A method for dynamically predicting quality of surrounding rock in front of a tunnel

Abstract

The application discloses a tunnel front surrounding rock quality dynamic prediction method. The application relates to the technical field of tunnel engineering and aims to solve the problem that the existing surrounding rock quality evaluation method cannot utilize the time sequence evolution law of continuous construction cycle data to perform multi-step advanced prediction on the surrounding rock quality of a front unexcavated section. The method collects a working face image and extracts a structure integrity index and a weathering degree index to determine the surrounding rock grade; a surrounding rock quality time sequence data is constructed and preprocessed; a training sample is constructed by using a sliding window method; a time sequence prediction model is used to learn the time sequence evolution law of the surrounding rock quality; a model is deployed to perform forward reasoning to obtain a front surrounding rock quality prediction result; a surrounding rock grade probability distribution graph and a key parameter change curve are outputted, and a support parameter dynamic adjustment suggestion is generated. The application realizes multi-step advanced prediction on the front surrounding rock quality, upgrades static grading to dynamic prediction, and can identify adverse geology in advance and optimize support parameters.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering technology, and in particular to a method for dynamic prediction of the quality of surrounding rock ahead of a tunnel. Background Technology

[0002] In tunnel drilling and blasting construction, surrounding rock quality assessment is the core basis for guiding support design and construction decisions. Current engineering practice primarily relies on two technical approaches for surrounding rock quality assessment: one is manual experience-based grading based on geological sketches of the tunnel face, and the other is automatic identification and grading based on digital images of the tunnel face. The former relies on geological engineers to make qualitative judgments based on on-site observations of lithology, joint and fracture development, weathering status, and other factors. The latter uses image processing algorithms to extract quantitative parameters such as structural integrity indicators and weathering degree indicators, which are then mapped to obtain the surrounding rock grade. Both methods are essentially static evaluations of the exposed rock mass at a single tunnel face. Their evaluation results only reflect the surrounding rock condition at the currently excavated location and cannot provide forward-looking judgments on the changing trends of the surrounding rock quality in the unexcavated sections ahead.

[0003] As the tunnel advances continuously along its mileage, the surrounding rock quality between adjacent construction cycles often exhibits a temporal pattern of gradual evolution or abrupt changes. This pattern implies an inherent connection between stratigraphic structure, lithological combinations, and geological structures in spatial distribution. However, existing technologies treat the surrounding rock evaluation results of each cycle as independent discrete samples, failing to correlate and integrate these data according to construction sequence and spatial mileage. Furthermore, they fail to explore the evolutionary patterns and transmission characteristics between cycles from a holistic temporal perspective, resulting in a significant waste of temporal information inherent in the continuous construction process. When fault fracture zones, lithological abrupt interfaces, or gradually changing weathering sections exist ahead of the tunnel, the lack of effective means to capture the temporal correlation patterns of multi-cycle data makes it difficult for construction units to obtain early warning signals of surrounding rock quality deterioration. They often can only react passively after excavation, which not only increases the risk of sudden geological disasters but also renders the pre-optimization and adjustment of support parameters technically unavailable. Therefore, it is necessary to propose a dynamic prediction method for the surrounding rock quality ahead of the tunnel to address these problems. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic prediction method for the quality of surrounding rock ahead of a tunnel, in order to solve the problem that existing methods for evaluating the quality of surrounding rock fail to utilize the temporal evolution of continuous construction cycle data to make multi-step advance predictions of the quality of surrounding rock in the unexcavated section ahead.

