Tunnel construction quality control method

Abstract

The application discloses a tunnel construction quality control method. The method comprises the following steps: obtaining physical property parameters of surrounding rock in front of a working face, real-time deformation data of surrounding rock of an excavated section and real-time stress data of a supporting structure; inputting the data into a pre-trained surrounding rock stability dynamic prediction network to output a stability grade of the current surrounding rock and a deformation prediction value of a next excavation cycle; obtaining supporting parameters of the current construction cycle, inputting the stability grade, the deformation prediction value and the current supporting parameters into a pre-built supporting decision reinforcement learning model to output a supporting parameter adjustment scheme of the next excavation cycle; and sending the supporting parameter adjustment scheme to a construction equipment control system through an industrial communication protocol so as to automatically perform work according to the scheme in the next excavation cycle construction. The application is mainly used for surrounding rock stability control and construction quality improvement in a tunnel construction process.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for tunnel construction, and specifically relates to a method for quality control in tunnel construction. Background Technology

[0002] During tunnel construction, the stability of the surrounding rock directly affects construction safety and project quality. As the tunnel face advances, the stress distribution and deformation characteristics of the surrounding rock are constantly changing, and the load borne by the support structure also changes accordingly. Existing tunnel construction quality control methods mainly rely on pre-established support plans and the experience judgment of on-site technicians, which have several shortcomings in actual construction.

[0003] First, the advanced identification and stability prediction of surrounding rock conditions are fundamental to support decisions. Traditional methods typically employ geological sketching, advanced drilling, or geophysical exploration to obtain geological information ahead of the tunnel face, and then combine this with engineering analogies or numerical simulations to classify the stability of the surrounding rock. However, geological conditions exhibit strong nonlinearity and spatiotemporal variability, and the physical property parameters obtained based on a limited number of measuring points cannot fully reflect the true state of the surrounding rock. Monitoring sensors deployed in the excavated section can provide real-time data on surrounding rock deformation and support stress, but these data are not synchronized with the unrevealed surrounding rock conditions ahead of the tunnel face in terms of time and space. How to effectively integrate the physical property parameters obtained from advanced exploration with the real-time monitoring data from the excavated section to improve the accuracy and timeliness of surrounding rock stability prediction is a challenge faced by existing technologies.

[0004] Secondly, regarding the dynamic adjustment of support parameters, current construction operations typically select the corresponding support scheme from a pre-set support parameter table based on the surrounding rock stability classification results. This table-based decision-making method suffers from response lag and insufficient precision. On the one hand, the pre-set support parameter tables are often based on limited typical geological conditions, making it difficult to cover the complex and variable situations encountered in actual construction. On the other hand, there is a coupling relationship between the steel arch spacing, shotcrete thickness, and anchor bolt arrangement parameters; adjustments to any of these parameters will affect the overall mechanical behavior of the surrounding rock-support system. There is a lack of systematic technical means to quickly determine reasonable support parameter adjustments under multi-parameter coupling conditions based on the current surrounding rock condition and future deformation trends.

[0005] Furthermore, at the construction execution level, instructions for adjusting support parameters need to be transmitted to each piece of construction equipment and the parameters need to be preset. In traditional workflows, this process relies on manual communication and manual setting, which is not only inefficient but may also lead to information transmission distortion or execution delays, affecting the timeliness and accuracy of support adjustments.

[0006] The aforementioned problems combined result in a lack of a systematic closed loop for controlling surrounding rock stability in current tunnel construction, encompassing perception, prediction, decision-making, and execution. This makes it difficult to achieve forward-looking and precise adjustments to support parameters. Establishing an effective linkage between dynamic prediction of surrounding rock deformation and adaptive decision-making of support parameters based on multi-source heterogeneous monitoring data, and automatically transmitting the decision results to construction equipment for execution, is a pressing technical problem that needs to be solved in this field. Summary of the Invention

[0007] One object of the present invention is to solve at least the above-mentioned problems and / or defects, and to provide at least the advantages described below.

[0008] The purpose of this invention is to provide a method for controlling the quality of tunnel construction. This method effectively integrates advanced detection information of the surrounding rock in front of the tunnel face with real-time monitoring data of the excavated section to dynamically predict the stability level and deformation trend of the surrounding rock. The prediction results are then converted into an adaptive adjustment scheme for support parameters, and the adjustment scheme is automatically executed through equipment communication protocols, thereby improving the quality of tunnel construction.

[0009] One object of the present invention is to provide a method for quality control in tunnel construction, comprising the following steps: To obtain the physical properties of the surrounding rock in front of the tunnel face during tunnel construction, as well as the real-time deformation data of the surrounding rock in the excavated section and the real-time stress data of the support structure; The physical property parameters, real-time deformation data, and real-time stress data are input into a pre-trained dynamic prediction network for surrounding rock stability. The dynamic prediction network integrates multi-source heterogeneous data through a long short-term memory network and outputs the current stability level of the surrounding rock and the deformation prediction value for the next excavation cycle. Obtain the support parameters of the current construction cycle, input the stability level, the predicted deformation value and the support parameters of the current construction cycle into a pre-built support decision reinforcement learning model, and output the support parameter adjustment scheme for the next excavation cycle from the support decision reinforcement learning model. The support parameter adjustment scheme includes at least one of the following: steel arch spacing adjustment, shotcrete thickness adjustment and anchor bolt arrangement parameter adjustment. The support parameter adjustment plan is sent to the construction equipment control system via an industrial communication protocol so that the operation is carried out in accordance with the support parameter adjustment plan in the next excavation cycle.

[0010] Preferably, in the tunnel construction quality control method, the training process of the dynamic prediction network for surrounding rock stability includes: The tunnel construction history data includes physical property parameters of the surrounding rock in front of the tunnel face collected at different construction stages, deformation data of the surrounding rock in the excavated section, stress data of the support structure, current support parameters used for each construction stage, and actual deformation recorded for each construction stage. Using the physical properties of the surrounding rock in front of the tunnel face, the deformation data of the excavated section of the surrounding rock, and the stress data of the support structure from the historical data as input samples, the actual deformation is mapped to the corresponding stability level label. The actual deformation and the stability level label are used together as training labels to train the long short-term memory network to obtain the dynamic prediction network for surrounding rock stability.

