A shield tunnel gas diffusion prediction method based on digital twinning

Abstract

The present application belongs to the field of underground engineering safety monitoring and early warning, and specifically relates to a shield tunnel gas diffusion prediction method based on digital twinning, which comprises the following steps: constructing a shield tunnel gas diffusion prediction model based on a digital twinning algorithm, performing deep optimization on the constructed shield tunnel gas diffusion prediction model through a multi-layer data processing mechanism, and performing shield tunnel gas diffusion prediction based on the optimized shield tunnel gas diffusion prediction model; and the present application achieves the effect of dynamic bidirectional fusion between a physical simulation model and real-time field sensing data by establishing a mapping relationship between physical monitoring points and virtual grid nodes and injecting real-time field sensor measured data into corresponding nodes of a virtual model through the mapping relationship.

Description

Technical Field

[0001] This invention belongs to the field of underground engineering safety monitoring and early warning, specifically a method for predicting gas diffusion in shield tunnels based on digital twins. Background Technology

[0002] When shield tunnels traverse gas-rich coal-bearing or organic strata, gas outbursts are frequent, and their diffusion behavior is influenced by multiple factors, including tunnel geometry, ventilation system layout, surrounding rock permeability distribution, and construction disturbance. Current technologies typically assess gas diffusion using concentration monitoring at fixed sensor locations combined with empirical formulas, or employ computational fluid dynamics (CFD) numerical simulations to perform offline simulations of gas diffusion fields under specific conditions. Some studies have also attempted to introduce machine learning models to statistically fit historical monitoring data to achieve short-term concentration prediction. These methods have all been applied to some extent in engineering practice and have provided fundamental support for tunnel gas safety management.

[0003] However, existing technologies are mainly limited by the lack of a dynamic two-way fusion mechanism between physical simulation models and real-time field perception data when realizing dynamic prediction of gas diffusion across the entire shield tunnel. This makes it difficult to balance the high timeliness of diffusion prediction results with the high fidelity of prediction models under complex and variable working conditions in practical applications, thus affecting the overall response accuracy and engineering practicality of gas disaster early warning systems. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for predicting gas diffusion in shield tunnels based on digital twins. This method solves the problems in existing technologies, such as dynamic bidirectional fusion between physical simulation models and real-time field perception data, high timeliness of prediction results, and high fidelity of prediction models under complex and variable working conditions.

[0005] A method for predicting gas diffusion in shield tunnels based on digital twins includes the following steps: constructing a gas diffusion prediction model for shield tunnels based on a digital twin algorithm;

[0006] The constructed shield tunnel gas diffusion prediction model is deeply optimized through a multi-layer data processing mechanism, specifically as follows:

[0007] Multi-source sensing data is time-aligned using a unified time reference. Outlier detection is performed on the aligned sensing data. Abnormal data points that exceed the physical reasonable range are marked and isolated to an abnormal buffer. Missing data is interpolated according to the time sequence of adjacent sensors in the same area.

[0008] The model is divided into multiple sub-model units. Independent optimization boundary conditions are set for each sub-model unit, and the internal parameter configuration of each sub-model unit is adjusted accordingly. After the partition optimization is completed, cross-regional physical continuity constraints are injected at the interface between adjacent sub-model units. The numerical solution of the global convection-diffusion equation is re-executed with the updated parameter set of each sub-model unit as input. The corresponding working condition label is bound to each sub-model unit.

[0009] The collected multi-source sensing data is divided into two categories according to the update frequency: high-frequency disturbance data and low-frequency trend data. The graded data is then input into the model through an asynchronous parallel mechanism. It is determined whether the high-frequency injection result and the low-frequency correction result cause state conflicts at the same model node. A physical constraint judgment function is constructed to uniformly revise the conflicting state. The unified state value is written into the corresponding grid node in the model through the established mapping relationship. The state synchronization deviation value of the node is calculated. The injection weight ratio of high-frequency and low-frequency data is dynamically adjusted according to the extracted state synchronization deviation value.

[0010] The optimized parameter set and the revised state set are fused according to the fusion weight. Consistency verification is performed on the fused model. After the consistency verification is completed, the fusion confidence score is calculated for each functional area of ​​the fused model. Targeted supplementary optimization is performed based on the calculation results.

[0011] The gas diffusion prediction for shield tunnels is performed based on the optimized shield tunnel gas diffusion prediction model.

[0012] Preferably, the method for constructing a gas diffusion prediction model for shield tunnels based on a digital twin algorithm is as follows:

[0013] Based on the actual construction drawings of the shield tunnel, establish a mapping relationship between each roadway segment, functional area division node, and sensor installation location in the physical entity of the tunnel and the corresponding grid node in the virtual model;

[0014] Based on the gas emission intensity parameters of each functional zone diffusion transport coefficient parameters and ventilation rate parameters Using the core parameters, the convection-diffusion equations are used for coupling, specifically:

[0015] ;

[0016] in, Indicates gas concentration. Indicates time, Represents the concentration gradient. Represents the concentration Laplace operator. This represents the wind speed field corresponding to the ventilation rate. Represents the diffusion transport coefficient. This indicates the source term of gas emission intensity.

[0017] Preferably, the outlier detection performed on the aligned sensing data is as follows:

[0018] Setting the first The original sampling time sequence of the sensor is The target sampling time sequence with a unified time reference is The aligned first digit is determined by linear interpolation. Sensor-like sensors at the target time Data values ​​at ;

[0019] After setting alignment, the first Sensor-like sensors at the target time The upper and lower bounds of the physical reasonable range of the data values ​​at that location are respectively and If abnormal data is isolated based on the defined upper and lower bounds of a reasonable physical range, then:

[0020] If the first Sensor-like sensors at the target time The aligned data values, when compared with the set upper and lower limits of the physical reasonable range, satisfy the formula. Then Data points marked as abnormal are isolated to an abnormal buffer and not included in subsequent channel inputs; otherwise, they are considered normal data.