[0005] This invention provides a method for dynamically predicting the quality of surrounding rock ahead of a tunnel, comprising the following steps: Step 1: In each construction cycle of tunnel construction, images of the tunnel face are collected, and image recognition processing is performed on the images of the tunnel face to extract the structural integrity index and weathering degree index corresponding to the construction cycle; based on the combination relationship between the structural integrity index and the weathering degree index and the tunnel surrounding rock grading standard, the surrounding rock grade corresponding to the construction cycle is determined. Step 2: Sort the structural integrity index, weathering degree index and surrounding rock grade of each construction cycle according to the order of construction mileage or construction time to construct the surrounding rock quality time series data, and perform outlier processing and normalization preprocessing on the surrounding rock quality time series data. Step 3: On the preprocessed surrounding rock quality time series data, a sliding window method is used to construct training samples. The historical window length and prediction length are set so that the input sequence of each training sample is the surrounding rock quality data of the continuous historical window length for construction cycles, and the label sequence of each training sample is the structural integrity index, weathering degree index and surrounding rock grade of the continuous prediction length for construction cycles. Step 4: Input the training samples into the time series prediction model for training. The time series prediction model takes the input sequence as input and the label sequence as output to learn the time series evolution law of the surrounding rock quality data between historical construction cycles and future construction cycles. Step 5: Deploy the trained time series prediction model at the tunnel construction site. After each new construction cycle is completed, use the surrounding rock quality data of the most recent consecutive historical window length of construction cycles as the model input to perform forward inference and obtain the predicted values ​​of structural integrity index, weathering degree index, and surrounding rock grade prediction probability distribution corresponding to the previous predicted length of construction cycles. Step six: Output the predicted probability distribution of the surrounding rock grade in the form of a surrounding rock grade probability distribution map, output the predicted values ​​of the structural integrity index and the predicted values ​​of the weathering degree index in the form of key parameter change curves, and generate dynamic adjustment suggestions for tunnel construction support parameters based on preset construction specifications and risk thresholds.

[0006] Furthermore, in step one, the structural integrity index is obtained by performing edge detection, crack identification, and statistical calculation of crack length and crack density on the face image; the weathering degree index is obtained by performing color space conversion, weathered region segmentation, and statistical calculation of weathered area ratio on the face image.

[0007] Furthermore, in step two, the outlier handling uses box plots to identify outlier data points, and interpolation correction or removal is applied to the identified outlier data points; the normalization preprocessing uses the minimum-maximum normalization method or the standard score normalization method.

[0008] Furthermore, in step two, the surrounding rock quality time series data also includes at least one additional feature from the advance length, burial depth, groundwater condition coding, and monitoring measurement data of each construction cycle. The additional feature, together with the structural integrity index, the weathering degree index, and the surrounding rock grade, constitutes a multidimensional time series feature and participates in the sample construction of the sliding window method.

[0009] Further, in step four, the time-series prediction model is one of a long short-term memory network, a gated recurrent unit, or a neural network model based on a self-attention mechanism; the time-series prediction model is trained using a joint loss function, which includes a cross-entropy loss term for predicting the surrounding rock grade and a mean square error loss term for predicting the numerical values ​​of structural integrity index and weathering degree index. The weights of each loss term are adjusted to simultaneously optimize the accuracy of grade prediction and parameter prediction.

[0010] Furthermore, in step five, the time-series prediction model adopts an encoder-decoder structure, wherein the encoder is used to receive and encode a multidimensional surrounding rock quality feature sequence of construction cycles with a historical window length, and the decoder is used to output the predicted values ​​of structural integrity index, weathering degree index, and surrounding rock grade prediction probability distribution of the predicted location with a predicted length.

[0011] Furthermore, the encoder and the decoder employ a neural network structure based on a self-attention mechanism, which includes a multi-head self-attention layer and a feedforward neural network layer, for capturing long-distance temporal dependencies in the surrounding rock quality data.

[0012] Furthermore, in step six, the risk threshold includes: when the predicted probability of a Class IV or Class V surrounding rock grade at a certain predicted location or within a continuous mileage interval exceeds a preset threshold, the system automatically issues an early warning and provides suggestions to increase the scope of advance support, strengthen support parameters, or optimize construction methods; the surrounding rock grade probability distribution map is a two-dimensional heat map with tunnel mileage as the horizontal axis and surrounding rock grade as the vertical axis, and color depth representing the predicted probability of each grade; the key parameter change curves are prediction curves of structural integrity index and weathering degree index changing with tunnel mileage, and an error band representing the confidence interval is added outside the prediction curves.

[0013] Furthermore, in step six, the dynamic adjustment suggestions for the support parameters include dynamically optimizing and adjusting at least one of the following support parameters: advanced small guide pipes, pipe roof, system anchor bolts, shotcrete thickness, and steel arch spacing.

[0014] Furthermore, the method also includes: after completing the actual excavation of a tunnel section, comparing the actual structural integrity index, weathering degree index and surrounding rock grade revealed for that section with the corresponding prediction results from the previous period, evaluating the performance of the time series prediction model using at least one of the following indicators: root mean square error, prediction accuracy, recall rate and F1 score, and updating, incrementally training or retraining the time series prediction model based on the evaluation results.