[0011] Preferably, in the tunnel construction quality control method, training the constructed long short-term memory network includes: The parameters of the Long Short-Term Memory network are optimized using the weighted sum of the deformation prediction loss and the stability level prediction loss as the objective function until the objective function converges.

[0012] Preferably, in the tunnel construction quality control method, the construction process of the support decision reinforcement learning model includes: The physical properties of the surrounding rock in front of the tunnel face, the deformation data of the excavated section of the surrounding rock, and the stress data of the support structure from the historical data are input into the dynamic prediction network for surrounding rock stability to obtain the stability level and the predicted deformation value. The current support parameters corresponding to each construction stage are obtained from the historical data. The stability level, the predicted deformation value, and the current support parameters are used together as the state, the support parameter adjustment scheme actually implemented in the historical data is used as the action, and the deviation between the actual deformation amount after the action and the preset safe deformation threshold is used to construct a reward function to train the support decision reinforcement learning model to be trained.

[0013] Preferably, in the tunnel construction quality control method, the weight of the deformation prediction loss in the objective function is greater than the weight of the stability level prediction loss.

[0014] Preferably, in the tunnel construction quality control method, the weight of the deformation prediction loss is 0.6 to 0.7, the weight of the stability level prediction loss is 0.3 to 0.4, and the sum of the weights of the deformation prediction loss and the stability level prediction loss is 1.

[0015] Preferably, in the tunnel construction quality control method, the construction equipment control system includes a central controller and multiple equipment controllers communicatively connected to the central controller. The central controller receives the support parameter adjustment scheme and distributes the steel arch spacing adjustment amount, the shotcrete thickness adjustment amount, and the anchor bolt arrangement parameter adjustment amount in the support parameter adjustment scheme to the corresponding equipment controllers.

[0016] Preferably, in the tunnel construction quality control method, the support parameter adjustment scheme includes the correspondence between the equipment identifier and the adjustment amount of the steel arch spacing, the adjustment amount of the shotcrete thickness, and the adjustment amount of the anchor bolt arrangement parameters. The central controller sends the adjustment amount of the steel arch spacing, the adjustment amount of the shotcrete thickness, and the adjustment amount of the anchor bolt arrangement parameters to the corresponding equipment controller according to the equipment identifier.

[0017] Preferably, in the tunnel construction quality control method, after sending the steel arch spacing adjustment amount, the shotcrete thickness adjustment amount, and the anchor bolt arrangement parameter adjustment amount to the corresponding equipment controller, the method further includes: the central controller receiving the execution confirmation signal returned by each of the equipment controllers, wherein the execution confirmation signal indicates that the construction equipment corresponding to the equipment controller has completed the parameter preset according to the received steel arch spacing adjustment amount, shotcrete thickness adjustment amount, or anchor bolt arrangement parameter adjustment amount.

[0018] Preferably, in the tunnel construction quality control method, after receiving the execution confirmation signal returned by each of the equipment controllers, the central controller further includes: the central controller associating and storing the support parameter adjustment scheme, the execution confirmation signal, and the sequence number of the excavation cycle to which the support parameter adjustment scheme is applied in the construction log database.

[0019] The present invention has at least the following beneficial effects: This invention acquires the physical properties of the surrounding rock ahead of the tunnel face, as well as real-time deformation and stress data of the excavated section. It then uses a long short-term memory network to fuse multi-source heterogeneous data, outputting the surrounding rock stability level and deformation prediction values. The prediction results, along with the current support parameters, are input into a reinforcement learning model to output an adjustment scheme for the support parameters in the next excavation cycle. This adjustment scheme is then sent to the construction equipment control system for automatic execution via an industrial communication protocol. This process links perception, prediction, decision-making, and execution into a complete closed-loop flow, reducing delays and deviations caused by manual judgment and instruction transmission. It enables support parameters to be adjusted according to dynamic changes in surrounding rock conditions, improving the stability of the surrounding rock during tunnel construction and enhancing construction quality.

[0020] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Detailed Implementation

[0021] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0022] This invention provides a method for quality control in tunnel construction, comprising the following steps: acquiring the physical properties of the surrounding rock ahead of the tunnel face during tunnel construction, as well as real-time deformation data of the excavated section of the surrounding rock and real-time stress data of the support structure; inputting the physical properties, real-time deformation data, and real-time stress data into a pre-trained dynamic prediction network for surrounding rock stability, wherein the dynamic prediction network fuses multi-source heterogeneous data through a long short-term memory network, and outputs the stability level of the current surrounding rock and the deformation prediction value for the next excavation cycle; acquiring the support parameters for the current construction cycle, inputting the stability level, the deformation prediction value, and the support parameters for the current construction cycle into a pre-constructed reinforcement learning model for support decision-making, wherein the reinforcement learning model for support decision-making outputs a support parameter adjustment scheme for the next excavation cycle, wherein the support parameter adjustment scheme includes at least one of steel arch spacing adjustment, shotcrete thickness adjustment, and anchor bolt arrangement parameter adjustment; and sending the support parameter adjustment scheme to the construction equipment control system via an industrial communication protocol, so that operations are performed according to the support parameter adjustment scheme in the next excavation cycle.

[0023] In existing technologies, there is a lack of systematic linkage between surrounding rock stability prediction and support parameter decision-making during tunnel construction, and support adjustments often lag behind changes in surrounding rock conditions. In traditional construction processes, surrounding rock stability assessment mainly relies on geological sketches, advanced drilling, and field experience. There is a temporal and spatial asynchrony between monitoring data from excavated sections and information about the unexposed surrounding rock ahead of the tunnel face, making it difficult to effectively integrate these multi-source data to support dynamic prediction. In the decision-making stage, current practices typically select corresponding schemes from a pre-set support parameter table based on surrounding rock classification results. This lookup-based method has limited response speed and struggles to provide precise adjustments under multi-parameter coupling conditions. In the execution stage, adjustment instructions rely on manual transmission and setting, posing risks of information delays and misoperation. These factors collectively make it difficult for existing methods to achieve closed-loop control from perception, prediction, decision-making to execution.