[0021] Preferably, the injection of cross-regional physical continuity constraints at the interface between adjacent sub-model units is as follows:

[0022] Setting the first The and the first At the interface of each sub-model unit, the calculated gas concentration values ​​on both sides are as follows: and At the same time, a maximum concentration continuity deviation threshold is set at the interface. ;

[0023] The maximum concentration continuity deviation threshold at the interface is compared with the gas concentration on both sides of the interface. Based on the comparison results, it is determined whether the overall model maintains state consistency at the interval boundary. Specifically:

[0024] If the comparison results satisfy the formula If we perform a coordinated correction on the parameters on both sides of the interface, then we have:

[0025] Prioritize adjusting the diffusion transport coefficient parameters and The concentrations on both sides converge toward the mean by shrinking or expanding in proportion to the diffusion coefficient.

[0026] If adjusting the diffusion coefficient alone still fails to meet the constraints, then the ventilation rate parameter at the boundary should be adjusted jointly. and The coordination correction is performed iteratively, with recalculation after each iteration. When the continuity constraint is met or the number of iterations reaches the preset maximum number of iterations, Stop when, if reached If the constraints are still not met, the current optimal correction result is retained and the current interface is recorded as the interval to be manually reviewed.

[0027] Preferably, the collected on-site multi-source sensing data is divided into two categories according to the update frequency: high-frequency disturbance data and low-frequency trend data, as follows:

[0028] The shortest time required for the ventilation status to change during tunnel boring machine excavation. The reciprocal of the value is used as a reference to set the frequency threshold. ;

[0029] Based on the set frequency threshold, the collected multi-source sensing data is divided into two categories according to the update frequency: high-frequency disturbance data and low-frequency trend data. Then:

[0030] Setting the first The data update frequency of this type of sensor is , will the The data update frequency of the sensor, relative to the set frequency threshold. By comparing the data and classifying it according to the comparison results, we have:

[0031] If the comparison results satisfy the formula ,but Then the first Sensor-like data Data classified as high-frequency disturbances ;

[0032] If the comparison results satisfy the formula ,but Then the first Sensor-like data Classified as low-frequency trend data .

[0033] Preferably, the determination of whether the high-frequency injection result and the low-frequency correction result cause a state conflict at the same model node is as follows:

[0034] The physical constraints constructed based on the principle of mass conservation are as follows:

[0035] node The state value at the point must satisfy the condition that the difference between the inflow gas mass and the outflow gas mass is equal to the change in gas concentration at the corresponding node over time.

[0036] Simultaneously set the allowable deviation threshold for mass conservation. and physical constraint determination function If the physical constraint determination function satisfies the formula , then it represents a node State value at Satisfy the physical constraints of the construction;

[0037] Set Node The high-frequency injection result is The low-frequency correction result is and conflict judgment tolerance threshold The set conflict judgment tolerance threshold is compared with the high-frequency injection results and the low-frequency correction results. Based on the comparison results, it is determined whether to perform conflict arbitration. Specifically:

[0038] If the comparison results satisfy the formula This indicates that the high-frequency injection result and the low-frequency correction result cause a state conflict at the same model node.

[0039] Preferably, the physical constraint determination function is used to uniformly revise the conflict state, as follows:

[0040] The high-frequency injection results and low-frequency correction results are input into the constructed physical constraint determination function. The unified state value after arbitration is determined based on the function's output. Specifically:

[0041] If the output of the physical constraint determination function satisfies the formula The unified state value after arbitration will be revised to... ;

[0042] If the output of the physical constraint determination function satisfies the formula The unified state value after arbitration will be revised to... ;

[0043] If the output of the physical constraint determination function satisfies the formula The unified state value after arbitration will be revised to... .

[0044] Preferably, the consistency check performed on the fused model is as follows:

[0045] Set the set of reconstruction parameters for the structure optimization channel output as follows: The real-time state set of the nodes output by the state synchronization channel is The fusion weights are determined based on the mean of the synchronization deviations of all nodes in the sub-model unit. The fused weights determine the fused digits. The integrated parameter set of each sub-model unit ;

[0046] According to the node fusion parameter set The model predicts the output value. Concentration value measured by sensor Calculate the consistency deviation based on the relative error between them. ;

[0047] Set a reasonable deviation threshold The set reasonableness deviation threshold is compared with the calculated consistency deviation, and a local callback correction is performed based on the comparison result. Specifically:

[0048] If the comparison results satisfy the formula Then a local callback correction is triggered for the current node, specifically:

[0049] Setting the first The step size of the next callback iteration is Initial step size After each iteration, check Check if the condition is met; if it is, retain the current step size and continue iterating.

[0050] If the condition is not met, the step size will be halved. The callback is then re-executed to ensure that after each valid iteration... Monotonically decreasing, iterative execution until satisfied. Or the number of callback iterations reaches the preset limit. Until then, if it reaches If the consistency check condition is still not met, the prediction result of the current node will be marked as a low confidence output.

[0051] Preferably, the targeted supplementation and optimization based on the calculation results is as follows:

[0052] Computational structure optimization depth scoring and status synchronization data coverage score Based on the calculated structure optimization depth score and state synchronization data coverage score, a fusion confidence score is calculated. ;

[0053] Set fusion confidence threshold By comparing the set fusion confidence threshold with the calculated fusion confidence score, and making targeted supplementary optimizations based on the comparison results, we have:

[0054] If the comparison results satisfy the formula This will trigger targeted supplementary optimization, specifically:

[0055] in accordance with and Those with lower scores are identified as potential candidates for further optimization, and thus:

[0056] like Then the drive structure optimization channel is for the first Each sub-unit will re-execute partition reconstruction;

[0057] like To determine the cause of the missing covered node, we have:

[0058] If the missing coverage node is caused by a data transmission interruption, the driver state synchronization channel will inject supplementary data.