[0015] The beneficial effects of this invention are as follows: The dynamic prediction method for the quality of surrounding rock ahead of a tunnel in this invention acquires images of the tunnel face during continuous construction cycles and extracts structural integrity and weathering indicators. It constructs a time series of surrounding rock quality data for each cycle according to the construction sequence, uses the sliding window method to construct training samples containing historical evolution patterns and future prediction targets, and employs a deep time series prediction model to learn the temporal correlation features between data from multiple cycles. This enables multi-step advance prediction of the quality of surrounding rock in the unexcavated section ahead, effectively overcoming the limitations of existing technologies that can only perform static evaluations of a single tunnel face and cannot capture the evolution of surrounding rock from an overall temporal perspective. The prediction results are presented intuitively as a probability distribution map of the surrounding rock grade and a curve showing the change of key parameters. Coupled with the risk threshold and support parameter library in the construction specifications, the prediction results can automatically generate dynamic adjustment suggestions for support parameters. This allows construction units to understand the changing trend of the surrounding rock quality before excavation and take enhanced support measures in areas with poor geological conditions in advance. They can also appropriately optimize support parameters in areas with improved surrounding rock conditions. This achieves a shift from passive response to proactive prevention and control. While reducing the risk of sudden geological disasters, it also reduces unnecessary over-support waste and significantly improves the safety and economy of tunnel construction. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the dynamic prediction method for the surrounding rock quality ahead of the tunnel according to the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The technical solutions provided by various embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0019] Please see Figure 1 This invention provides a method for dynamically predicting the quality of surrounding rock ahead of a tunnel, comprising: Step 1: In each construction cycle of tunnel construction, images of the tunnel face are collected, and image recognition processing is performed on the images of the tunnel face to extract the structural integrity index and weathering degree index corresponding to the construction cycle; based on the combination relationship between the structural integrity index and the weathering degree index and the tunnel surrounding rock grading standard, the surrounding rock grade corresponding to the construction cycle is determined.

[0020] Specifically, the structural integrity index is obtained by performing edge detection, crack identification, and statistical calculation of crack length and crack density on the face image; the weathering degree index is obtained by performing color space conversion, weathered region segmentation, and statistical calculation of weathered area ratio on the face image.

[0021] By acquiring images of the excavation face during each construction cycle, raw visual information reflecting the current state of the rock mass at the excavation face can be obtained. Image acquisition utilizes fixed or mobile industrial cameras, automatically or semi-automatically triggered under dust-controlled and supplemental lighting conditions to ensure image quality meets subsequent recognition requirements. Image recognition processing is performed on the face images to extract structural integrity indicators. These indicators are obtained through edge detection, crack identification, and statistical calculation of crack length and density. They quantify the degree of development of structural planes in the rock mass; a higher structural integrity index value indicates a more intact rock mass with fewer cracks, while a lower value indicates a higher degree of rock fragmentation.

[0022] The formula for calculating the structural integrity index is as follows: ,in This represents the structural integrity index for the t-th construction cycle. This represents the total length of the cracks identified in the image. The total pixel area of ​​the face image is represented by this formula, which intuitively reflects the degree to which fracture development weakens the integrity of the rock mass. Image recognition processing is performed on the face image to extract a weathering degree index. This index is obtained by performing color space conversion, weathering region segmentation, and statistical calculation of the proportion of weathered area in the face image. It can quantitatively characterize the degree of weathering on the rock mass. The higher the weathering degree index value, the fresher the rock mass and the less weathered it is, and vice versa.

[0023] Formula for calculating weathering degree index ,in This represents the weathering degree index for the t-th construction cycle. This represents the pixel area of ​​the weathered region identified in the image. This formula represents the effective analysis pixel area of ​​the tunnel face image. It accurately measures the weathering degree of the rock mass by quantifying the proportion of weathered areas. The surrounding rock grade is determined based on the combined relationship between structural integrity and weathering degree indices, as well as the tunnel surrounding rock grading standards. This determination process integrates two complementary quantitative indicators into a unified surrounding rock quality grade. The structural integrity indicator focuses on the development of the rock mass's structural surfaces, while the weathering degree indicator focuses on the physicochemical weathering state of the rock mass. The combination of these two indicators provides a more comprehensive characterization of the surrounding rock quality, avoiding potential misjudgments caused by a single indicator, and thus providing a reliable grade label for subsequent time-series modeling.