[0024] The method provided in this embodiment includes the following steps.

[0025] Step one involves acquiring the physical properties of the surrounding rock ahead of the tunnel face during tunnel construction, as well as real-time deformation data of the excavated section and real-time stress data of the support structure. The physical properties of the surrounding rock ahead of the tunnel face can be obtained through advanced geological prediction methods, such as using seismic wave methods or ground-penetrating radar to detect physical indicators reflecting the integrity and weathering degree of the surrounding rock, such as wave velocity and resistivity, within a certain range ahead of the tunnel face. Specifically, the physical properties include at least one or more of the following indicators: seismic wave velocity, resistivity, rock integrity coefficient (Kv), rock quality index (RQD), and elastic modulus. These indicators are acquired through an advanced geological prediction system. Real-time deformation data of the excavated section's surrounding rock can be collected using displacement sensors embedded in the arch and sidewalls, while real-time stress data of the support structure can be collected using stress gauges installed in the steel arch frame or shotcrete layer. These data reflect the state of the surrounding rock and support system from both spatial and temporal dimensions, with the physical properties being predictive and the deformation and stress data being real-time.

[0026] To clarify the meaning of the terms, in this invention, the real-time deformation data refers to the unprocessed raw deformation reading sequence continuously collected by monitoring sensors during construction, which is used to reflect the dynamic evolution process of surrounding rock deformation; the actual deformation amount refers to the final stable value or maximum deformation value measured on the preset monitoring section after the completion of one construction cycle of excavation and before the start of the next cycle of excavation, which serves as the true result label of this construction stage.

[0027] Step two involves inputting the aforementioned physical property parameters, real-time deformation data, and real-time stress data into a pre-trained dynamic prediction network for surrounding rock stability. This dynamic prediction network uses a Long Short-Term Memory (LSTM) network as its basic architecture, which is capable of handling long-term dependencies in time-series data. When fusing multi-source heterogeneous data, physical property parameters serve as the static feature input of the surrounding rock ahead of the tunnel face, while real-time deformation and stress data serve as the time-series feature input of the excavated section. The network uses internal memory units and gating mechanisms to fuse these two types of data from different sources with different physical meanings, extracting feature representations related to the evolution of surrounding rock stability. The network outputs the current stability level of the surrounding rock and the predicted deformation value for the next excavation cycle. The stability level can be divided into several levels according to the degree of stability to instability of the surrounding rock, and the predicted deformation value gives the maximum possible deformation of the surrounding rock in the next excavation cycle, providing a quantitative basis for subsequent support decisions.

[0028] Specifically, the Long Short-Term Memory (LSTM) network in this embodiment adopts a multi-input channel structure: the first channel receives the physical properties (static features) of the surrounding rock in front of the tunnel face, and repeatedly inputs this feature vector at each time step; the second channel receives the real-time deformation data (time series) of the surrounding rock in the excavated section, and inputs it sequentially according to the time steps; the third channel receives the real-time stress data (time series) of the support structure, and inputs it sequentially according to the time steps. At each time step, the LSTM unit concatenates the feature vectors of the three channels to form a fused feature vector, which is then input to the input gate, forget gate, and output gate respectively. Through the gating mechanism, the network adaptively learns the coupling relationship between static geological parameters and dynamic monitoring data. The output layer of the LSTM connects two branches: one branch outputs the surrounding rock stability level through a softmax classification layer; the other branch outputs the deformation prediction value for the next excavation cycle through a linear regression layer. This network structure achieves effective fusion of multi-source heterogeneous ports.

[0029] Step 3: Obtain the support parameters for the current construction cycle. Input the stability level and deformation prediction values ​​obtained in the previous steps, along with the current support parameters, into the pre-constructed support decision reinforcement learning model. The support parameters for the current construction cycle include the currently used steel arch spacing, shotcrete thickness, and anchor bolt arrangement parameters, reflecting the current state of the support system. For the first construction cycle, the support parameters for the current cycle use the initial support parameter values ​​preset in the design. The reinforcement learning model uses the stability level and deformation prediction values ​​as state inputs to perceive the current and future risk level of the surrounding rock, while using the current support parameters as a component of the state, enabling the model to fully understand the existing support conditions when making decisions. The model outputs a support parameter adjustment scheme for the next excavation cycle. The adjustment scheme includes at least one of the following: steel arch spacing adjustment, shotcrete thickness adjustment, and anchor bolt arrangement parameter adjustment. The steel arch spacing adjustment represents the increase or decrease in the arch installation spacing in the next cycle relative to the current spacing; the shotcrete thickness adjustment represents the increase or decrease in the shotcrete layer thickness relative to the current thickness; and the anchor bolt arrangement parameter adjustment includes adjustments to the anchor bolt length, spacing, or number. The reinforcement learning model has learned the mapping relationship between the surrounding rock state and the support adjustment action during the training process, and can output an adjustment scheme that controls the deformation of the surrounding rock within a safe range under given input conditions.

[0030] It should be noted that for the first construction cycle, since no support structure has been constructed yet, the 'current support parameters' are the initial support parameter values ​​from the tunnel design documents (e.g., the initial spacing of the steel arches, the initial thickness of the shotcrete, and the initial arrangement parameters of the anchor bolts specified in the design). In subsequent construction cycles, the current support parameters are obtained from the parameters obtained after the previous construction cycle was executed.

[0031] Step four involves sending the support parameter adjustment plan to the construction equipment control system via an industrial communication protocol, ensuring that the operation is carried out according to this plan in the next excavation cycle. The industrial communication protocol can be a commonly used communication protocol in tunnel construction equipment, connecting the upper control system with each piece of construction equipment. Upon receiving the adjustment plan, the construction equipment control system sends the adjustments for the steel arch spacing, shotcrete thickness, and anchor bolt arrangement parameters to the corresponding construction equipment. Each piece of equipment automatically presets its parameters before the start of the next excavation cycle, thus achieving automatic execution of the support plan. Compared to manual communication and settings, automatic execution reduces information transmission links and time delays, and also lowers the possibility of human error.

