Machine learning and domain knowledge embedded dynamic prediction method for surrounding rock classification

Abstract

This invention provides a dynamic prediction method for surrounding rock classification based on machine learning and domain knowledge embedding, relating to the field of tunnel engineering technology. The method includes establishing basic engineering information and surrounding rock prediction units; dividing the tunnel into multiple mileage prediction units along the tunnel excavation direction; and establishing a time-mileage-excavation cycle correspondence for each mileage prediction unit. It also involves acquiring multi-source surrounding rock data corresponding to each mileage prediction unit, including tunnel face image data, advanced geological prediction data, construction equipment operation data, monitoring and measurement data, and groundwater data. In this invention, domain knowledge embedding ensures that the prediction results of the machine learning model conform to the surrounding rock classification specifications and rock mechanics laws. Furthermore, multi-source data credibility evaluation reduces the impact of dust obstruction, equipment malfunctions, sensor drift, and human error on the prediction results, thereby ensuring the accuracy of the prediction results.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering technology, specifically to a dynamic prediction method for surrounding rock classification using machine learning and domain knowledge embedding. Background Technology

[0002] In the construction of tunnels, underground caverns, and mine roadways, the surrounding rock grade directly affects excavation methods, support parameters, construction progress, and safety risk control. In tunnel construction, surrounding rock classification is the core basis for determining support parameters and assessing construction risks, directly impacting project safety and economy. Traditional surrounding rock classification methods mainly rely on static means such as drilling sampling and geological logging. By obtaining the physical and mechanical parameters of rock core samples and combining them with classification standards in specifications, such as the BQ value calculation method, the surrounding rock grade is determined. However, existing surrounding rock classification predictions rely on manual classification and engineering experience, resulting in strong subjectivity. Static judgments cannot reflect dynamic changes during construction in a timely manner, and ordinary machine learning methods are prone to becoming black box models, potentially producing prediction results that do not conform to the laws of rock mechanics. Existing models typically only output a single surrounding rock grade, lacking explanations of prediction uncertainties, risk sources, and proactive supplementary measurement suggestions. After model training, they are often used in a fixed manner, making it difficult to continuously update based on the actual exposed surrounding rock grade and support feedback, which can easily lead to problems such as model learning bias.

[0003] Therefore, this invention provides a dynamic prediction method for surrounding rock classification based on machine learning and domain knowledge embedding. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a dynamic prediction method for surrounding rock grading based on machine learning and domain knowledge embedding, thereby solving the problems mentioned in the background. This invention can reduce the impact of dust obstruction, equipment malfunction, sensor drift, and human error on prediction results, thus ensuring accuracy. It avoids misjudgments by single black-box models, improving prediction reliability. It prevents erroneous surrounding rock grade labels from contaminating training data, can continuously adapt to new geological conditions during tunnel excavation, and allows for self-correction of domain knowledge weights. Furthermore, it can output grade probability, confidence level, risk type, influencing factors, support recommendations, and supplementary measurement recommendations, enhancing its construction guidance value.

[0005] To achieve the above objectives, the present invention provides a method for dynamic prediction of surrounding rock grading using machine learning and domain knowledge embedding, comprising the following steps: S1. Establish basic engineering information and surrounding rock prediction units. Divide the tunnel into multiple mileage prediction units along the tunnel excavation direction and establish the time-mileage-excavation cycle correspondence for each mileage prediction unit. S2. Obtain multi-source surrounding rock data corresponding to each mileage prediction unit. The multi-source surrounding rock data includes face image data, advanced geological prediction data, construction equipment operation data, monitoring and measurement data, groundwater data, artificial geological sketch data, support parameter data, and post-support deformation response data. The data is then preprocessed and aligned to the mileage. S3. Construct a knowledge base for surrounding rock classification based on knowledge in fields such as surrounding rock classification standards, rock mechanics, and support feedback rules; S4. Construct a knowledge graph of surrounding rock classification based on the knowledge base of surrounding rock classification, and convert the knowledge graph of surrounding rock classification into the domain knowledge embedding vector corresponding to each mileage prediction unit. S5. Extract features from multi-source surrounding rock data to obtain features of the working face image, advanced geological prediction features, construction equipment operation features, monitoring and measurement features, and artificial geological sketch features; S6. Calculate the data credibility of each data source based on data quality, time freshness, multi-source consistency, device status, and domain knowledge consistency; S7. Input the features of the working face image, advanced geological prediction features, construction equipment operation features, monitoring and measurement features, artificial geological sketch features, domain knowledge embedding vectors and data credibility into the domain knowledge embedded multi-task prediction model to obtain the machine learning surrounding rock grade prediction results, surrounding rock grade probability, prediction uncertainty and adverse geological risk results. S8. Calculate the surrounding rock grade according to the surrounding rock classification specification, and perform conflict verification between the surrounding rock grade result and the machine learning surrounding rock grade prediction result. S9. When the degree of conflict is less than the preset threshold, the surrounding rock grade prediction result is fused and output; when the degree of conflict is greater than or equal to the preset threshold, the surrounding rock grade prediction result is conservatively corrected based on data credibility, domain knowledge consistency, adverse geological risks and support feedback results. S10. Calculate the prediction uncertainty based on the probability of surrounding rock grade, and determine the priority of active learning based on the prediction uncertainty, conflict degree, engineering risk weight, sample representativeness and domain knowledge anomaly. S11. Add mileage prediction units with active learning priority higher than the preset threshold to the active learning sample pool, and generate expert review or retest suggestions. S12. After excavation is completed, obtain the actual exposed surrounding rock grade and deformation response data after support, and calculate the label credibility based on the manual judgment results, deformation response after support, consistency of surrounding rock grade between adjacent mileages, consistency of rule classification and consistency of domain knowledge. S13. Based on the credibility of the labels, the samples are divided into strong-label samples, weak-label samples, or samples to be verified. S14. Add the prediction error samples, level boundary samples, rule conflict samples, low confidence samples, support feedback abnormal samples and domain knowledge abnormal samples to the difficult sample library, and increase the training weight of difficult samples when updating the model. S15. Based on newly revealed samples, difficult samples, historical representative samples, active supplementary test samples, support feedback samples, and domain knowledge anomaly samples, the domain knowledge embedded multi-task prediction model is continuously learned and updated. S16. Based on the actual exposed surrounding rock grade, prediction error and support feedback results, dynamically correct the domain knowledge rule weights in the surrounding rock classification domain knowledge base; S17. Output the surrounding rock grade, grade probability, confidence level, main influencing factors, adverse geological risk type, support adjustment suggestions, and supplementary measurement suggestions for the current segment and the segment to be excavated ahead.