[0024] Step 2: Sort the structural integrity index, weathering degree index, and surrounding rock grade of each construction cycle according to the order of construction mileage or construction time to construct the surrounding rock quality time series data, and perform outlier processing and normalization preprocessing on the surrounding rock quality time series data.

[0025] Specifically, the outlier handling employs box plotting to identify anomalous data points, and interpolates or removes these points. The normalization preprocessing uses minimum-maximum normalization or standard score normalization. The surrounding rock quality time series data also includes at least one additional feature from each construction cycle's advance length, burial depth, groundwater condition coding, and monitoring measurement data. This additional feature, along with the structural integrity index, weathering degree index, and surrounding rock grade, constitutes a multidimensional time series feature, collectively participating in the sample construction using the sliding window method.

[0026] This step transforms the discrete surrounding rock quality data obtained cycle by cycle into a sequential form with chronological order, thus revealing the inherent law of the evolution of surrounding rock quality as the tunnel excavation progresses. Sorting according to the chronological order of construction mileage or construction time makes the data of adjacent construction cycles continuous in space and relevant in geological structure, laying a foundation for capturing the sequential evolution law in the follow-up. Conducting outlier processing on the surrounding rock quality time series data, using the box plot method to identify outlier data points, and interpolating and correcting or removing the identified outlier data points can effectively remove the outlier data caused by factors such as poor image acquisition quality, sudden light changes, and equipment failures, avoiding the interference of outliers on the subsequent model training and ensuring the smoothness and reliability of the sequential data.

[0027] The decision criterion of the box plot method adopts the formula Q1 - 1.5×IQR < x < Q3 + 1.5×IQR, where x is the surrounding rock quality characteristic value to be detected, Q1 is the first quartile, Q3 is the third quartile, IQR = Q3 - Q1 is the interquartile range, and the data points falling outside this interval will be marked as outliers. Conducting normalization preprocessing on the surrounding rock quality time series data, using the minimum-maximum normalization method or the standard score normalization method, this processing method eliminates the dimensional difference and numerical range difference between different features, enabling features of different scales such as the structural integrity index and the weathering degree index to participate in model training within the same numerical range, accelerating model convergence and improving training stability.

[0028] The normalization process can follow the minimum-maximum normalization formula , where is the normalized feature value, is the original feature value, is the minimum value of the feature sequence, is the maximum value of the feature sequence. Through this transformation, the feature value is mapped to the interval from 0 to 1. In addition, the surrounding rock quality time series data also includes at least one additional feature such as the advance length, buried depth, groundwater condition code of each construction cycle, and monitoring measurement data. The additional features and the structural integrity index, weathering degree index, and surrounding rock grade together constitute multi-dimensional sequential features, jointly participating in the sample construction of the sliding window method, which can introduce more dimensions of geological and construction information, enabling the model to not only learn the self-evolution law of surrounding rock quality but also comprehensively consider the influence of external factors such as construction parameters, geological environment, and hydrological conditions on the change of surrounding rock quality, thereby significantly enhancing the model's ability to depict the evolution law of surrounding rock under complex geological conditions.

[0029] Step 3: On the preprocessed surrounding rock quality time series data, a sliding window method is used to construct training samples. The historical window length and prediction length are set so that the input sequence of each training sample is the surrounding rock quality data of the continuous historical window length for construction cycles, and the label sequence of each training sample is the structural integrity index, weathering degree index and surrounding rock grade of the continuous prediction length for construction cycles.

[0030] Dividing continuous time-series data into standard input-output pairs suitable for deep learning models is a crucial step in transitioning from raw data to model training. Utilizing the sliding window method to construct training samples allows for the efficient generation of a large number of training samples from limited time-series data. By sliding the window, multiple sets of training samples can be constructed over a long time series, effectively addressing the issues of slow data accumulation and limited sample size in tunnel engineering. Setting a historical window length ensures that the input sequence for each training sample consists of surrounding rock quality data from consecutive construction cycles within that historical window length. This historical window length determines the amount of historical information the model can trace back; a longer historical window length allows the model to capture more distant temporal dependencies, but also increases computational complexity. Conversely, a shorter historical window length focuses more on recent trends. Therefore, a reasonable setting based on tunnel geological characteristics and computational resources is necessary. A prediction length is set so that the label sequence for each training sample consists of structural integrity indicators, weathering degree indicators, and surrounding rock grades for a continuous prediction length of construction cycles. This prediction length directly determines the number of construction cycles the model can predict in advance, and thus the advance warning distance. A reasonable prediction length setting achieves a balance between prediction accuracy and advance warning. The clear separation between the input sequence and the label sequence allows the model to establish a mapping relationship between historical and future data during the training phase, thereby enabling it to perform multi-step advance predictions of the surrounding rock quality in unexcavated sections.

