.001. After every 50 training rounds, the learning rate is multiplied by 0.9. The batch size can be 64. The training is stopped when the aggregate loss no longer decreases. The training rounds are usually around 150 rounds.

[0122] During training, robust local normalization, nonlinear feature recombination, forward propagation, joint loss calculation, and backpropagation update are performed sequentially for each batch of samples.

[0123] S6. Intelligent identification of weak surrounding rock in tunnels: Utilizing an intelligent surrounding rock identification model to output the surrounding rock level identification results and failure probability prediction values. Once the model has been trained and all parameters have been saved, it can be used for intelligent identification of new tunnel cross sections.

[0124] For each new sample requiring identification, 23 original data items are first collected in the same feature order as during the training phase. If individual features are temporarily missing, robust center estimates of the corresponding features saved during the training phase are used to fill in the missing features. Then, robust local normalization is performed: the robust center estimates and adaptive scaling parameters of the 23 features stored during the training phase are called, and the corresponding robust center estimate is subtracted from each original feature of the new sample, then divided by the sum of the adaptive scaling parameter and the minimum constant to obtain the normalized feature vector. Then, nonlinear feature recombination is performed: using the normalized features of the training reference samples, kernel width parameters, and kernel mapping coefficients of each embedding dimension saved during the training phase, the Gaussian kernel similarity between the new sample and all training reference samples is first calculated. Then, for each embedding dimension, the weighted sum of the kernel similarity and kernel mapping coefficients is used as the kernel response, and the normalized value of the corresponding dynamically sensitive feature is multiplied by the guiding coefficient to form the original recombination value. Finally, the mean and standard deviation of each embedding dimension saved during the training phase are used to statistically correct the original recombination value to obtain the recombination feature vector.

[0125] The recombined feature vector is input into a geological stratification adaptive deep neural network: First, based on the normalized values ​​of the burial depth and groundwater flow of the new sample, combined with the trained burial depth weight parameters, groundwater flow weight parameters, and complexity bias parameters, the geological complexity coefficient is calculated using the Sigmoid function. Then, each geological stratification unit is traversed sequentially. For each layer, a linear mapping is performed to obtain a linear response vector. No random perturbation is injected during the inference stage (i.e., the perturbation vector is set to zero). Then, each component of the linear response vector is activated by complexity modulation according to the geological complexity coefficient. Finally, the nonlinear response of the current layer is fused with the input of the previous layer according to the geological complexity coefficient through gated skip fusion, and the feature vector of the current layer is output. After processing through all L layers, the final hidden feature vector is obtained. Based on this hidden feature vector, the ordered output head calculates the main score of the surrounding rock level and generates the predicted probability of 6 surrounding rock levels by combining five ordered boundary thresholds. The one with the highest probability is the identified surrounding rock level. At the same time, the failure probability output head linearly maps the hidden feature vector and then compresses it with Sigmoid to output the failure probability prediction value between 0 and 1.

[0126] Finally, the system outputs the surrounding rock level identification results (levels 1 to 6) and the failure probability prediction value (e.g., 0.35 represents a 35% risk of instability) to the tunnel construction management platform, so that on-site engineers can adjust construction parameters or take reinforcement measures.

[0127] For multiple consecutive new cross sections, repeating the above process can achieve real-time intelligent identification of weak surrounding rock in tunnels.

Claims

1. A method for intelligent identification of weak surrounding rock in tunnels, characterized in that, Includes the following steps: S1. Collect multi-source data related to the surrounding rock during tunnel construction, manually label the surrounding rock level and damage risk, and construct a labeled training dataset. S2. Perform geological strength index softening weighting, local block density scale estimation and robust standardization transformation on the training dataset, and robust local normalization processing on multi-source heterogeneous data of tunnel surrounding rock. S3. Through dynamic sensitive feature screening, kernel mapping embedding, and statistical correction output, the normalized features are reorganized into a low-dimensional representation with geodynamic consistency. S4. Based on the sample-level geological complexity coefficient, dynamically adjust the activation steepness and jump fusion ratio of each layer to achieve sample-level adaptive forward propagation; S5. Construct an intelligent rock identification model that includes an ordered classification head and a continuous probability output head. Combine ordered classification loss, continuous risk loss, complexity constraints and weight regularization to construct a unified loss function, and use the Adam optimizer for backpropagation update. S6. Utilize the intelligent rock identification model to output the rock level identification results and failure probability prediction values.

2. The intelligent identification method for weak surrounding rock in tunnels according to claim 1, characterized in that, The multi-source data related to the surrounding rock in S1 include geological exploration data, construction response data, and field monitoring data.

3. The intelligent identification method for weak surrounding rock in tunnels according to claim 1, characterized in that, S2 specifically refers to: S21. Huber weighted center estimation based on softening of geological strength index: First, construct the softening factor of geological strength index for each sample based on the rock quality index Q value and uniaxial compressive strength. Then, combine the factor with Huber robustness loss to make a stable estimate of the center position of each feature. S22. Adaptive scaling estimation based on local block density: First, estimate the local block density around each sample in the multidimensional feature space, and then use the local block density to adjust the contribution of absolute bias to obtain adaptive scaling parameters that better match the distribution of the surrounding rock block structure. S23. Robust local normalization output: After obtaining the robust center estimate and adaptive scale parameter for each feature, all original features are translated and scaled to obtain the standardized input for network training.