[0059] If the missing coverage node is due to objective limitations, the spatial interpolation estimate of the adjacent covered node is used as a substitute, and a data coverage insufficiency warning is generated.

[0060] In addition, after each supplementary optimization is completed, channel fusion embedding, consistency verification, and fusion confidence assessment are re-executed, with the supplementary optimization being performed at most once. Wheel, if reached The comparison results still satisfy the formula. If the current optimal fusion result is retained, the current interval is marked as an interval with insufficient confidence and an early warning is generated.

[0061] Preferably, the step of performing shield tunnel gas diffusion prediction based on the optimized shield tunnel gas diffusion prediction model is as follows:

[0062] Before the prediction is executed, the optimized prediction model is verified using reserved historical monitoring data. Based on the verified shield tunnel gas diffusion prediction model, the current real-time sensing data is used as the driving input to perform the full-area gas diffusion prediction of the shield tunnel, output the spatiotemporal distribution prediction results of gas concentration within the future time window, and issue gas disaster early warning based on the prediction results.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] 1. This invention achieves dynamic bidirectional fusion between the physical simulation model and real-time field perception data by establishing a mapping relationship between physical monitoring points and virtual grid nodes and injecting the measured data of field sensors into the corresponding nodes of the virtual model in real time through the mapping relationship. This solves the core problem of the lack of a dynamic bidirectional fusion mechanism between the physical simulation model and real-time perception data in the prior art, enabling the digital twin model to continuously track the evolution of the actual working conditions on site and improve the real-time response capability of the prediction model.

[0065] 2. This invention achieves the effect of shortening the overall optimization convergence cycle and improving the timeliness of diffusion prediction results without sacrificing the model optimization depth by adopting a technical means of splitting the optimization process into a dual-channel parallel advancement of structural optimization channel and state synchronization channel, and an asynchronous injection mechanism that injects high-frequency disturbance data into the boundary input layer in a short period and low-frequency trend data into the global parameter layer in a long period. This solves the problem of excessive prediction response delay caused by serial optimization path in the prior art and meets the prediction timeliness requirements of gas disaster early warning system.

[0066] 3. This invention achieves high fidelity of the prediction model under complex and variable working conditions by adopting technical means such as partitioning and decoupling the prediction model into three functional zones: gas outburst zone, diffusion and transmission zone, and ventilation and dilution zone; setting independent optimization boundary conditions according to the differences in physical mechanisms of each zone; and combining working condition tag binding to realize the rapid reuse of historical optimization models. This overcomes the shortcomings of the existing technology, such as insufficient generalization ability of a single global model under variable working conditions and a sharp decline in prediction accuracy as the working conditions change.

[0067] 4. This invention achieves an adaptive dynamic balance between physical simulation results and real-time perceived state during the fusion stage by adopting a two-way collaborative iterative mechanism that pushes the latest optimized parameters to the state synchronization channel in real time through the structural optimization channel, and pushes the arbitration results back to the structural optimization channel to update the boundary conditions through the state synchronization channel, as well as a technical means of dynamically calculating adaptive fusion weights based on node synchronization deviation and consistency verification deviation. This avoids the problem that the fixed weight fusion method in the prior art cannot adapt to the differences in data credibility under different nodes and different working conditions, and further improves the overall response accuracy and engineering practicality of the prediction model.

[0068] 5. This invention achieves the effect of a digital twin closed-loop operation mechanism of prediction-control-feedback-update by taking the prediction results into action when they exceed the time limit to generate ventilation adjustment commands and tunneling speed adjustment suggestions, and feeding back the control execution results to the digital twin virtual model through a mapping relationship to trigger a new round of optimization. This enables the system to not only have the ability to predict and warn of gas diffusion, but also to actively drive ventilation control and construction scheduling decisions, thereby improving the overall response accuracy and engineering practicality of the gas disaster early warning system. Attached Figure Description

[0069] Figure 1 This is a schematic diagram of the overall method steps of the shield tunnel gas diffusion prediction method based on digital twins of the present invention. Detailed Implementation

[0070] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0071] Example 1

[0072] Reference Figure 1 As an embodiment of the present invention, a method for predicting gas diffusion in shield tunnels based on digital twins is provided, comprising the following steps:

[0073] S1: Construct a gas diffusion prediction model for shield tunnels based on digital twin algorithm.

[0074] Specifically, the shield tunnel gas diffusion prediction model, based on the digital twin algorithm, is constructed using the shield tunnel's geometry, geological parameters, and ventilation system layout as inputs. It employs the convection-diffusion equation based on the mass conservation principle as the modeling framework. The model consists of two parts: a physical entity mapping layer and a virtual simulation calculation layer. The physical entity mapping layer establishes the correspondence between physical monitoring points and virtual grid nodes, driving the virtual model state to maintain dynamic synchronization with the actual on-site entities. The virtual simulation calculation layer performs numerical solutions on discrete grid nodes based on the convection-diffusion equation, outputting the calculated gas concentration results for each node. This includes establishing entity mapping relationships and initializing the virtual model, as detailed below:

[0075] Establishing entity mapping relationships involves mapping the tunnel sections, functional area division nodes, and sensor installation locations in the physical tunnel entity to the corresponding mesh nodes in the virtual model, based on the actual construction drawings of the shield tunnel. This results in:

[0076] ;

[0077] in, Indicates the first Spatial coordinates of each physical monitoring point This represents the coordinates of the corresponding mesh node in the virtual model. Indicates the first The first physical monitoring point and the virtual model The mapping relationship between each grid node is used to determine the basis for state synchronization between the physical entity and the virtual model. Measured data collected by field sensors is processed through this mapping relationship. Real-time injection of corresponding virtual nodes drives the virtual model state to maintain dynamic consistency with the physical entity.