[0031] Step four: Input the training samples into the time series prediction model for training. The time series prediction model takes the input sequence as input and the label sequence as output to learn the time series evolution law of the surrounding rock quality data between historical construction cycles and future construction cycles.

[0032] Specifically, the time-series prediction model is one of a long short-term memory network, a gated recurrent unit, or a neural network model based on a self-attention mechanism; the time-series prediction model is trained using a joint loss function, which includes a cross-entropy loss term for predicting the surrounding rock grade and a mean square error loss term for predicting the numerical values ​​of structural integrity index and weathering degree index. The weights of each loss term are adjusted to simultaneously optimize the accuracy of grade prediction and parameter prediction.

[0033] By training a time-series prediction model, the model can automatically mine and memorize the temporal correlation features between data from multiple construction cycles, thereby achieving intelligent learning of the evolution law of surrounding rock quality. The time-series prediction model can be one of the following: Long Short-Term Memory (LSTM) networks, gated recurrent units (GRUs), or neural network models based on self-attention mechanisms. All these models have the ability to process sequential data. LSMs, through forgetting gates, input gates, and output gates, can effectively capture long-term temporal dependencies. GRUs have a simpler structure and are suitable for edge deployment. Neural network models based on self-attention mechanisms can process global temporal information in parallel. Each of the three has its advantages, and the appropriate model can be flexibly selected based on actual engineering conditions. The time-series prediction model is trained using a joint loss function, which includes a cross-entropy loss term for predicting the surrounding rock grade and a mean squared error loss term for predicting the structural integrity and weathering degree indices. By adjusting the weights of each loss term, the accuracy of grade prediction and parameter prediction are optimized simultaneously. This joint training method enables the model to simultaneously handle classification and regression tasks. Surrounding rock grade prediction is a classification task, while the prediction of structural integrity and weathering degree indices is a regression task. The joint optimization of the two results in the model output prediction results containing both discrete grade information and continuous parameter values, providing richer quantitative basis for subsequent construction decisions.

[0034] The expression for the joint loss function is: ,in Let represent the total loss value, α, β, and γ be the weighting coefficients of each loss term, MSE be the mean squared error function, and CE be the cross-entropy loss function. and These are the predicted values ​​for structural integrity and weathering degree indices, respectively. and These are the corresponding true values. The probability distribution for predicting the surrounding rock grade. This is a true label for the surrounding rock grade.

[0035] The temporal prediction model employs an encoder-decoder structure. The encoder receives and encodes a multi-dimensional sequence of surrounding rock quality features spanning several construction cycles within a historical window. The decoder outputs predicted values ​​for structural integrity, weathering, and surrounding rock grade at several predicted locations within a prediction window. The encoder focuses on extracting high-level temporal feature representations from historical data, while the decoder transforms these representations into predictions for future multiple steps. This clear division of labor enhances the model's structural rationality and prediction stability when handling long-sequence prediction tasks. The encoder and decoder utilize a neural network structure based on a self-attention mechanism. This mechanism includes a multi-head self-attention layer and a feedforward neural network layer, used to capture long-distance temporal dependencies in the surrounding rock quality data. The self-attention mechanism assigns different attention weights to each position in the sequence, enabling the model to adaptively focus on the most relevant positions in the historical sequence when processing the current prediction. This effectively captures long-distance dependencies in changes in surrounding rock quality, such as precursory information about fault fracture zones that may appear multiple cycles prior. The multi-head mechanism allows for parallel learning of various dependency patterns from different perspectives, further enhancing the model's expressive power.

[0036] Step 5: Deploy the trained time-series prediction model at the tunnel construction site. After each new construction cycle is completed, use the surrounding rock quality data of the most recent consecutive historical window length of construction cycles as the model input to perform forward inference, and obtain the predicted values ​​of structural integrity index, weathering degree index, and surrounding rock grade prediction probability distribution corresponding to the previous predicted length of construction cycles.