[0032] Through the above steps, this implementation method integrates advanced surrounding rock detection, real-time monitoring of excavated sections, dynamic prediction of surrounding rock stability, adaptive decision-making of support parameters, and automatic execution of construction equipment into a complete closed-loop process. Compared with existing lookup-based decision-making and manual execution methods, this method enables support parameters to be dynamically adjusted according to actual changes in surrounding rock conditions, improving surrounding rock stability and construction quality during tunnel construction.

[0033] In a preferred embodiment, the tunnel construction quality control method includes the following training process for the dynamic prediction network of surrounding rock stability: acquiring historical tunnel construction data, which includes physical property parameters of the surrounding rock in front of the tunnel face, deformation data of the excavated section of the surrounding rock, stress data of the support structure, current support parameters used for each construction stage, and actual deformation recorded for each construction stage; using the physical property parameters of the surrounding rock in front of the tunnel face, deformation data of the excavated section of the surrounding rock, and stress data of the support structure from the historical data as input samples, mapping the actual deformation to a corresponding stability level label, and using the actual deformation and the stability level label together as training labels to train the long short-term memory network to obtain the dynamic prediction network of surrounding rock stability.

[0034] Rock stability prediction models can be classified and judged solely based on geological descriptions near the working face or a single geophysical index, or they can be extrapolated based solely on displacement monitoring data from the excavated section. The former ignores the actual response information of the rock mass revealed by the excavated section, while the latter lacks advance perception of the geological conditions of the unexcavated section ahead. When attempting to use advance detection data and monitoring data from the excavated section together to train the prediction model, a difficulty arises because the two types of data are not synchronized in time. Advance detection acquires information about the unexcavated rock mass ahead, while deformation and stress monitoring records the response of the excavated rock mass behind, and there is no clear alignment between the two.

[0035] This implementation provides a method for constructing training data and training models to address the aforementioned problems. First, historical tunnel construction data is acquired, which comes from multiple construction stages of the same tunnel or neighboring tunnels with similar geological conditions. The historical data includes physical properties of the surrounding rock ahead of the tunnel face, deformation data of the excavated section of the surrounding rock, and stress data of the support structure collected at different construction stages. The physical properties of the surrounding rock ahead of the tunnel face can be indicators such as wave velocity and resistivity extracted from advanced geological prediction records. The deformation data of the excavated section of the surrounding rock can be displacement values ​​obtained from arch settlement and perimeter convergence monitoring records. The stress data of the support structure can be load values ​​obtained from steel arch stress gauges or concrete strain gauge records. The historical data also includes the current support parameters used for each construction stage, i.e., the specific values ​​of the actual steel arch spacing, shotcrete thickness, and anchor bolt arrangement parameters used in that construction stage. These parameters reflect the support conditions at each construction stage. Importantly, the historical data also includes the actual deformation recorded for each construction stage. This actual deformation refers to the actual measured value of the surrounding rock after the excavation of that construction stage is completed. It reflects the true response of the surrounding rock under the conditions of the surrounding rock and the support conditions at that construction stage.

[0036] The input samples consist of the physical properties of the surrounding rock ahead of the tunnel face, the deformation data of the excavated section of the surrounding rock, and the stress data of the support structure, all from historical data. These three types of data constitute the input part of a training sample. The physical properties provide the inherent geological attributes of the surrounding rock ahead, the deformation data reflects the current motion state of the surrounding rock, and the stress data reflects the current bearing state of the support structure. Together, they describe all the surrounding rock and support status information that can be obtained before the start of a construction cycle. The actual deformation recorded for that construction stage is mapped to the corresponding stability level label. The mapping method can be based on a preset deformation classification threshold. For example, deformation less than the first threshold is mapped to a stable level, deformation between the first and second thresholds is mapped to a basic stable level, and so on, thus converting continuous deformation values ​​into discrete stability levels. Specifically, the first threshold is set to 15 mm, the second threshold to 30 mm, and the third threshold to 50 mm. Actual deformation less than 15mm is mapped to Level I (stable); 15mm to 30mm to Level II (basically stable); 30mm to 50mm to Level III (understood); and greater than 50mm to Level IV (unstable). These thresholds can be adjusted according to tunnel design specifications and engineering experience; different tunnel projects may use different grading standards. The actual deformation and the stability level label mapped from it are used together as training labels. Thus, the input to a training sample consists of advanced physical property parameters, excavated section deformation data, and stress data for a specific construction stage, with the labels being the actual deformation and stability level corresponding to that construction stage.

[0037] The deformation data of the surrounding rock in the excavated section and the stress data of the support structure are both taken from the monitoring data within one hour before the completion of the excavation of that construction stage. The actual deformation is taken from the final stable deformation value or the maximum deformation value from the completion of the excavation of that construction stage to the start of the next excavation stage. There is no time overlap between the input data and the label data.

[0038] Using the training dataset constructed above, the Long Short-Term Memory (LSTM) network to be trained is trained. During training, the network receives input samples and fuses physical property parameters, deformation data, and stress data through internal LSM units and gating mechanisms. It outputs predicted deformation and stability levels, compares these with the actual deformation and stability level labels, calculates the loss function, and updates the network parameters accordingly. Through iterative training with a large number of historical samples, the network gradually learns the mapping relationship between advance detection information and monitoring information of excavated sections to future deformation and stability levels. When the loss function converges or the training reaches a preset number of rounds, the trained dynamic prediction network for surrounding rock stability is obtained. This network can receive real-time data acquired in the current construction cycle during the online prediction phase and output the current stability level of the surrounding rock and the predicted deformation value for the next excavation cycle.

[0039] In a preferred embodiment, the tunnel construction quality control method includes training the constructed long short-term memory network by optimizing the parameters of the long short-term memory network with the objective function of minimizing the weighted sum of deformation prediction loss and stability level prediction loss, until the objective function converges.