[0006] Furthermore, the establishment of engineering basic information and surrounding rock prediction units includes establishing a tunnel engineering coordinate system and establishing a time-mileage-excavation cycle correspondence. Among them, establishing a tunnel engineering coordinate system includes establishing a unified spatial coordinate system with the tunnel line mileage as the main axis, dividing the tunnel into several mileage prediction units according to the excavation direction, and each prediction unit includes prediction unit number, starting mileage, ending mileage, excavation time, support time, construction cycle number, tunnel face number, corresponding prediction point number, and corresponding advance prediction interval, unifying data from different sources into the same mileage unit.

[0007] Furthermore, the data preprocessing and mileage alignment include data cleaning, time alignment, spatial alignment, and data standardization. Specifically, data cleaning includes removing obvious erroneous values, correcting sensor drift, deleting duplicate data, filling in missing time points, removing invalid images, standardizing data units, and standardizing sampling frequencies. For data from different data sources with varying sampling times, the data from different data sources are aligned according to excavation cycle numbers and mileage intervals. Advanced forecast data, monitoring point data, and face image data are uniformly mapped to corresponding prediction units. Data with different dimensions are standardized to obtain standardized data that can be input into machine learning models.

[0008] Furthermore, the construction of the knowledge base for surrounding rock classification includes establishing standardized classification knowledge, transforming the rules in the surrounding rock classification specifications into computable knowledge, transforming field engineering experience into rules, determining support parameters through the surrounding rock grade, and then verifying the surrounding rock grade in reverse based on support feedback; then setting knowledge graph nodes and knowledge graph relationships, and then for each prediction unit, extracting relevant subgraphs from the knowledge graph based on its corresponding lithology, groundwater, fissures, support feedback and other information.

[0009] Furthermore, the knowledge-guided data credibility evaluation includes calculating the data quality coefficient, where different data sources have different quality indicators, and calculating the freshness coefficient based on time. The closer the time is to the current working face, the more credible the data. Then, the system judges whether different data sources corroborate each other, and then judges the data credibility based on the equipment operating status. If a data source obviously conflicts with the knowledge of surrounding rock engineering, its credibility is reduced.

[0010] Furthermore, the fusion of multi-source features and domain knowledge includes constructing a model input vector, combining various features, domain knowledge embedding vectors, and data credibility as input, weighting and fusing different data sources according to credibility, and then establishing a domain knowledge embedded multi-task prediction model. The model includes multiple branches, such as image recognition branch, advanced prediction branch, construction parameter time series branch, etc., and then setting multi-task outputs and multi-task constraint relationships.

[0011] Furthermore, the model training process includes establishing training samples, which include historical project samples, current project excavated samples, manually verified samples, and actual exposed samples. Then, domain knowledge-constrained loss functions and monotonicity constraints are set, support feedback consistency constraints are set, and the rule-based surrounding rock grade is calculated according to the standard grading method. Finally, the confidence level is calculated.

[0012] Furthermore, the machine learning channel prediction includes outputting the surrounding rock grade from the model, and then outputting interpretable results to facilitate engineers' judgment.

[0013] Furthermore, when a conflict occurs, the surrounding rock grade is digitized, the degree of conflict is calculated, and the conflict level is determined. Then, the uncertainty is calculated based on the probability of the surrounding rock grade, and the confidence zone is determined.

[0014] Furthermore, the learning and retesting process includes prioritizing active learning and filtering active learning samples, and generating retesting suggestions when the active learning priority is high, with the retesting data re-entering the multi-source data collection.