[0078] The virtual model is initialized using the gas emission intensity parameters of each functional zone. diffusion transport coefficient parameters and ventilation rate parameters Using the core parameters, the convection-diffusion equations are used for coupling, specifically:

[0079] ;

[0080] in, Indicates gas concentration. Indicates time, Represents the concentration gradient. Represents the concentration Laplace operator. This represents the wind speed field corresponding to the ventilation rate. Represents the diffusion transport coefficient. Indicates the source term of gas emission intensity;

[0081] Based on the constructed convection-diffusion equations, spatial discretization is performed on discrete grid nodes of each functional section of the shield tunnel using the finite volume method. The time term is time-progressed using an implicit Euler scheme, and the convection term is performed using a first-order upwind scheme to ensure numerical stability. Thus, the solution is obtained node-by-node on each grid node, forming a predictive underlying model with the ability to express physical mechanisms, and outputting a set of parameters. Meanwhile, after normalizing each parameter in the parameter set, we have:

[0082] ;

[0083] in, , , , , , These are the minimum and maximum values ​​of each parameter in the historical operating data. After normalization, each parameter is mapped to... The interval is defined, and a parameter set is constructed based on the normalized data. , where represents the normalized parameter set, providing the data foundation for subsequent processing steps.

[0084] S2: The constructed shield tunnel gas diffusion prediction model is deeply optimized using a multi-layer optimization mechanism.

[0085] Specifically, the deep optimization of the constructed shield tunnel gas diffusion prediction model is carried out using a multi-layer optimization mechanism. This involves a multi-layer data processing mechanism, including: sensing data preprocessing, dual-channel parallel optimization, and dual-channel fusion. The specific implementation is as follows:

[0086] Furthermore, the preprocessing of the sensing data addresses the timestamp misalignment issue caused by inconsistent sampling frequencies among different types of sensors. It aligns the multi-source sensing data using a unified time base, eliminating spurious errors introduced by time misalignment. The preprocessing is achieved through anomaly isolation and missing data completion, as detailed below:

[0087] Timing alignment of multi-source sensing data using a unified time base is performed as follows:

[0088] Setting the first The original sampling time sequence of the sensor is The target sampling time sequence with a unified time reference is After alignment Sensor-like sensors at the target time Data values ​​at Determined by linear interpolation, we have:

[0089] ;

[0090] in, , They represent the first The time immediately adjacent to the target in the original sampling sequence of the sensor-like device The two sampling times before and after, Indicates the first after alignment Sensor-like sensors at the target time The data value at that location, Indicates the first Sensor-like sensors at the target time The data value at that location, Indicates the first Sensor-like sensors at the target time The data value at that location.

[0091] Abnormal data isolation involves performing outlier detection on the aligned perceptual data, marking and isolating outlier data points that exceed physically reasonable limits into an anomaly buffer. Specifically:

[0092] Regarding the first Sensor-like sensors at the target time Aligned data values The upper and lower bounds of its physical reasonable range are set as follows: and If abnormal data is isolated based on the defined upper and lower bounds of a reasonable physical range, then:

[0093] If the first Sensor-like sensors at the target time The aligned data values, when compared with the set upper and lower limits of the physical reasonable range, satisfy the formula. Then Data points marked as abnormal are isolated to an abnormal buffer and not included in subsequent channel inputs; otherwise, they are considered normal data.

[0094] Missing data completion addresses localized data loss caused by sensor failure. It involves interpolating missing data based on the temporal patterns of adjacent sensors in the same area, as detailed below:

[0095] Setting the first Sensor-like sensors at the target time Data is missing at this location; the set of adjacent sensors in the same area is... Based on the weighted average of data from adjacent sensors at the same time, we have:

[0096] ;

[0097] in, For the first The adjacent sensor pairs The interpolation weights of the sensor-like structures are determined by the spatial distance between the two sensors. The reciprocal is determined, that is spatial distance The implementation personnel shall conduct the measurements themselves based on the actual application scenario, and the measured data shall prevail. This represents the set of adjacent sensors in the same region where data is missing. This represents data that has been imputed for missing data.

[0098] It should be noted that after the perception data preprocessing is completed, the preprocessing results are used as the unified input data source for both the subsequent structure optimization channel and the state synchronization channel, and the boundary condition sets of each sub-model unit are refreshed synchronously. In addition, after the structure optimization channel completes the local optimization of each sub-model unit, it pushes the latest optimization parameters to the state synchronization channel in real time. The state synchronization channel refreshes the current synchronization benchmark based on the latest received structure parameters to ensure that the state injection calculation is performed based on the latest structure state. After completing each round of conflict arbitration and obtaining the unified state of the nodes, the state synchronization channel pushes the arbitration result back to the structure optimization channel, triggering the local correction and update of the boundary conditions of the corresponding sub-model unit. This allows the two channels to maintain data collaboration while advancing independently, forming a two-way collaborative iteration mechanism.