[0040] When a neural network needs to output both continuous numerical values ​​and discrete categories simultaneously, the two tasks can be trained separately: first, train the deformation prediction task, and then train the stability level classification task with fixed network parameters. This phased training approach has a drawback: subsequent training phases may overwrite the feature representations learned in the previous phase, leading to a decrease in the performance of the first trained task. Alternatively, the two losses can be simply added together. However, because the deformation prediction loss and the stability level prediction loss differ in numerical magnitude and convergence characteristics, simple addition may cause one loss to dominate the training process, making it difficult for the other task to achieve the expected prediction results.

[0041] This implementation provides a method for constructing the objective function for multi-task joint training. When training the constructed Long Short-Term Memory (LSTM) network, the objective function is the weighted sum of the deformation prediction loss and the stability level prediction loss. The deformation prediction loss measures the deviation between the network's output predicted deformation and the actual deformation in the training labels; this loss term constrains the network's ability to regress and predict deformation values. The stability level prediction loss measures the difference between the network's output predicted stability level and the stability level labels in the labels; this loss term constrains the network's ability to classify and discriminate stability levels. Each of the two losses reflects the degree of completion of different prediction tasks; the weighted sum of the two losses is used as the overall objective function.

[0042] In each training iteration, the input portion of the training samples is fed into the network. The network outputs predicted deformation and predicted stability level during forward propagation. The deformation prediction loss and stability level prediction loss are calculated separately, and then multiplied by their respective weights and summed to obtain the total loss value. Based on this total loss value, the gradient is calculated through backpropagation, and the network parameters are updated. As the number of training iterations increases, the network parameters are gradually adjusted, and the total loss value gradually decreases. When the total loss value no longer decreases significantly in multiple consecutive training iterations, i.e., the objective function converges, training is stopped, and the trained dynamic prediction network for surrounding rock stability is obtained.

[0043] Compared to staged training, the weighted summation objective function ensures that both tasks are simultaneously optimized with each parameter update, sharing the network parameters from the underlying feature extraction part and avoiding feature overlap issues caused by sequential training. Compared to direct addition, introducing weights reconciles the magnitude differences between the two losses, achieving a reasonable balance between deformation prediction accuracy and stability classification accuracy during training. The network trained in this way can provide relatively reliable deformation predictions and stability classification results simultaneously when used online.

[0044] In a preferred embodiment, the tunnel construction quality control method includes the following steps in constructing the support decision reinforcement learning model: inputting the physical property parameters of the surrounding rock ahead of the tunnel face, the deformation data of the excavated section of the surrounding rock, and the stress data of the support structure from the historical data into the dynamic prediction network for surrounding rock stability to obtain the stability level and the predicted deformation value; obtaining the current support parameters recorded for each construction stage from the historical data; using the stability level, the predicted deformation value, and the current support parameters together as the state, using the support parameter adjustment scheme actually implemented in the historical data as the action, and constructing a reward function based on the deviation between the actual deformation amount corresponding to the action and the preset safe deformation threshold, to train the support decision reinforcement learning model to be trained.

[0045] This embodiment provides a method for constructing and training a reinforcement learning model for support decision-making. The construction process utilizes previously collected historical tunnel construction data, including physical properties of the surrounding rock ahead of the tunnel face, deformation data of the excavated section of the surrounding rock, and stress data of the support structure collected at different construction stages. First, the physical properties of the surrounding rock ahead of the tunnel face, the deformation data of the excavated section of the surrounding rock, and the stress data of the support structure from the historical data are input into a pre-trained dynamic prediction network for rock stability. This prediction network then outputs a set of stability levels and deformation prediction values ​​for the input data corresponding to each construction stage. In this way, each sample from the historical data for each construction stage corresponds to state information generated by the prediction network, namely, a stability level and a deformation prediction value.

[0046] Simultaneously, the current support parameters corresponding to each construction stage are obtained from historical data. These current support parameters refer to specific values ​​such as the actual steel arch spacing, shotcrete thickness, and anchor bolt arrangement parameters used in that construction stage, reflecting the existing support conditions. The stability level, deformation prediction value, and current support parameters are collectively used as the state of the reinforcement learning model. The stability level describes the current stability of the surrounding rock from a discrete level perspective; the deformation prediction value describes the expected deformation of the surrounding rock in the next construction stage from a continuous numerical perspective; and the current support parameters describe the specific situation of the applied support from an engineering measure perspective. Together, these three constitute a complete representation of the current state and recent trends of the surrounding rock and support system, enabling the model to simultaneously perceive surrounding rock risks and existing support conditions during decision-making, thereby outputting an adjustment plan adapted to the current support situation.

[0047] The reinforcement learning model uses actual support parameter adjustment schemes implemented in historical data as actions. Each construction stage in the historical construction records corresponds to actual support parameter adjustments, including one or more adjustments to steel arch spacing, shotcrete thickness, and anchor bolt arrangement parameters. These implemented adjustment schemes are used as the basis for defining the action space, ensuring that the model learns feasible adjustment operations in real-world construction scenarios during training.

[0048] A reward function is constructed based on the deviation between the actual deformation after an action is performed and a preset safe deformation threshold. Historical data records the actual deformation for each construction stage. When the model outputs a support parameter adjustment action during a historical construction stage, the actual deformation corresponding to that action refers to the measured surrounding rock deformation value after the action is performed in subsequent construction. The design principle of the reward function is to provide a positive reward when the deviation between the actual deformation and the safe deformation threshold is small and the actual deformation does not exceed the safe deformation threshold; and a negative reward is given based on the extent of the excess when the actual deformation exceeds the safe deformation threshold. This reward function guides the model to learn and select support parameter adjustment actions that keep the surrounding rock deformation within a safe range.

[0049] Specifically, the reward function is constructed based on the deviation between the actual deformation after the action and the preset safe deformation threshold corresponding to the current surrounding rock grade. The preset safe deformation threshold is set in segments according to the surrounding rock grade and tunnel depth, based on tunnel design specifications. Different surrounding rock grades correspond to different allowable upper limits for deformation. For example, for Grade I surrounding rock, the preset safe deformation threshold is 10mm; for Grade II surrounding rock, it is 20mm; for Grade III surrounding rock, it is 30mm; and so on. In practical applications, the corresponding threshold is called based on the currently output surrounding rock stability grade to participate in the reward function calculation.