[0037] Specifically, the time-series prediction model employs an encoder-decoder structure. The encoder receives and encodes a multi-dimensional sequence of surrounding rock quality features spanning several construction cycles within a historical window. The decoder outputs predicted values ​​for structural integrity indices, weathering degree indices, and the probability distribution of surrounding rock grade predictions for several predicted locations within a predicted length. Both the encoder and decoder utilize a neural network structure based on a self-attention mechanism. This self-attention mechanism includes a multi-head self-attention layer and a feedforward neural network layer, used to capture long-range temporal dependencies in the surrounding rock quality data.

[0038] The trained time-series prediction model is deployed at the tunnel construction site, enabling it to run in real-time as construction progresses. A prediction update is triggered after each new construction cycle, ensuring the prediction results are always based on the latest construction data. The model input uses the surrounding rock quality data from the most recent consecutive historical window length for several construction cycles. This input method ensures the timeliness and continuity of the historical information upon which the predictions are based. As construction progresses, the prediction window moves forward, achieving dynamic rolling predictions of the surrounding rock quality ahead. Forward inference yields the predicted values ​​of structural integrity indicators, weathering degree indicators, and the probability distribution of surrounding rock grade predictions for the preceding predicted construction cycles. The probability distribution of surrounding rock grade predictions outputs the likelihood of each grade in probabilistic form, rather than a simple single-point prediction, quantifying the uncertainty of the prediction and providing more comprehensive risk information for construction decisions. A relatively concentrated probability distribution indicates high model prediction confidence, while a relatively dispersed probability distribution suggests the need for cautious decision-making.

[0039] Step six: Output the predicted probability distribution of the surrounding rock grade in the form of a surrounding rock grade probability distribution map, output the predicted values ​​of the structural integrity index and the predicted values ​​of the weathering degree index in the form of key parameter change curves, and generate dynamic adjustment suggestions for tunnel construction support parameters based on preset construction specifications and risk thresholds.

[0040] Specifically, the risk threshold includes: when the predicted probability of a Class IV or V surrounding rock grade within a certain predicted location or continuous mileage interval exceeds a preset threshold, the system automatically issues an early warning and provides suggestions to increase the scope of advance support, strengthen support parameters, or optimize construction methods; the surrounding rock grade probability distribution map is a two-dimensional heat map with tunnel mileage as the horizontal axis and surrounding rock grade as the vertical axis, using color depth to represent the predicted probability of each grade; the key parameter change curves are predicted curves of structural integrity index and weathering degree index changing with tunnel mileage, with an error band representing the confidence interval added outside the predicted curves. The dynamic adjustment suggestions for support parameters include dynamic optimization and adjustment of at least one of the following support parameters: advance small guide pipe, pipe roof, system anchor bolts, shotcrete thickness, and steel arch spacing.

[0041] Preferably, after the actual excavation of a tunnel section is completed, the structural integrity index, weathering degree index and surrounding rock grade actually revealed for that section are compared with the corresponding prediction results from the previous stage. The performance of the time series prediction model is evaluated using at least one of the following indicators: root mean square error, prediction accuracy, recall rate and F1 score. Based on the evaluation results, the time series prediction model is updated with parameters, incrementally trained or retrained.

[0042] The predicted probability distribution of surrounding rock grade is output as a probability distribution map, a two-dimensional heatmap with tunnel mileage on the horizontal axis and surrounding rock grade on the vertical axis, using color depth to represent the predicted probability of each grade. This visually presents the probability distribution of different surrounding rock grades at different mileage locations, allowing engineers to easily identify high-risk and low-risk sections. The color depth also reflects the spatial variation trend of surrounding rock quality. The predicted values ​​of structural integrity and weathering indices are output as key parameter variation curves, showing the changes in structural integrity and weathering indices with tunnel mileage. An error band representing the confidence interval is added outside the prediction curves, demonstrating the continuous variation trend of key surrounding rock parameters. The error band quantifies the uncertainty range of the prediction, allowing decision-makers to simultaneously know the predicted value and its reliability.