[0099] Furthermore, the dual-channel parallel optimization involves establishing a dual-channel parallel optimization mechanism for the constructed shield tunnel gas diffusion prediction model. The optimization process is divided into a structural optimization channel and a state synchronization channel, both operating independently within their respective channels. These channels interact periodically through a bidirectional collaborative mechanism defined by the sensing data preprocessing layer. Upon completion of each channel, they collectively enter the dual-channel fusion stage. The specific implementation is as follows:

[0100] The structural optimization channel targets the internal hierarchical structure of the constructed shield tunnel gas diffusion prediction model. It sequentially performs hierarchical decomposition, partition reconstruction, cross-region constraint injection, global model re-solution, and working condition label binding. The specific implementation is as follows:

[0101] Hierarchical decomposition involves dividing the shield tunnel gas diffusion prediction model into three functional zones: the gas emission zone, the diffusion and transmission zone, and the ventilation and dilution zone, and decoupling them to form independent sub-model units. Specifically:

[0102] The global gas diffusion prediction model for shield tunnels is divided into three parts. The first sub-model unit, Each sub-model unit is denoted as ( The corresponding normalized parameter subset is denoted as:

[0103] ;

[0104] in, , , They represent the first Each sub-model unit corresponds to a normalized gas emission intensity parameter, a normalized diffusion transport coefficient parameter, and a normalized ventilation rate parameter. Furthermore, the parameter subsets of each sub-model unit are independent of each other and do not overlap. This represents a subset of normalized parameters.

[0105] Partition reconstruction involves setting independent optimization boundary conditions for each sub-model unit and making targeted adjustments to the internal parameter configurations of each sub-unit based on the differences in the physical mechanisms of each functional area. Specifically:

[0106] Based on the classification of physical mechanisms for each functional zone, the set of boundary conditions for each sub-model unit is established, then:

[0107] For the gas outburst zone, the set of boundary conditions The main constraint is the boundary constraint of gas release, that is, the constraint normalized gas emission intensity parameter. The feasible region is the historical outflow statistics interval;

[0108] For the diffusion transport region, the set of boundary conditions The main constraint is the concentration continuity boundary, that is, the concentration difference between adjacent grid nodes is constrained not to exceed the preset continuity tolerance.

[0109] This is the set of boundary conditions for the ventilation dilution zone. The primary constraint is the ventilation pressure boundary, i.e., the constraint is on the normalized ventilation rate parameter. The feasible domain is the current rated operating range of the ventilation system.

[0110] By making targeted adjustments to the internal parameter configurations of each subunit, we can obtain:

[0111] Setting the first The normalized parameter subset of each sub-model unit before partition reconstruction is: Based on the concentration prediction error term With boundary continuity error term The weighted local optimization objective function is constructed, and the reconstructed parameter subset is determined based on the constructed local objective function and the normalized parameter subset before partition reconstruction. Specifically:

[0112] ;

[0113] ;

[0114] in, This represents the concentration prediction error term, which is the term of the first element. The mean square error between the model-predicted concentration and the sensor-measured concentration at each grid node within a sub-model unit. The boundary continuity error term is the term of the first generation. The sum of squares of the concentration differences at the interface between each sub-model unit and its adjacent sub-units , The distribution represents the corresponding weighting coefficients, which are set by the implementers according to the importance of each error term. This represents the constructed local optimization objective function. This represents a subset of normalized parameters before partition reconstruction. This represents the reconstructed subset of parameters.

[0115] Cross-region constraint injection involves injecting cross-region physical continuity constraints at the interfaces between adjacent sub-model units after partition optimization. This ensures that the overall model, after reassembling all sub-units, maintains state consistency at the interval boundaries. Specifically:

[0116] Setting the first The and the first At the interface of each sub-model unit, the calculated gas concentration values ​​on both sides are as follows: and At the same time, a maximum concentration continuity deviation threshold is set at the interface. ;

[0117] The maximum concentration continuity deviation threshold at the interface is compared with the gas concentration on both sides of the interface. Based on the comparison results, it is determined whether the overall model maintains state consistency at the interval boundary. Specifically:

[0118] If the comparison results satisfy the formula If we perform a coordinated correction on the parameters on both sides of the interface, then we have:

[0119] Prioritize adjusting the diffusion transport coefficient parameters and The concentrations on both sides converge toward the mean by shrinking or expanding in proportion to the diffusion coefficient.

[0120] If adjusting the diffusion coefficient alone still fails to meet the constraints, then further joint adjustments should be made to the ventilation rate parameter at the boundary. and The reconciliation is performed iteratively, with recalculation after each iteration. When the continuity constraint is met or the number of iterations reaches the preset maximum number of iterations, Stop when, if reached If the constraints are still not met, the current optimal correction result is retained and the interface is recorded as the interval to be manually reviewed.

[0121] The global model re-solution is performed after all sub-model units have completed partition reconstruction and cross-region constraint injection, using the updated parameter set of each sub-model unit. Using this as input, the numerical solution of the global convection-diffusion equation is re-executed to obtain the updated global gas concentration field, specifically:

[0122] ;

[0123] in, This represents the global numerical solution operator based on the finite volume method and the implicit Euler scheme; after the global re-solution is completed, This represents the updated set of parameters for each input sub-model unit. This represents the set of input boundary conditions. This represents the updated global gas concentration field, which is the final output of the current structural optimization channel, and includes the updated parameter set for each sub-model unit. The status is pushed to the status synchronization channel in real time so that it can refresh the synchronization benchmark.