4. The intelligent identification method for weak surrounding rock in tunnels according to claim 1, characterized in that, S3 specifically refers to: S31. Dynamic sensitive feature screening: Evaluate the statistical dependence strength between each input feature and the surrounding rock level label and failure probability label, and then select the feature with the highest comprehensive sensitivity to enter the embedding process. S32, Kernel Mapping Embedding: Low-dimensional embedding is constructed by superimposing kernel mapping and linear guidance. The Gaussian kernel captures global nonlinear similarity, while dynamic sensitive features are used as linear guidance terms, so that each embedding dimension has both physical orientation and retains complex coupling relationships. S33. Reorganize the statistical correction output of the features, and perform zero mean and unit variance correction on each embedding dimension to obtain a scale-consistent reorganized feature matrix.

5. The intelligent identification method for weak surrounding rock in tunnels according to claim 1, characterized in that, S4 specifically refers to: S41, Geological complexity coefficient generation, compresses the burial depth and groundwater flow into a geological complexity coefficient between 0 and 1, which is used to uniformly adjust the nonlinear intensity and jump fusion ratio of each layer; S42, Linear mapping and perturbation injection of geological stratification units: linear mapping is performed on the input features of each layer, and a random perturbation term with adjustable intensity is injected according to the complexity of the sample. S43. Geological complexity modulated activation: The steepness of the activation curve of each layer is dynamically adjusted according to the geological complexity coefficient. By introducing an adaptive activation function that modulates complexity, each sample has a nonlinear mapping intensity that matches its own geological complexity. S44, Gated Jump Fusion Output: The geological complexity coefficient is used to gated the fusion of two pathways, so that different samples can automatically select a more suitable effective depth. S45. The surrounding rock level is output in an ordered manner, using an ordered classification output method to generate the probability distribution of the surrounding rock level based on the final implicit feature vector. S46, Destruction probability output, for the first... For each sample, the final latent feature vector and weight vector are linearly mapped, and then compressed to between 0 and 1 using the Sigmoid function to obtain the predicted failure probability value of the sample. The larger the value, the higher the risk of surrounding rock instability or failure.

6. The intelligent identification method for weak surrounding rock in tunnels according to claim 1, characterized in that, S5 specifically refers to: S51. Calculation of ordered classification loss of surrounding rock level: The ordered classification loss of surrounding rock level is constructed using the negative log-likelihood form. S52. Calculation of regression loss for probability of destruction: The regression loss for probability of destruction is constructed using the mean square error method. S53. Geological complexity constraint loss calculation: constrain the unreasonable overestimation of the relationship between the geological complexity coefficient and the actual surrounding rock level. By comparing samples one by one and penalizing the excess, the complexity coefficient is made to be basically consistent with the weakness of the surrounding rock. S54. Construction of total loss and parameter update: The multiple losses are summed according to their weights to form a unified total loss, and the Adam optimizer is used to update all trainable parameters.

7. The intelligent identification method for weak surrounding rock in tunnels according to claim 4, characterized in that, S31 specifically refers to: S311. For each training sample, read the true label of the surrounding rock level and the true label of the failure probability. S312, For each feature ,Will and The average is calculated to obtain the comprehensive mutual information score. , The larger the value, the more sensitive the feature is to both the identification of the surrounding rock grade and the prediction of instability risk, and the more worthwhile it is to retain it in subsequent feature recombination. S313. Calculate the 23 features according to their comprehensive mutual information scores. Sort by largest to smallest, then select the first... These features form a dynamic sensitive feature index set. , The feature order remains fixed in subsequent steps and is used to construct the correspondence between linear guiding terms and embedding dimensions.

8. The intelligent identification method for weak surrounding rock in tunnels according to claim 4, characterized in that, S32 specifically refers to: S321, using the normalized characteristic matrix Based on this, calculate the pairwise Euclidean distance between all training samples, and take the median of all pairwise Euclidean distances. Multiples as kernel width parameter ; S322, Set up the dynamic sensitive feature index The features in the data are ranked from highest to lowest based on their comprehensive mutual information scores. , No. The first embedding dimension corresponds to the first... One dynamic sensitive feature index This ensures that each embedded dimension has a clear physical orientation; S323, For each embedding dimension , with the first The normalized value sequence of each dynamic sensitive feature on all training samples is used as the target response, and the kernel mapping coefficient of this embedding dimension is solved by kernel ridge regression. S324. For each sample and embedding dimension First, calculate the weighted sum of kernel similarities between the sample and all training samples to obtain the kernel response.

9. The intelligent identification method for weak surrounding rock in tunnels according to claim 8, characterized in that, S33 specifically refers to: S331, For each embedding dimension The mean and standard deviation of the original recombined values ​​of all training samples in this dimension are calculated. S332. Arrange all corrected embedding values ​​in dimensional order to form a reconstructed feature matrix. , dimension , As input data for the geological stratification adaptive deep neural network, and the first The first sample The values ​​of each embedding dimension are denoted as... , No. The recombined feature vector of each sample is denoted as . .

Images