[0015] The beneficial effects of this invention are: 1. By embedding domain knowledge, the prediction results of the machine learning model conform to the surrounding rock classification specifications and the laws of rock mechanics; by evaluating the credibility of multi-source data, the impact of dust obstruction, equipment malfunction, sensor drift and human misjudgment on the prediction results is reduced, thereby ensuring the accuracy of the prediction results; by using dual verification through rule channels and machine learning channels, misjudgment by a single black box model is avoided, and the reliability of the prediction is improved.

[0016] 2. By quantifying the uncertainty of prediction, the system can identify low-confidence surrounding rock sections and proactively propose supplementary measurement suggestions to prevent errors. Through an active learning mechanism, the model prioritizes learning high-risk, high-uncertainty, and high-value samples to improve learning efficiency. By correcting the credibility of labels, it avoids erroneous surrounding rock grade labels from contaminating the training data. Through reinforcement training with difficult samples, it improves the model's ability to identify surrounding rock grade boundary sections, fault fracture zones, water and mud inrush sections, and soft rock with large deformation sections, thus making it more flexible to use and expanding its application scope.

[0017] 3. Through a continuous learning mechanism, the model can adapt to new geological conditions as the tunnel is excavated, and the domain knowledge weights can be self-corrected, allowing fixed empirical rules to be dynamically adjusted according to the actual situation of the current project. The output results include not only the surrounding rock grade, but also the grade probability, confidence level, risk type, influencing factors, support suggestions, and supplementary measurement suggestions, which improves the construction guidance value. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the principle of a dynamic prediction method for surrounding rock classification based on machine learning and domain knowledge embedding according to the present invention. Figure 2 This is a flowchart of a dynamic prediction method for surrounding rock classification using machine learning and domain knowledge embedding, according to the present invention. Figure 3 This is a flowchart of multi-source data acquisition in a dynamic prediction method for surrounding rock classification using machine learning and domain knowledge embedding, as described in this invention. Figure 4 This is a flowchart of the training model in the dynamic prediction method for surrounding rock classification using machine learning and domain knowledge embedding of the present invention. Figure 5 This is a flowchart of the conflict verification process in the dynamic prediction method for surrounding rock classification using machine learning and domain knowledge embedding, as described in this invention. Detailed Implementation

[0019] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0020] Please see Figures 1 to 5 This invention provides a technical solution: a dynamic prediction method for surrounding rock grading using machine learning and domain knowledge embedding, comprising the following steps: S1. Establish basic engineering information and surrounding rock prediction units. Divide the tunnel into multiple mileage prediction units along the tunnel excavation direction and establish the time-mileage-excavation cycle correspondence for each mileage prediction unit. S2. Obtain multi-source surrounding rock data corresponding to each mileage prediction unit. The multi-source surrounding rock data includes face image data, advanced geological prediction data, construction equipment operation data, monitoring and measurement data, groundwater data, artificial geological sketch data, support parameter data, and post-support deformation response data. The data is then preprocessed and aligned to the mileage. S3. Construct a knowledge base for surrounding rock classification based on knowledge in fields such as surrounding rock classification standards, rock mechanics, and support feedback rules; S4. Construct a knowledge graph of surrounding rock classification based on the knowledge base of surrounding rock classification, and convert the knowledge graph of surrounding rock classification into the domain knowledge embedding vector corresponding to each mileage prediction unit. S5. Extract features from multi-source surrounding rock data to obtain features of the working face image, advanced geological prediction features, construction equipment operation features, monitoring and measurement features, and artificial geological sketch features; S6. Calculate the data credibility of each data source based on data quality, time freshness, multi-source consistency, device status, and domain knowledge consistency; S7. Input the features of the working face image, advanced geological prediction features, construction equipment operation features, monitoring and measurement features, artificial geological sketch features, domain knowledge embedding vectors and data credibility into the domain knowledge embedded multi-task prediction model to obtain the machine learning surrounding rock grade prediction results, surrounding rock grade probability, prediction uncertainty and adverse geological risk results. S8. Calculate the surrounding rock grade according to the surrounding rock classification specification, and perform conflict verification between the surrounding rock grade result and the machine learning surrounding rock grade prediction result. S9. When the degree of conflict is less than the preset threshold, the surrounding rock grade prediction result is fused and output; when the degree of conflict is greater than or equal to the preset threshold, the surrounding rock grade prediction result is conservatively corrected based on data credibility, domain knowledge consistency, adverse geological risks and support feedback results. S10. Calculate the prediction uncertainty based on the probability of surrounding rock grade, and determine the priority of active learning based on the prediction uncertainty, conflict degree, engineering risk weight, sample representativeness and domain knowledge anomaly. S11. Add mileage prediction units with active learning priority higher than the preset threshold to the active learning sample pool, and generate expert review or retest suggestions. S12. After excavation is completed, obtain the actual exposed surrounding rock grade and deformation response data after support, and calculate the label credibility based on the manual judgment results, deformation response after support, consistency of surrounding rock grade between adjacent mileages, consistency of rule classification and consistency of domain knowledge. S13. Based on the credibility of the labels, the samples are divided into strong-label samples, weak-label samples, or samples to be verified. S14. Add the prediction error samples, level boundary samples, rule conflict samples, low confidence samples, support feedback abnormal samples and domain knowledge abnormal samples to the difficult sample library, and increase the training weight of difficult samples when updating the model. S15. Based on newly revealed samples, difficult samples, historical representative samples, active supplementary test samples, support feedback samples, and domain knowledge anomaly samples, the domain knowledge embedded multi-task prediction model is continuously learned and updated. S16. Based on the actual exposed surrounding rock grade, prediction error and support feedback results, dynamically correct the domain knowledge rule weights in the surrounding rock classification domain knowledge base; S17. Output the surrounding rock grade, grade probability, confidence level, main influencing factors, adverse geological risk type, support adjustment suggestions, and supplementary measurement suggestions for the current segment and the segment to be excavated ahead.