[0043] Based on preset construction specifications and risk thresholds, the system generates dynamic adjustment suggestions for tunnel construction support parameters, directly translating prediction results into actionable engineering decisions. This avoids a disconnect between prediction and decision-making. The risk thresholds include an automatic early warning system that issues suggestions to increase the scope of advance support, strengthen support parameters, or optimize construction methods when the predicted probability of a Class IV or V surrounding rock grade at a specific predicted location or within a continuous mileage range exceeds a preset threshold. This tiered early warning mechanism enables proactive identification and prevention of adverse geological conditions. The dynamic adjustment suggestions for support parameters include dynamic optimization of at least one of the following: advance small guide pipes, pipe roofs, system anchors, shotcrete thickness, and steel arch spacing. This covers the main parameter types for initial tunnel support, allowing prediction results to comprehensively guide the refined adjustment of construction parameters. Appropriate optimization of support parameters in sections with good surrounding rock quality can reduce costs, while strengthening support in sections with poor surrounding rock quality can ensure safety.

[0044] The method of the present invention further includes, after the actual excavation of a tunnel section is completed, comparing the actual structural integrity index, weathering degree index and surrounding rock grade revealed for that section with the corresponding prediction results from the previous stage, evaluating the performance of the time series prediction model using at least one of the following indicators: root mean square error, prediction accuracy, recall rate and F1 score, and updating the parameters, incrementally training or retraining the time series prediction model based on the evaluation results, so that the model can continuously self-correct and optimize in engineering practice, thereby continuously enhancing the model's adaptability to specific engineering geological conditions and gradually improving the prediction accuracy over time.

[0045] The formula for calculating the root mean square error can be expressed as follows: Where n is the number of samples, Let be the predicted value for the i-th sample. Let be the true value of the i-th sample. This indicator is used to quantify the degree of deviation between the predicted value and the true value.

[0046] The formula for calculating the F1 value is: Precision is Recall is TP represents true positives, FP represents false positives, FN represents false negatives, and the F1 score combines precision and recall to comprehensively evaluate the performance of the classification model.

[0047] In summary, this invention overcomes the limitations of traditional static grading of single tunnel face operations, establishing temporal correlations of multi-construction cycle data. It transforms discrete surrounding rock evaluation results into a temporal sequence with evolutionary patterns, fully leveraging the spatial distribution information of geological structures inherent in continuous construction. This invention employs a deep temporal prediction model to learn the temporal evolution of surrounding rock quality. Through an encoder-decoder structure and a self-attention mechanism, it effectively captures long-distance temporal dependencies. The design of the joint loss function enables the model to simultaneously possess classification and regression capabilities, resulting in richer and more practical output results. This invention deeply couples prediction results with construction specifications. Through probability distribution maps, variation curves, and automated support parameter adjustment suggestions, it achieves seamless integration from prediction to decision-making, transforming tunnel construction from a passive response to proactive pre-control. This invention constructs a closed-loop mechanism encompassing online prediction, accuracy verification, and model updates, enabling the system to possess adaptive evolution capabilities, continuously optimize as the project progresses, and maintain high prediction reliability over the long term.

[0048] The embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention.

Claims

1. A method for dynamic prediction of the quality of surrounding rock ahead of a tunnel, characterized in that, Includes the following steps: Step 1: In each construction cycle of tunnel construction, images of the tunnel face are collected, and image recognition processing is performed on the images of the tunnel face to extract the structural integrity index and weathering degree index corresponding to the construction cycle; based on the combination relationship between the structural integrity index and the weathering degree index and the tunnel surrounding rock grading standard, the surrounding rock grade corresponding to the construction cycle is determined. Step 2: Sort the structural integrity index, weathering degree index and surrounding rock grade of each construction cycle according to the order of construction mileage or construction time to construct the surrounding rock quality time series data, and perform outlier processing and normalization preprocessing on the surrounding rock quality time series data. Step 3: On the preprocessed surrounding rock quality time series data, a sliding window method is used to construct training samples. The historical window length and prediction length are set so that the input sequence of each training sample is the surrounding rock quality data of the continuous historical window length for construction cycles, and the label sequence of each training sample is the structural integrity index, weathering degree index and surrounding rock grade of the continuous prediction length for construction cycles. Step 4: Input the training samples into the time series prediction model for training. The time series prediction model takes the input sequence as input and the label sequence as output to learn the time series evolution law of the surrounding rock quality data between historical construction cycles and future construction cycles. Step 5: Deploy the trained time series prediction model at the tunnel construction site. After each new construction cycle is completed, use the surrounding rock quality data of the most recent consecutive historical window length of construction cycles as the model input to perform forward inference and obtain the predicted values ​​of structural integrity index, weathering degree index, and surrounding rock grade prediction probability distribution corresponding to the previous predicted length of construction cycles. Step six: Output the predicted probability distribution of the surrounding rock grade in the form of a surrounding rock grade probability distribution map, output the predicted values ​​of the structural integrity index and the predicted values ​​of the weathering degree index in the form of key parameter change curves, and generate dynamic adjustment suggestions for tunnel construction support parameters based on preset construction specifications and risk thresholds.