[0124] Working condition label binding involves assigning a corresponding working condition label to each sub-model unit after partition reconstruction and global re-solution verification. This labels record the tunneling speed range, ventilation mode, and surrounding rock permeability range suitable for the current optimization results. This provides an index for rapid model retrieval during subsequent working condition switching, enhancing the reusability of structural optimization channel output results under varying working conditions. Specifically:

[0125] Setting the first The operating condition label of each sub-model unit is denoted as Then we have:

[0126] ;

[0127] in, This indicates the tunneling speed range that the current optimization result is suitable for. Indicates ventilation mode. Indicates the range of permeability of the surrounding rock. Indicates the first The operating condition label for each sub-model unit is indicated by the sub-model unit number. The index is stored in the working condition index library;

[0128] Furthermore, during subsequent operating condition switching, the current measured operating condition parameter vector is used as the basis. Labels for each sub-model unit's operating condition We use weighted Euclidean distance for matching, calculate the matching distance, and determine the nearest cell based on the calculated matching distance. Then we have:

[0129] ;

[0130] in, and The first The center value of the tunneling speed range and the surrounding rock permeability range in each label. and Indicates the width of the corresponding interval. This is an indicator function; it takes the value 1 if the condition is met, and 0 if the condition is not met. , , These are the weighting coefficients for each operating condition component. , , These represent the tunneling speed range, ventilation mode indicator, and surrounding rock permeability range in the current measured working condition parameter vector, respectively. This represents the calculated matching distance, used to determine the nearest cell, specifically:

[0131] Set the failure threshold for operating condition matching When the minimum matching distance satisfies When the current working condition is determined to be beyond the applicable scope of all historical models, the historical models are no longer retrieved, and the structural optimization channel is re-triggered to perform a complete partition reconstruction for the current working condition, so as to ensure that model reuse is always based on the similarity of working conditions.

[0132] When there are multiple matching distances that are the same and none exceed When selecting sub-model units, prioritize the unit whose working condition label binding time is closest to the current time.

[0133] The state synchronization channel is responsible for mapping the real-time sensing data and the twin model. It sequentially performs data hierarchical processing, asynchronous injection, conflict arbitration, arbitration result writing, and synchronization deviation feedback. The specific implementation is as follows:

[0134] Data classification involves dividing the collected multi-source sensing data from the field into two categories based on their update frequency: high-frequency disturbance data and low-frequency trend data. These categories are then entered into different synchronization queues. Specifically:

[0135] The shortest time required for the ventilation status to change during tunnel boring machine excavation. The reciprocal of the value is used as a reference to set the frequency threshold. ;

[0136] Based on the set frequency threshold, the collected multi-source sensing data is divided into two categories according to the update frequency: high-frequency disturbance data and low-frequency trend data. Then:

[0137] Setting the first The data update frequency of this type of sensor is , will the The data update frequency of the sensor, relative to the set frequency threshold. By comparing the data and classifying it according to the comparison results, we have:

[0138] If the comparison results satisfy the formula ,but Then the first Sensor-like data Data classified as high-frequency disturbances ;

[0139] If the comparison results satisfy the formula ,but Then the first Sensor-like data Classified as low-frequency trend data .

[0140] It should be noted that, and These are separate high-frequency disturbance data synchronization queues and low-frequency trend data synchronization queues. Different types of sensors update independently based on their respective update frequencies and frequency thresholds. Comparisons will be made, but a uniform classification result will not be mandatory.

[0141] At the same time, set periodic reclassification intervals. Every The equivalent frequency is recalculated based on the actual sampling change rate of each sensor within the period. and with Alternative Re-execute the classification judgment and remove sensors with increased rate of change from... Migrate to The rate of change remains below 0.5% for three consecutive periods. The sensor from Migrate to This enables data classification to have dynamic adjustment capabilities.

[0142] Asynchronous injection refers to inputting the graded data into the shield tunnel gas diffusion prediction model through an asynchronous parallel mechanism. This includes: continuously injecting high-frequency disturbance data into the boundary input layer of the prediction model in a short-period manner to drive the model to respond quickly to the current operating conditions; and injecting low-frequency trend data into the global parameter layer of the model in a long-period manner to drive the model to slowly correct the overall evolution trend. The specific implementation is as follows:

[0143] Set the high-frequency injection cycle as The low-frequency injection cycle is And both satisfy the formula ;

[0144] The nearest neighbor interpolation method is used to interpolate high-frequency perturbation data. If the latest sampled values ​​of each sensor are directly mapped to the boundary input layer of the corresponding grid node, then:

[0145] ;

[0146] in, Represents a high-frequency mapping function. This indicates high-frequency disturbance data. Indicates high-frequency disturbance data The sensor corresponds to the boundary input layer of the grid node, and is injected at intervals of one high-frequency injection cycle. Update once;

[0147] Using a sliding window mean filter for low-frequency trend data The parameters in the recent The average of the data from each low-frequency sampling time is mapped to the global parameter layer, and the latest parameter set is pushed by the structure optimization channel during asynchronous injection. Refresh the current synchronization baseline to ensure that the injection calculation is based on the latest structural parameters, then:

[0148] ;

[0149] in, Represents the low-frequency mapping function. Indicates low-frequency trend data, Indicating low-frequency trend data The parameters in the recent The data from each low-frequency sampling time is averaged and mapped to the global parameter layer, with each high-frequency injection cycle having a certain interval. Update once, This indicates the length of the sliding window, which can be set by the implementer based on the actual application scenario.

[0150] Conflict arbitration is a mechanism initiated when high-frequency injection results and low-frequency correction results cause state conflicts at the same model node. It uses the principle of physical constraint priority to adjudicate and unify the state values ​​of conflicting nodes. Specifically:

[0151] The physical constraints constructed based on the principle of mass conservation are as follows:

[0152] node The state value at the point must satisfy the condition that the difference between the inflow gas mass and the outflow gas mass is equal to the change in gas concentration at the corresponding node over time.