[0021] Specifically, surrounding rock classification knowledge is incorporated into the model through knowledge features, knowledge graph embedding vectors, rule-constrained loss functions, and a dual-channel verification mechanism. This ensures that the model's predictions align with engineering experience and rock mechanics principles. Evaluations based on data quality, time freshness, multi-source consistency, equipment status, and domain knowledge consistency mitigate the impact of low-quality images, abnormal sensors, and human error on the model. Both rule-based and machine learning channels are evaluated simultaneously; when conflicts arise, verification is triggered to prevent overly optimistic or non-compliant model outputs. By prioritizing the learning of high-uncertainty, high-conflict, high-risk, and domain-knowledge-abnormal samples, blind learning is avoided. Using all samples improves model iteration efficiency. The reliability of labels is judged by support feedback, consistency of adjacent mileage, and consistency of domain knowledge, avoiding model learning bias caused by human error in classification. Weighted training can also be performed on grade boundary segments, abrupt change segments, fault fracture zones, and low-confidence samples to improve the accuracy of the model under complex working conditions. The model is continuously updated based on newly revealed data and support feedback, while retaining historical representative samples. This allows the model to learn new geology while avoiding forgetting old rules. Furthermore, the system automatically adjusts the rule weights of groundwater, structural surfaces, faults, and support feedback based on actual construction feedback, making the domain knowledge more adaptable to the current project.

[0022] In this embodiment, establishing basic engineering information and surrounding rock prediction units includes establishing a tunnel engineering coordinate system and establishing a time-mileage-excavation cycle correspondence. Establishing the tunnel engineering coordinate system includes establishing a unified spatial coordinate system with the tunnel route mileage as the main axis, dividing the tunnel into several mileage prediction units according to the excavation direction. Each prediction unit includes a prediction unit number, starting mileage, ending mileage, excavation time, support time, construction cycle number, tunnel face number, corresponding prediction point number, and corresponding advance prediction interval, unifying data from different sources into the same mileage unit.

[0023] Specifically, the prediction unit can be set to 1m, 3m, or 5m, or a single excavation cycle can be considered as a unit. Information is established for each prediction unit, such as: U36: Mileage K120+360 ~ K120+365; 72nd excavation cycle; face number F72; monitoring points M36-1 and M36-2; advance prediction interval K120+360 ~ K120+390 unifies data from different sources into a single mileage unit. After each excavation cycle, high-resolution images or video frames of the tunnel face are acquired. The acquired content includes overall images of the tunnel face, left and right shoulder images, crown images, left and right wall images, magnified images of local fractures, and images of water inflow points. Then, data such as fracture quantity, fracture length, fracture width, fracture density, joint spacing, joint dip angle, rock fragmentation degree, area of ​​weak interlayers, moist areas of the tunnel face, number of water inflow points, rock color, and texture characteristics are extracted from the images. Advanced forecasting is then performed within a certain range ahead of the tunnel face, utilizing data from ground-penetrating radar and TSP (Thunderbolt Spray). Data, transient electromagnetic data, advanced drilling data, seismic wave reflection data, and resistivity anomaly data are collected. Then, the anomalous reflection intensity, anomaly location, anomaly length, anomaly burial depth, low resistivity anomaly zone, water-bearing anomaly zone, possible location of fracture zone, possible location of fault, and location of cavity or weak interlayer are extracted from the advanced prediction data. Then, construction data, monitoring and measurement data, artificial geological sketch data, and support parameters and support feedback data are collected and processed.

[0024] Data preprocessing and mileage alignment include data cleaning, temporal alignment, spatial alignment, and data standardization. Data cleaning involves removing obvious errors, correcting sensor drift, deleting duplicate data, filling in missing time points, removing invalid images, standardizing data units, and standardizing sampling frequencies. For data from different data sources with varying sampling times, the data is aligned according to excavation cycle numbers and mileage intervals. Advanced forecast data, monitoring point data, and face image data are uniformly mapped to corresponding prediction units. Data with different dimensions is standardized to obtain standardized data that can be input into machine learning models.

[0025] Specifically, when a monitoring point suddenly exhibits an extreme value that does not conform to the continuous change pattern and there is no similar change at adjacent monitoring points, the data is marked as a suspected outlier. For data from different data sources, such as: face images: collected once per cycle; construction parameters: collected once per second; monitoring data: collected once per hour; and advance prediction: collected once every 30m or 50m, these data are then aligned according to the excavation cycle number and mileage interval. When spatial anomalies occur, such as when ground-penetrating radar detects abnormal reflections between K120+370 and K120+382, this anomaly information is assigned to prediction units such as U74, U75, and U76. Then, data of different dimensions are standardized: advance speed, cutterhead torque, settlement, water inflow, and fracture density are normalized to obtain standardized data that can be input into machine learning models.