[0050] The support decision reinforcement learning model to be trained is trained using the aforementioned defined state, action, and reward function. During training, the model takes sample states from historical data as input and outputs corresponding support parameter adjustment actions. The reward value is calculated based on the actual deformation and safe deformation threshold corresponding to each action, and the model parameters are then updated using the reinforcement learning algorithm. Through iterative training with a large number of historical samples, the model gradually learns the ability to select appropriate support parameter adjustment strategies to obtain the maximum cumulative reward under different surrounding rock conditions and existing support conditions, ultimately resulting in a fully trained support decision reinforcement learning model. In the online decision-making phase, this model can output support parameter adjustment schemes that control surrounding rock deformation within a safe range based on the current surrounding rock condition and current support parameters.

[0051] In a preferred embodiment, in the tunnel construction quality control method, the weight of the deformation prediction loss in the objective function is greater than the weight of the stability level prediction loss.

[0052] The objective function for multi-task learning typically involves directly summing the losses from each task, assuming equal weights for each loss. However, the deformation prediction loss and the stability level prediction loss often differ in magnitude. The deformation prediction loss is a regression loss, its value depending on the dimensions and range of the deformation itself, while the stability level prediction loss is a classification loss, its value depending on the degree of difference between the classification probability distribution and the label. When the magnitudes of the two losses differ significantly, direct summation leads to the larger loss term dominating the training process, causing network parameter updates to favor optimizing this loss term while suppressing the optimization of the other loss term. Ultimately, this results in one loss term performing reasonably well while the other performs significantly poorly.

[0053] In this implementation, when constructing the objective function, the weight of the deformation prediction loss is set to be greater than the weight of the stability level prediction loss. This weight allocation is based on the fact that, in the context of surrounding rock stability control during tunnel construction, the deformation prediction value provides direct quantitative input for subsequent support parameter decisions. The more accurate the deformation prediction value, the more reliable the calculation basis for support parameter adjustments. Although the stability level also provides a macroscopic judgment of the surrounding rock condition for decision-making, the continuous numerical information contained in the deformation prediction value provides more direct support for refined decision-making. Therefore, appropriately increasing the penalty weight for the deformation prediction loss during training helps guide the network to focus more on improving the deformation prediction accuracy during the learning process.

[0054] In practice, the objective function is constructed by multiplying the deformation prediction loss by its weights and adding the stability level prediction loss by its weights. During training, after calculating the two loss values ​​in each iteration, they are weighted and summed according to preset weights. Then, gradient backpropagation and parameter updates are performed based on the weighted sum. Because the weight of the deformation prediction loss is greater than that of the stability level prediction loss, the network responds more strongly to the deformation prediction error during parameter updates, causing the network parameters to be updated in a direction that reduces the deformation prediction error. At the same time, the stability level prediction loss still participates in training, ensuring that the network does not completely ignore classification performance. The prediction network trained in this way has higher deformation prediction accuracy while maintaining basic stability level discrimination ability.

[0055] In a preferred embodiment, in the tunnel construction quality control method, the weight of the deformation prediction loss is 0.6 to 0.7, the weight of the stability level prediction loss is 0.3 to 0.4, and the sum of the weights of the deformation prediction loss and the stability level prediction loss is 1.

[0056] In this embodiment, the deformation prediction loss weight is between 0.6 and 0.7, meaning that the deformation prediction loss accounts for 60% to 70% of the total loss, while the stability level prediction loss accounts for 30% to 40%. This proportional allocation allows the deformation prediction task to play a dominant role in parameter updates, while the stability level prediction task retains sufficient participation, ensuring that the network does not lose its classification and discrimination capabilities during training.

[0057] In the specific training process, after calculating the deformation prediction loss and the stability level prediction loss in each iteration, the deformation prediction loss is multiplied by a preset weight value, such as 0.65, and the stability level prediction loss is multiplied by a preset weight value, such as 0.35. The two are then added together to obtain the total loss value, which is then used for backpropagation and parameter updates. When the training converges, the obtained prediction network achieves a balance between deformation prediction accuracy and stability level classification accuracy.

[0058] In a preferred embodiment, in the tunnel construction quality control method, the construction equipment control system includes a central controller and multiple equipment controllers communicatively connected to the central controller. The central controller receives the support parameter adjustment scheme and distributes the steel arch spacing adjustment, shotcrete thickness adjustment, and anchor bolt arrangement parameter adjustment in the support parameter adjustment scheme to the corresponding equipment controllers.

[0059] This embodiment adopts a layered architecture for the construction equipment control system. The control system includes a central controller and multiple device controllers communicatively connected to it. The central controller is deployed at the tunnel construction site control center, responsible for receiving support parameter adjustment plans from the upper-level decision-making system, and parsing and distributing the plans. Each device controller is installed on its corresponding construction equipment or in a control cabinet, and each device controller is responsible for controlling one or a group of construction devices. The device controllers and the central controller establish a communication connection through an industrial communication network, which can be built using fieldbus or industrial Ethernet, among other methods.

[0060] After receiving the support parameter adjustment plan, the central controller parses the adjustment amounts for the steel arch spacing, shotcrete thickness, and anchor bolt arrangement parameters. The central controller maintains a mapping table between each equipment controller and the type of equipment it controls. Based on this mapping, it sends the steel arch spacing adjustment amount to the equipment controller corresponding to the arch installation equipment, the shotcrete thickness adjustment amount to the equipment controller corresponding to the shotcrete equipment, and the anchor bolt arrangement parameter adjustment amount to the equipment controller corresponding to the anchor bolt construction equipment. This distribution process is completed automatically by the central controller, requiring no manual intervention. Upon receiving the adjustment amounts, each equipment controller writes them into its respective execution buffer as the preset values ​​for the next excavation cycle.