[0153] Simultaneously set the allowable deviation threshold for mass conservation. and physical constraint determination function If the physical constraint determination function satisfies the formula , then it represents a node State value at Satisfy the physical constraints of the construction;

[0154] Set Node The high-frequency injection result is The low-frequency correction result is and conflict judgment tolerance threshold The set conflict judgment tolerance threshold is compared with the high-frequency injection results and the low-frequency correction results. Based on the comparison results, it is determined whether to perform conflict arbitration. Specifically:

[0155] If the comparison results satisfy the formula This indicates that conflict arbitration is being performed. The high-frequency injection result and the low-frequency correction result are input into the constructed physical constraint decision function. The unified state value after arbitration is determined based on the output of the function. Specifically:

[0156] If the output of the physical constraint determination function satisfies the formula The unified state value after arbitration will be revised to... ;

[0157] If the output of the physical constraint determination function satisfies the formula The unified state value after arbitration will be revised to... ;

[0158] If the output of the physical constraint determination function satisfies the formula The unified state value after arbitration will be revised to... .

[0159] The arbitration result is written after the conflict arbitration is completed, and the unified state value is entered into the system. Through the established mapping relationship The corresponding first step in the prediction model of gas diffusion in shield tunnels Each grid node, covering the current state of the current node, serves as the initial state for the next round of state synchronization and prediction calculations, specifically:

[0160] ;

[0161] in, This represents the unified state value after conflict arbitration. Indicates the relationship between the virtual model and the first The coordinates of the grid nodes corresponding to each physical monitoring point.

[0162] Furthermore, after the arbitration result is written, As a reverse push signal, it is transmitted to the structural optimization channel, triggering the local correction and update of the boundary conditions of the corresponding sub-model unit, thus completing this round of bidirectional collaborative data interaction.

[0163] Synchronization deviation feedback involves extracting the state synchronization deviation value of the shield tunnel gas diffusion prediction model node after each asynchronous injection, and transmitting the extracted deviation value back to the data grading stage of the state synchronization channel as a feedback signal. The injection weight ratio of high-frequency and low-frequency data is dynamically adjusted according to the deviation magnitude, enabling the state synchronization channel to have self-adjusting capabilities. This further enhances the dynamic bidirectional fusion effect between the physical simulation model and real-time field perception data. Specifically:

[0164] Extracting the state synchronization deviation values ​​of the nodes in the shield tunnel gas diffusion prediction model, we have:

[0165] ;

[0166] in, Represents a node The current calculated values ​​of the gas diffusion prediction model for shield tunnels. This represents the unified state value after conflict arbitration. This represents the extracted state synchronization deviation value.

[0167] Based on the extracted state synchronization deviation value, the injection weight ratio of high-frequency and low-frequency data is dynamically adjusted, resulting in:

[0168] ;

[0169] in, This represents the baseline constant for weight adjustment, and its value is determined by the average synchronization deviation of each node during historical stable operation. This indicates the adjusted injection weight ratio, when the ratio satisfies the formula... At the same time, the injection weights of high-frequency and low-frequency frequencies are kept balanced;

[0170] And adjust the injection weight ratio As a feedback signal, it is sent back to the data classification stage to adjust the high-frequency disturbance data in the next injection cycle. Low-frequency trend data The proportion of each entity's data contribution to that node.

[0171] Dual-channel fusion occurs after the structural optimization channel and the state synchronization channel have completed their respective optimizations, entering the fusion and convergence phase. This phase sequentially executes channel fusion embedding, consistency verification, fusion confidence assessment, and targeted supplementary optimization, as detailed below:

[0172] Channel fusion embedding uses the reconstructed model of the structure-optimized channel output as a skeleton, and embeds the real-time state mapping relationship of the state synchronization channel output into the corresponding hierarchical nodes of the skeleton, specifically:

[0173] Set the set of reconstruction parameters for the structure optimization channel output as follows: The real-time state set of the nodes output by the state synchronization channel is ;

[0174] By constructing state mapping embedding functions For those belonging to the first The node set of each sub-model unit Real-time status values ​​of each node According to the spatial position weight of the corresponding node within the sub-unit After weighted averaging, replacing the parameter values ​​at the corresponding grid positions in the reconstructed parameter set, we have:

[0175] ;

[0176] in, The spatial location weights are represented by the nodes. Distance from The distance to the center of each sub-model unit is determined by inverse normalization.

[0177] The fusion weights are determined based on the mean of the synchronization deviations of all nodes in the sub-model unit, then:

[0178] ;

[0179] in, Indicates the first Synchronization deviation of all nodes within a sub-model unit The mean, Indicates the first In the previous round of consistency verification of each sub-model unit, the deviation of each node was... The mean, To prevent small constants with a denominator of zero, This represents the determined fusion weight. The larger the value, the greater the synchronization deviation of the current state of the current sub-unit, and the greater the fusion weight of the real-time state mapping should be. The smaller the value, the greater the weight of the structural reconstruction parameters.

[0180] Based on the determined fusion weights, the fused i-th The combined parameter set of each sub-model unit is then:

[0181] ;

[0182] in, This indicates a defined fusion weight. Indicates belonging to the The set of nodes of each sub-model unit This represents the state mapping embedding function being constructed. Indicates the fused first The comprehensive parameter set of each sub-model unit.

[0183] Consistency verification involves performing a consistency check on the fused model and triggering local callback corrections based on the detection results, as detailed below:

[0184] According to the node fusion parameter set The model predicts the output value. Concentration value measured by sensor The relative error between them is used to calculate the consistency deviation, specifically:

[0185] ;

[0186] in, Nodes representing computation The consistency deviation at that point is used to perform consistency checks, specifically:

[0187] Set a reasonable deviation threshold The set reasonableness deviation threshold is compared with the calculated consistency deviation, and a local callback correction is performed based on the comparison result. Specifically:

[0188] If the comparison results satisfy the formula Then a local callback correction is triggered for the current node, specifically:

[0189] Using the Armijo step size criterion to determine the callback step size to ensure iteration convergence, we have:

[0190] Setting the first The step size of the next callback iteration is Initial step size After each iteration, check If the condition is met, retain the current step size and continue iterating; otherwise, halve the step size. The callback is then re-executed to ensure that after each valid iteration... Monotonically decreasing, iterative execution until satisfied. Or the number of callback iterations reaches the preset limit. Until then, if it reaches If the consistency verification condition is still not met, the prediction result of the current node will be marked as a low confidence output. In the subsequent early warning system, the low confidence mark will be used for reference by operation and maintenance personnel, without interrupting the continued execution of the overall prediction process.