[0026] Constructing a knowledge base for the field of surrounding rock classification includes establishing standardized classification knowledge, transforming the rules in the surrounding rock classification standards into computable knowledge, transforming field engineering experience into rules, determining support parameters through the surrounding rock grade, and then verifying the surrounding rock grade based on support feedback; then setting knowledge graph nodes and knowledge graph relationships, and for each prediction unit, extracting relevant subgraphs from the knowledge graph based on its corresponding lithology, groundwater, fissures, support feedback and other information.

[0027] Specifically, the rules in the surrounding rock classification specifications are transformed into computable knowledge, including the relationship between rock strength and surrounding rock quality, the relationship between rock mass integrity and surrounding rock quality, the correction relationship of groundwater on surrounding rock grade, the correction relationship of structural plane occurrence on surrounding rock grade, the correction relationship of initial in-situ stress on surrounding rock grade, and the correspondence between classification indicators such as BQ, RMR, and Q and surrounding rock grade. Then, the basic rock mass quality indicators, groundwater correction values, structural plane correction values, in-situ stress correction values, and corrected surrounding rock quality indicators are calculated. Rules can then be established, such as: increased fracture density reduces surrounding rock integrity; increased water inflow reduces surrounding rock stability; the appearance of fault fracture zones corrects the surrounding rock grade towards a worse state; increased crown settlement rate increases support risk; abnormally increased anchor bolt axial force increases surrounding rock pressure; abnormal fluctuations in construction torque may indicate the presence of fracture zones or alternating layers of hard and soft rock. Support parameters are then determined based on the surrounding rock grade, and the surrounding rock grade is verified in reverse based on support feedback. For example, if the predicted surrounding rock is Grade III, but it is classified as Grade III... If settlement still exceeds the limit after grade 1 support, it indicates that the original prediction level may be too optimistic, and the probability of surrounding rock deterioration in this section and similar strata ahead should be increased.

[0028] The knowledge graph nodes include lithology nodes, rock strength nodes, rock mass integrity nodes, fracture density nodes, joint spacing nodes, structural plane orientation nodes, groundwater nodes, fault fracture zone nodes, initial geostress nodes, abnormal construction parameter nodes, monitored deformation nodes, support parameter nodes, support feedback nodes, surrounding rock grade nodes, and adverse geological risk nodes. The knowledge graph relationships include influence, correction, cause, correspondence, enhancement, reduction, early warning, and reverse verification; for example: rock strength affects surrounding rock grade, rock mass integrity affects surrounding rock stability, groundwater corrects surrounding rock grade, unfavorable structural planes reduce surrounding rock stability, fault fracture zones cause surrounding rock deterioration, abnormal deformation after support reverse verifies surrounding rock grade, and surrounding rock grade corresponds to support parameters. Then, for each prediction unit, the system extracts relevant subgraphs from the knowledge graph based on its corresponding lithology, groundwater, fracture, support feedback, and other information.

[0029] Knowledge-guided data credibility evaluation includes calculating data quality coefficients, where different data sources have different quality indicators, and calculating a freshness coefficient based on time. The closer the time is to the current working face, the more credible the data. Then the system judges whether different data sources corroborate each other, and then judges the data credibility based on the equipment operating status. If a data source obviously conflicts with the knowledge of surrounding rock engineering, its credibility is reduced.

[0030] For each prediction unit, calculations are performed, including rock strength index, rock mass integrity index, groundwater correction index, structural surface correction index, geostress correction index, BQ category index, RMR category index, Q system category index, and support matching index. Then, adverse geological risk characteristics are calculated, including rock strength index, rock mass integrity index, groundwater correction index, structural surface correction index, geostress correction index, BQ category index, RMR category index, Q system category index, and support matching index. Each type of risk is expressed with a value between 0 and 1. Then, support feedback characteristics are calculated, which can also be used to determine whether the original surrounding rock grade is too optimistic.

[0031] When the face image shows dense fissures, increased construction torque fluctuations, and a higher monitoring convergence rate, and these three are consistent, the reliability is improved. When the image shows intact surrounding rock, but radar anomalies, torque anomalies, and convergence anomalies all indicate a high risk, the consistency of the image data is reduced, and the reliability of the data is judged based on the equipment's operating status. For example, if the camera is contaminated, the image reliability is reduced. If the surrounding rock is manually determined to be Class II, but groundwater increases, fissures are dense, and support deformation is abnormal, the consistency of the domain knowledge of the manually determined data is reduced.

[0032] The fusion of multi-source features and domain knowledge includes constructing model input vectors, combining various features, domain knowledge embedding vectors, and data credibility as inputs, weighting and fusing different data sources according to credibility, and then establishing a domain knowledge embedded multi-task prediction model. The model includes multiple branches, such as image recognition branch, advanced prediction branch, construction parameter time series branch, etc., and then setting multi-task outputs and multi-task constraint relationships.