[0061] The centralized distribution method by the central controller reduces intermediate links in information transmission. The three adjustment parameters can be sent to each device controller in parallel, avoiding the time lag and information distortion risks caused by manual notification of each parameter. At the same time, the central controller manages the distribution process uniformly and has the ability to record distribution results, which facilitates subsequent tracking of the sending status of each adjustment command.

[0062] In a preferred embodiment, in the tunnel construction quality control method, the support parameter adjustment scheme includes the correspondence between the equipment identifier and the adjustment amount of the steel arch spacing, the adjustment amount of the shotcrete thickness, and the adjustment amount of the anchor bolt arrangement parameters. The central controller sends the adjustment amount of the steel arch spacing, the adjustment amount of the shotcrete thickness, and the adjustment amount of the anchor bolt arrangement parameters to the corresponding equipment controller according to the equipment identifier.

[0063] This implementation introduces a correspondence between equipment identifiers and adjustment quantities in the support parameter adjustment scheme. When generating the support parameter adjustment scheme, it includes not only the numerical values ​​of the three types of adjustment quantities—steel arch spacing adjustment, shotcrete thickness adjustment, and anchor bolt arrangement parameter adjustment—but also the equipment identifier corresponding to each adjustment quantity. The equipment identifier is a unique identification code pre-assigned to each construction device, which is registered with the central controller when the device controller connects to the system. The equipment identifier can include equipment type information and equipment number information. For example, the prefix of the identifier indicates whether the equipment is an arch installation device, a shotcrete device, or an anchor bolt construction device, and the suffix is ​​the equipment number, used to distinguish different devices of the same type.

[0064] After receiving the support parameter adjustment plan, the central controller parses the correspondence between the adjustment quantities and equipment identifiers in each plan. For the steel arch spacing adjustment, the central controller reads its corresponding equipment identifier, searches for the matching equipment controller in the registered equipment controller list, establishes a communication connection, and sends the steel arch spacing adjustment to that equipment controller. Similarly, based on the equipment identifier corresponding to the shotcrete thickness adjustment, the central controller sends the adjustment quantity to the corresponding concrete spraying equipment controller. Based on the equipment identifier corresponding to the anchor bolt arrangement parameter adjustment, the central controller sends the adjustment quantity to the corresponding anchor bolt construction equipment controller. Even if multiple pieces of the same type of equipment are operating simultaneously, as long as the identifiers of each piece of equipment are different, the central controller can accurately send the adjustment quantity to the target equipment.

[0065] In a preferred embodiment, the tunnel construction quality control method, after sending the steel arch spacing adjustment amount, the shotcrete thickness adjustment amount, and the anchor bolt arrangement parameter adjustment amount to the corresponding equipment controller, further includes: the central controller receiving the execution confirmation signal returned by each of the equipment controllers, the execution confirmation signal indicating that the construction equipment corresponding to the equipment controller has completed the parameter preset according to the received steel arch spacing adjustment amount, shotcrete thickness adjustment amount, or anchor bolt arrangement parameter adjustment amount.

[0066] In this implementation, after the central controller sends the steel arch spacing adjustment, shotcrete thickness adjustment, and anchor bolt arrangement parameter adjustment to each equipment controller, an execution confirmation signal receiving stage is added. Upon receiving the corresponding adjustment, each equipment controller writes the adjustment to the control unit of the construction equipment. After the equipment completes the preset parameter actions, the equipment controller generates an execution confirmation signal and returns this signal to the central controller via the communication network.

[0067] The execution confirmation signal contains information that the construction equipment corresponding to the equipment controller has completed the parameter preset according to the received adjustment amount. Specifically, when the arch frame installation equipment completes the preset adjustment amount of the steel arch frame spacing, the execution confirmation signal returned by its corresponding equipment controller indicates that the arch frame installation equipment has completed the parameter preset according to the received steel arch frame spacing adjustment amount. When the concrete spraying equipment completes the preset adjustment amount of the sprayed concrete thickness, the execution confirmation signal returned by its corresponding equipment controller indicates that the concrete spraying equipment has completed the parameter preset according to the received sprayed concrete thickness adjustment amount. When the anchor bolt construction equipment completes the preset adjustment amount of the anchor bolt arrangement parameters, the execution confirmation signal returned by its corresponding equipment controller indicates that the anchor bolt construction equipment has completed the parameter preset according to the received anchor bolt arrangement parameter adjustment amount. If an adjustment amount is empty or the plan does not include a parameter adjustment, the corresponding equipment controller does not need to return an execution confirmation signal for that adjustment amount.

[0068] After receiving execution confirmation signals from each device controller, the central controller summarizes the feedback status. If all required confirmation signals are received from the device controllers within the preset waiting time, the current adjustment plan is confirmed to have been fully executed, and the next cycle of construction can proceed as planned. If no confirmation signal is received from a device controller within the preset waiting time, the central controller can mark that device as unconfirmed, prompting construction management personnel to check the device to avoid starting construction before the parameters are ready.

[0069] In a preferred embodiment, in the tunnel construction quality control method, after receiving the execution confirmation signals returned by each of the equipment controllers, the central controller further includes: the central controller associating and storing the support parameter adjustment scheme, the execution confirmation signals, and the sequence number of the excavation cycle to which the support parameter adjustment scheme is applied in the construction log database.

[0070] In this implementation, after the central controller receives the execution confirmation signals from each device controller, a data association and storage step is added. The central controller associates the support parameter adjustment scheme executed in this round, the execution confirmation signals returned by each device controller, and the sequence number of the excavation cycle to which the support parameter adjustment scheme is applied, and stores the associated data in the construction log database.

[0071] The support parameter adjustment plan records the complete content of this round of decisions, including the specific values ​​of adjustments to the steel arch spacing, shotcrete thickness, and anchor bolt arrangement parameters. This data reflects the decision-making intent of this round of adjustments. The execution confirmation signal records whether each construction device has completed the parameter presets according to the plan, indicating whether the decision intent has been implemented at the device level. The excavation cycle number identifies the construction cycle applied in this round of adjustments, establishing a correspondence between the adjustment plan and specific construction locations and time nodes. These three data items are organized using the excavation cycle number as the primary key or association key, ensuring that the support adjustment plan content, implementation status, and construction location for any given excavation cycle form a complete data record.