2. The method for dynamic prediction of surrounding rock quality ahead of a tunnel according to claim 1, characterized in that, In step one, the structural integrity index is obtained by performing edge detection, crack identification, and statistical calculation of crack length and crack density on the face image; the weathering degree index is obtained by performing color space conversion, weathered region segmentation, and statistical calculation of weathered area ratio on the face image.

3. The method for dynamic prediction of surrounding rock quality ahead of a tunnel according to claim 2, characterized in that, In step two, the outlier handling uses box plots to identify outlier data points, and interpolation correction or removal is applied to the identified outlier data points; the normalization preprocessing uses the minimum-maximum normalization method or the standard score normalization method.

4. The method for dynamic prediction of surrounding rock quality ahead of a tunnel according to claim 3, characterized in that, In step two, the surrounding rock quality time series data also includes at least one additional feature from the following: advance length, burial depth, groundwater condition coding, and monitoring measurement data for each construction cycle. This additional feature, together with the structural integrity index, the weathering degree index, and the surrounding rock grade, constitutes a multidimensional time series feature. They jointly participate in the sample construction of the sliding window method.

5. The method for dynamic prediction of surrounding rock quality ahead of a tunnel according to claim 4, characterized in that, In step four, the time-series prediction model is one of a long short-term memory network, a gated recurrent unit, or a neural network model based on a self-attention mechanism. The time-series prediction model is trained using a joint loss function, which includes a cross-entropy loss term for predicting the surrounding rock grade and a mean square error loss term for predicting the numerical values ​​of structural integrity index and weathering degree index. The weights of each loss term are adjusted to simultaneously optimize the accuracy of grade prediction and parameter prediction.

6. The method for dynamic prediction of surrounding rock quality ahead of a tunnel according to claim 5, characterized in that, In step five, the time-series prediction model adopts an encoder-decoder structure, wherein the encoder is used to receive and encode a multidimensional surrounding rock quality feature sequence of construction cycles with a historical window length, and the decoder is used to output the predicted values ​​of structural integrity index, weathering degree index, and surrounding rock grade prediction probability distribution of the predicted location with a prediction length.

7. The method for dynamic prediction of surrounding rock quality ahead of a tunnel according to claim 6, characterized in that, The encoder and the decoder employ a neural network structure based on a self-attention mechanism, which includes a multi-head self-attention layer and a feedforward neural network layer, used to capture long-distance temporal dependencies in the surrounding rock quality data.

8. The method for dynamic prediction of surrounding rock quality ahead of a tunnel according to claim 7, characterized in that, In step six, the risk threshold includes: when the predicted probability of a Class IV or Class V surrounding rock grade at a certain predicted location or within a continuous mileage interval exceeds a preset threshold, the system automatically issues an early warning and provides suggestions to increase the range of advance support, strengthen support parameters, or optimize construction methods; the surrounding rock grade probability distribution map is a two-dimensional heat map with tunnel mileage as the horizontal axis and surrounding rock grade as the vertical axis, and color depth representing the predicted probability of each grade; the key parameter change curves are the predicted curves of structural integrity index and weathering degree index changing with tunnel mileage, and an error band representing the confidence interval is added outside the predicted curves.

9. The method for dynamic prediction of surrounding rock quality ahead of a tunnel according to claim 8, characterized in that, In step six, the proposed dynamic adjustment of support parameters includes dynamically optimizing at least one of the following support parameters: advanced small guide pipes, pipe roof, system anchor bolts, shotcrete thickness, and steel arch spacing.

10. The method for dynamic prediction of surrounding rock quality ahead of a tunnel according to claim 9, characterized in that, The method further includes: after completing the actual excavation of a tunnel section, comparing the actual structural integrity index, weathering degree index and surrounding rock grade revealed for that section with the corresponding prediction results from the previous period, evaluating the performance of the time series prediction model using at least one of the following indicators: root mean square error, prediction accuracy, recall rate and F1 score, and updating, incrementally training or retraining the time series prediction model based on the evaluation results.

Images