[0033] Specifically, the model includes image recognition branches, advanced prediction branches, construction parameter time series branches, monitoring feedback branches, manual sketching branches, domain knowledge embedding branches, and fusion prediction branches. The model outputs include: surrounding rock grade probability, rule correction index, fault fracture risk, water and mud inrush risk, soft rock large deformation risk, rockburst risk, surrounding rock grade sudden change probability, support grade recommendation, and prediction uncertainty. Multiple tasks are mutually constrained. For example, if the water inrush risk is high, the surrounding rock grade should not be too optimistic; if the soft rock large deformation risk is high, the support recommendation should be strengthened; if the fault fracture risk is high, the probability of surrounding rock deterioration in the adjacent unit should be increased; if the support feedback is abnormal, the original prediction grade should be corrected.

[0034] The model training process includes establishing training samples, which include historical project samples, current project excavated samples, manually verified samples, and actual exposed samples. Then, domain knowledge-constrained loss functions and monotonicity constraints are set, support feedback consistency constraints are set, and the surrounding rock grade is calculated according to the standard grading method. Finally, the confidence level is calculated.

[0035] Specifically, each sample includes input features Xi, surrounding rock grade label, adverse geological condition label, support feedback label, label credibility, and sample weight. The model training loss is the sum of surrounding rock grade classification loss, domain knowledge rule consistency loss, monotonicity constraint loss, feedback consistency loss, hard sample reinforcement loss, and uncertainty loss according to their weight ratios. When rock mass integrity decreases, the predicted grade should not change significantly towards a better direction; when water inflow increases, the surrounding rock risk should not decrease; when fracture density increases, the surrounding rock integrity evaluation should not increase; when support deformation increases, the stability evaluation should not increase. If the model violates the above rules, the loss penalty is increased. If a segment is predicted to be Grade III, but the actual support results in a deformation response that would only occur with Grade IV surrounding rock, the model should be penalized. For example, if the predicted grade is good, but the actual support feedback is poor, the support feedback consistency loss is increased.

[0036] By judging whether the rock strength data is complete, whether the integrity index is reliable, whether the groundwater data is real-time, whether the structural surface occurrence is clear, whether the artificial sketch is sufficient, and whether the advanced drilling has been verified, if there are many missing data in the rule calculation, the confidence of the rule channel will decrease.

[0037] Machine learning channel prediction includes the model outputting the surrounding rock grade, and then outputting interpretable results to facilitate engineers' judgment. When a conflict occurs, the surrounding rock grade is digitized, the degree of conflict is calculated, and the conflict level is determined. Then, the uncertainty is calculated based on the probability of the surrounding rock grade, and the confidence zone is determined. Learning and retesting include active learning priority and screening of active learning samples. When the active learning priority is high, retesting suggestions are generated, and the retesting data is re-entered into multi-source data collection.

[0038] Specifically, when the rule channel is poor and the machine learning channel is good, if there are simultaneously issues such as increased groundwater levels, increased fault risk, increased convergence rate, and poor image quality, the final result will be corrected to the poorer level. When the machine learning channel is poor and the rule channel is good, if both construction parameters and monitoring data are abnormal, a risk warning will be output and supplementary measurements will be recommended. When the two channels conflict and the data reliability is low, a level range will be output instead of a single level.

[0039] Uncertainty is calculated based on the probability of surrounding rock grade. If the probability of a certain grade is significantly the highest, the uncertainty is low. If the probabilities of multiple grades are close, the uncertainty is high.

[0040] When active learning has a high priority, the system generates supplementary measurement suggestions, such as: suggesting to conduct advance drilling 8m ahead of K120+365, suggesting to re-scan the right arch waist area with ground-penetrating radar, and suggesting to re-acquire high-definition images of the working face. The supplementary measurement data then re-enters the multi-source data acquisition module.

[0041] When it is necessary to calculate the credibility of the labels, the judgment is made through manual judgment, support feedback, rule classification, and domain knowledge consistency. Then the samples are divided into strong label samples, weak label samples, and samples to be reviewed.

[0042] Establish a continuous learning sample set to prevent the model from forgetting. If only the latest samples are used for training, the model may only be adapted to the current stratum and forget the historical stratum patterns. Therefore, historical representative samples are added with each update to achieve self-correction of domain knowledge weights.