[0072] The construction log database is a dedicated database for storing construction process data and can be integrated with the tunnel construction management information system. After each adjustment of support parameters, the central controller automatically writes the associated data into the database without manual intervention. The records in the database are arranged in sequence according to the excavation cycle number, forming a complete data chain of the construction process.

[0073] The associative storage method encapsulates the decision content, execution results, and construction location into a complete record stored in a unified database. This allows for subsequent queries of the complete process of a particular support adjustment, providing the entire scheme, execution status of each piece of equipment, and construction location information for that cycle simply by using the excavation cycle number. This historical data also provides a structured source of training samples for subsequent optimization of the prediction network and decision model. Specifically, physical property parameters, deformation data, stress data, actual deformation, and corresponding support parameter adjustment schemes for a specific excavation cycle can be extracted from the construction log database and added to the training dataset for iterative model updates.

[0074] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily implemented by those skilled in the art. Therefore, the present invention is not limited to the specific details without departing from the general concept defined by the claims and their equivalents.

Claims

1. A method for quality control in tunnel construction, characterized in that, Includes the following steps: To obtain the physical properties of the surrounding rock in front of the tunnel face during tunnel construction, as well as the real-time deformation data of the surrounding rock in the excavated section and the real-time stress data of the support structure; The physical property parameters, real-time deformation data, and real-time stress data are input into a pre-trained dynamic prediction network for surrounding rock stability. The dynamic prediction network integrates multi-source heterogeneous data through a long short-term memory network and outputs the current stability level of the surrounding rock and the deformation prediction value for the next excavation cycle. Obtain the support parameters of the current construction cycle, input the stability level, the predicted deformation value and the support parameters of the current construction cycle into a pre-built support decision reinforcement learning model, and output the support parameter adjustment scheme for the next excavation cycle from the support decision reinforcement learning model. The support parameter adjustment scheme includes at least one of the following: steel arch spacing adjustment, shotcrete thickness adjustment and anchor bolt arrangement parameter adjustment. The support parameter adjustment plan is sent to the construction equipment control system via an industrial communication protocol so that the operation is carried out in accordance with the support parameter adjustment plan in the next excavation cycle.

2. The tunnel construction quality control method according to claim 1, characterized in that, The training process of the dynamic prediction network for surrounding rock stability includes: The tunnel construction history data includes physical property parameters of the surrounding rock in front of the tunnel face collected at different construction stages, deformation data of the surrounding rock in the excavated section, stress data of the support structure, current support parameters used for each construction stage, and actual deformation recorded for each construction stage. Using the physical properties of the surrounding rock in front of the tunnel face, the deformation data of the excavated section of the surrounding rock, and the stress data of the support structure from the historical data as input samples, the actual deformation is mapped to the corresponding stability level label. The actual deformation and the stability level label are used together as training labels to train the long short-term memory network to obtain the dynamic prediction network for surrounding rock stability.

3. The tunnel construction quality control method according to claim 2, characterized in that, The training of the constructed Long Short-Term Memory network includes: The parameters of the Long Short-Term Memory network are optimized using the weighted sum of the deformation prediction loss and the stability level prediction loss as the objective function until the objective function converges.

4. The tunnel construction quality control method according to claim 2, characterized in that, The construction process of the support decision reinforcement learning model includes: The physical properties of the surrounding rock in front of the tunnel face, the deformation data of the excavated section of the surrounding rock, and the stress data of the support structure from the historical data are input into the dynamic prediction network for surrounding rock stability to obtain the stability level and the predicted deformation value. The current support parameters corresponding to each construction stage are obtained from the historical data. The stability level, the predicted deformation value, and the current support parameters are used together as the state, the support parameter adjustment scheme actually implemented in the historical data is used as the action, and the deviation between the actual deformation amount after the action and the preset safe deformation threshold is used to construct a reward function to train the support decision reinforcement learning model to be trained.

5. The tunnel construction quality control method according to claim 3, characterized in that, In the objective function, the weight of the deformation prediction loss is greater than the weight of the stability level prediction loss.

6. The tunnel construction quality control method according to claim 5, characterized in that, The weight of the deformation prediction loss is 0.6 to 0.7, the weight of the stability level prediction loss is 0.3 to 0.4, and the sum of the weights of the deformation prediction loss and the stability level prediction loss is 1.

7. The tunnel construction quality control method according to claim 1, characterized in that, The construction equipment control system includes a central controller and multiple device controllers that are communicatively connected to the central controller. The central controller receives the support parameter adjustment scheme and distributes the adjustment amounts of the steel arch spacing, the shotcrete thickness, and the anchor bolt arrangement parameters in the support parameter adjustment scheme to the corresponding device controllers.

8. The tunnel construction quality control method according to claim 7, characterized in that, The support parameter adjustment scheme includes the correspondence between the equipment identifier and the adjustment amount of the steel arch spacing, the adjustment amount of the shotcrete thickness, and the adjustment amount of the anchor bolt arrangement parameters. The central controller sends the adjustment amount of the steel arch spacing, the adjustment amount of the shotcrete thickness, and the adjustment amount of the anchor bolt arrangement parameters to the corresponding equipment controller according to the equipment identifier.

9. The tunnel construction quality control method according to claim 8, characterized in that, After sending the steel arch spacing adjustment amount, the shotcrete thickness adjustment amount, and the anchor bolt arrangement parameter adjustment amount to the corresponding equipment controller, the method further includes: the central controller receiving the execution confirmation signal returned by each of the equipment controllers, wherein the execution confirmation signal indicates that the construction equipment corresponding to the equipment controller has completed the parameter preset according to the received steel arch spacing adjustment amount, shotcrete thickness adjustment amount, or anchor bolt arrangement parameter adjustment amount.

10. The tunnel construction quality control method according to claim 9, characterized in that, After receiving the execution confirmation signals returned by each of the device controllers, the central controller further includes: the central controller associating and storing the support parameter adjustment scheme, the execution confirmation signals, and the sequence number of the excavation cycle to which the support parameter adjustment scheme is applied in the construction log database.