[0043] Workflow: Establish basic engineering information and surrounding rock prediction units. Divide the tunnel into multiple mileage prediction units along the tunnel excavation direction and establish the time-mileage-excavation cycle correspondence for each mileage prediction unit; acquire multi-source surrounding rock data corresponding to each mileage prediction unit, including tunnel face image data, advanced geological prediction data, construction equipment operation data, monitoring and measurement data, groundwater data, artificial geological sketch data, support parameter data, and post-support deformation response data, and preprocess and mileage-align the data; construct a surrounding rock classification knowledge base based on knowledge in the fields of surrounding rock classification specifications, rock mechanics, and support feedback rules; construct a surrounding rock classification knowledge graph based on the surrounding rock classification knowledge base, and classify the surrounding rock classification... The knowledge graph is converted into domain knowledge embedding vectors corresponding to each mileage prediction unit; feature extraction is performed on multi-source surrounding rock data to obtain face image features, advanced geological prediction features, construction equipment operation features, monitoring and measurement features, and artificial geological sketch features; data credibility of each data source is calculated based on data quality, time freshness, multi-source consistency, equipment status, and domain knowledge consistency; face image features, advanced geological prediction features, construction equipment operation features, monitoring and measurement features, artificial geological sketch features, domain knowledge embedding vectors, and data credibility are input into a domain knowledge embedded multi-task prediction model to obtain machine learning surrounding rock grade prediction results, surrounding rock grade probability, prediction uncertainty, and adverse geological risk results; based on the surrounding rock classification... The system calculates the surrounding rock grade according to the standard and performs conflict verification with the machine learning surrounding rock grade prediction results. When the conflict level is less than a preset threshold, the surrounding rock grade prediction results are merged and output. When the conflict level is greater than or equal to the preset threshold, the surrounding rock grade prediction results are conservatively corrected based on data credibility, domain knowledge consistency, adverse geological risks, and support feedback results. The prediction uncertainty is calculated based on the surrounding rock grade probability, and the active learning priority is determined based on prediction uncertainty, conflict level, engineering risk weight, sample representativeness, and domain knowledge anomaly. Mileage prediction units with an active learning priority higher than the preset threshold are added to the active learning sample pool, and expert review or supplementary measurement suggestions are generated. After excavation is completed, the actual mileage prediction results are obtained. The system discloses surrounding rock grade and deformation response data after support, and calculates label credibility based on manual judgment results, deformation response after support, consistency of surrounding rock grade between adjacent mileages, consistency of rule classification, and consistency of domain knowledge. Based on the label credibility, samples are divided into strongly labeled samples, weakly labeled samples, or samples awaiting verification. Predicted error samples, grade boundary samples, rule conflict samples, low-confidence samples, abnormal support feedback samples, and abnormal domain knowledge samples are added to a difficult sample library, and the training weight of difficult samples is increased during model updates. Based on newly disclosed samples, difficult samples, historical representative samples, actively supplemented test samples, support feedback samples, and abnormal domain knowledge samples, the domain knowledge embedded multi-task prediction model is continuously learned and updated.Based on the actual exposed surrounding rock grade, prediction errors, and support feedback results, the weights of the domain knowledge rules in the surrounding rock classification domain knowledge base are dynamically corrected; the surrounding rock grade, grade probability, confidence level, main influencing factors, adverse geological risk types, support adjustment suggestions, and supplementary measurement suggestions are output for the current segment and the segment to be excavated ahead.

[0044] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A dynamic prediction method for surrounding rock grading using machine learning and domain knowledge embedding, characterized in that, Includes the following steps: S1. Establish basic engineering information and surrounding rock prediction units. Divide the tunnel into multiple mileage prediction units along the tunnel excavation direction and establish the time-mileage-excavation cycle correspondence for each mileage prediction unit. S2. Obtain multi-source surrounding rock data corresponding to each mileage prediction unit. The multi-source surrounding rock data includes face image data, advanced geological prediction data, construction equipment operation data, monitoring and measurement data, groundwater data, artificial geological sketch data, support parameter data, and post-support deformation response data. The data is then preprocessed and aligned to the mileage. S3. Construct a knowledge base for surrounding rock classification based on knowledge in fields such as surrounding rock classification standards, rock mechanics, and support feedback rules; S4. Construct a knowledge graph of surrounding rock classification based on the knowledge base of surrounding rock classification, and convert the knowledge graph of surrounding rock classification into the domain knowledge embedding vector corresponding to each mileage prediction unit. S5. Extract features from multi-source surrounding rock data to obtain features of the working face image, advanced geological prediction features, construction equipment operation features, monitoring and measurement features, and artificial geological sketch features; S6. Calculate the data credibility of each data source based on data quality, time freshness, multi-source consistency, device status, and domain knowledge consistency; S7. Input the features of the working face image, advanced geological prediction features, construction equipment operation features, monitoring and measurement features, artificial geological sketch features, domain knowledge embedding vectors and data credibility into the domain knowledge embedded multi-task prediction model to obtain the machine learning surrounding rock grade prediction results, surrounding rock grade probability, prediction uncertainty and adverse geological risk results. S8. Calculate the surrounding rock grade according to the surrounding rock classification specification, and perform conflict verification between the surrounding rock grade result and the machine learning surrounding rock grade prediction result. S9. When the degree of conflict is less than the preset threshold, the predicted results of the surrounding rock grade are fused and output. When the degree of conflict is greater than or equal to the preset threshold, the prediction results of the surrounding rock grade are conservatively corrected based on data credibility, consistency of domain knowledge, adverse geological risks and support feedback results. S10. Calculate the prediction uncertainty based on the probability of surrounding rock grade, and determine the priority of active learning based on the prediction uncertainty, conflict degree, engineering risk weight, sample representativeness and domain knowledge anomaly. S11. Add mileage prediction units with active learning priority higher than the preset threshold to the active learning sample pool, and generate expert review or retest suggestions. S12. After excavation is completed, obtain the actual exposed surrounding rock grade and deformation response data after support, and calculate the label credibility based on the manual judgment results, deformation response after support, consistency of surrounding rock grade between adjacent mileages, consistency of rule classification and consistency of domain knowledge. S13. Based on the credibility of the labels, the samples are divided into strong-label samples, weak-label samples, or samples to be verified. S14. Add the prediction error samples, level boundary samples, rule conflict samples, low confidence samples, support feedback abnormal samples and domain knowledge abnormal samples to the difficult sample library, and increase the training weight of difficult samples when updating the model. S15. Based on newly revealed samples, difficult samples, historical representative samples, active supplementary test samples, support feedback samples, and domain knowledge anomaly samples, the domain knowledge embedded multi-task prediction model is continuously learned and updated. S16. Based on the actual exposed surrounding rock grade, prediction error and support feedback results, dynamically correct the domain knowledge rule weights in the surrounding rock classification domain knowledge base; S17. Output the surrounding rock grade, grade probability, confidence level, main influencing factors, adverse geological risk type, support adjustment suggestions, and supplementary measurement suggestions for the current segment and the segment to be excavated ahead.

2. The method for dynamic prediction of surrounding rock grading based on machine learning and domain knowledge embedding according to claim 1, characterized in that: The establishment of basic engineering information and surrounding rock prediction unit includes establishing a tunnel engineering coordinate system and establishing the correspondence between time, mileage and excavation cycle. The established tunnel engineering coordinate system includes a unified spatial coordinate system based on the tunnel mileage as the main axis. The tunnel is divided into several mileage prediction units according to the excavation direction. Each prediction unit includes prediction unit number, starting mileage, ending mileage, excavation time, support time, construction cycle number, tunnel face number, corresponding prediction point number, and corresponding advance prediction interval, unifying data from different sources into the same mileage unit.

3. The method for dynamic prediction of surrounding rock grading based on machine learning and domain knowledge embedding according to claim 1, characterized in that: The data preprocessing and mileage alignment include data cleaning, time alignment, spatial alignment, and data standardization. Specifically, data cleaning involves removing obvious errors, correcting sensor drift, deleting duplicate data, filling in missing time points, removing invalid images, standardizing data units, and standardizing sampling frequencies. For data from different data sources with varying sampling times, the data is aligned according to excavation cycle numbers and mileage intervals. Advanced forecast data, monitoring point data, and face image data are uniformly mapped to corresponding prediction units. Data with different dimensions are standardized to obtain standardized data that can be input into machine learning models.

4. The method for dynamic prediction of surrounding rock grading based on machine learning and domain knowledge embedding according to claim 1, characterized in that: The construction of the knowledge base for surrounding rock classification includes establishing standardized classification knowledge, transforming the rules in the surrounding rock classification specifications into computable knowledge, transforming field engineering experience into rules, determining support parameters through the surrounding rock grade, and then verifying the surrounding rock grade based on support feedback; then setting knowledge graph nodes and knowledge graph relationships, and for each prediction unit, extracting relevant subgraphs from the knowledge graph based on its corresponding lithology, groundwater, fissures, support feedback and other information.

5. The method for dynamic prediction of surrounding rock grading based on machine learning and domain knowledge embedding according to claim 1, characterized in that: The knowledge-guided data credibility evaluation includes calculating the data quality coefficient, where different data sources have different quality indicators, and calculating the freshness coefficient based on time. The closer the time is to the current working face, the more credible the data. Then the system judges whether different data sources corroborate each other, and then judges the data credibility based on the equipment operating status. If a data source obviously conflicts with the knowledge of surrounding rock engineering, its credibility is reduced.

6. The method for dynamic prediction of surrounding rock grading based on machine learning and domain knowledge embedding according to claim 1, characterized in that: The process of fusing multi-source features and domain knowledge includes constructing a model input vector, combining various features, domain knowledge embedding vectors, and data credibility as input, weighting and fusing different data sources according to credibility, and then establishing a domain knowledge embedded multi-task prediction model. The model includes multiple branches, such as image recognition branch, advanced prediction branch, construction parameter time series branch, etc., and then setting multi-task outputs and multi-task constraint relationships.

7. The method for dynamic prediction of surrounding rock grading based on machine learning and domain knowledge embedding according to claim 1, characterized in that: The model training process includes establishing training samples, which include historical project samples, current project excavated samples, manually verified samples, and actual exposed samples. Then, domain knowledge constraint loss function and monotonicity constraint are set, support feedback consistency constraint is set, and then the surrounding rock grade is calculated according to the standard grading method. Finally, the confidence level is calculated.

8. The method for dynamic prediction of surrounding rock grading based on machine learning and domain knowledge embedding according to claim 1, characterized in that: The machine learning channel prediction includes outputting the surrounding rock grade from the model, and then outputting interpretable results to facilitate engineers' judgment.

9. The method for dynamic prediction of surrounding rock grading based on machine learning and domain knowledge embedding according to claim 1, characterized in that: When a conflict occurs, the surrounding rock grade is digitized, the degree of conflict is calculated, and the conflict level is determined. Then, the uncertainty is calculated based on the probability of the surrounding rock grade, and the confidence zone is determined.

10. The method for dynamic prediction of surrounding rock grading based on machine learning and domain knowledge embedding according to claim 1, characterized in that: The learning and retesting process includes prioritizing active learning and selecting active learning samples. When the active learning priority is high, retesting suggestions are generated, and the retesting data is re-entered into multi-source data collection.

Images