[0191] Fusion confidence assessment and targeted supplementary optimization involve calculating the fusion confidence score for each functional region of the fusion model after consistency verification, and then performing targeted supplementary optimization based on the calculation results, as follows:

[0192] By calculating the structural optimization depth score and the state synchronization data coverage score, we have:

[0193] ;

[0194] Based on the calculated structural optimization depth score and status synchronization data coverage score Calculating the fusion confidence score, we have:

[0195] ;

[0196] in, This represents the calculated structural optimization depth score. This indicates that the calculated state synchronization data covers the score. This represents the weighting coefficient, which is set by the implementers based on the actual application scenario. The calculated fusion confidence score is used for targeted supplementary optimization, specifically:

[0197] Set fusion confidence threshold By comparing the set fusion confidence threshold with the calculated fusion confidence score, and making targeted supplementary optimizations based on the comparison results, we have:

[0198] If the comparison results satisfy the formula This will trigger targeted supplementary optimization, specifically:

[0199] in accordance with and Those with lower scores are identified as potential candidates for further optimization, and thus:

[0200] like Then the drive structure optimization channel is for the first Each sub-unit will re-execute partition reconstruction;

[0201] like To determine the cause of the missing covered node, we have:

[0202] If the missing coverage node is caused by a temporary sensor failure or data transmission interruption, the drive state synchronization channel will inject supplementary data.

[0203] If the missing coverage node is due to objective limitations such as the sensor not being physically deployed or hardware damage, then the spatial interpolation estimate of the adjacent covered node is used as a substitute, and a data coverage insufficiency warning is generated.

[0204] In addition, after each supplementary optimization is completed, channel fusion embedding, consistency verification, and fusion confidence assessment are re-executed, with the supplementary optimization being performed at most once. Wheel, if reached The comparison results still satisfy the formula. If the current optimal fusion result is retained, the current interval is marked as an interval with insufficient confidence and an early warning is generated. In subsequent prediction processes, the prediction results corresponding to the current interval are output with low confidence.

[0205] After all functional areas have completed confidence assessment, a shield tunnel gas diffusion prediction model with complete depth optimization is output, providing a data foundation for shield tunnel gas diffusion prediction.

[0206] S3: Perform gas diffusion prediction for shield tunnels based on the optimized shield tunnel gas diffusion prediction model.

[0207] Before executing the prediction, the optimized prediction model is validated using reserved historical monitoring data, as follows:

[0208] A validation dataset was constructed based on the optimized sensing dataset, and all historical monitoring data were arranged in chronological order of construction time. The last 30% is reserved for the optimization of the gas diffusion prediction model for shield tunnels, and the remaining 30% is reserved for the validation of the gas diffusion prediction model for shield tunnels. This ensures that the validation dataset and the optimization dataset do not overlap in time and avoids overfitting that could lead to artificially low validation metrics.

[0209] To construct three validation metrics—one for evaluating the overall prediction accuracy of the global gas concentration field, one for evaluating the relative error level of the time-series concentration prediction at each monitoring node, and one for evaluating the prediction accuracy of the peak location and magnitude of the gas concentration—we have:

[0210] ;

[0211] ;

[0212] ;

[0213] in, This represents the constructed validation dataset. Indicates the first The model predicts the output value of the group of samples. Indicates the first The historical measured values ​​corresponding to the model's predicted output values ​​for a set of samples. This indicates the overall prediction accuracy used to evaluate the global gas concentration field. This indicates the relative error level used to assess the time-series concentration predictions at each monitoring node. This indicates the accuracy of predictions used to assess the location and magnitude of peak gas concentrations.

[0214] The accuracy of the shield tunnel gas diffusion prediction model was verified based on three calculated validation indicators, specifically:

[0215] If the calculated three verification indicators satisfy the formula If the result is 0, it means that the accuracy of the gas diffusion prediction model for shield tunnels has been verified.

[0216] Based on a validated shield tunnel gas diffusion prediction model, and using real-time on-site sensing data as the driving input, a full-area gas diffusion prediction for the shield tunnel is performed, outputting the prediction results within the future time window. Prediction results of spatiotemporal distribution of gas concentration ;

[0217] And based on the forecast results, a gas disaster early warning will be issued, specifically as follows:

[0218] Set a safe threshold for gas concentration The set safe threshold for gas concentration is compared with the predicted gas concentration values ​​for each node in the prediction results. Based on the comparison results, a gas disaster prediction is made, which yields the following:

[0219] If the comparison results satisfy the formula It automatically triggers a gas early warning signal and simultaneously outputs the location of the exceeding node, the predicted time of exceeding the limit, and the predicted peak concentration to the gas disaster early warning system.

[0220] At the same time, based on the predicted location and concentration gradient of the over-limit area, ventilation adjustment instructions and shield tunneling speed adjustment suggestions are generated in conjunction. The ventilation adjustment instructions are based on the current value of the ventilation rate parameter of the functional area where the over-limit area is located, and the air volume output is increased according to the predicted over-limit level.

[0221] The tunneling speed of the shield is adjusted based on the predicted duration of the over-limit. When the predicted duration of the over-limit exceeds the preset tolerance time, the tunneling speed is reduced to reduce the intensity of the gas outburst.

[0222] After the early warning control command is executed, the actual execution results of ventilation adjustment and tunneling adjustment are fed back to the corresponding node of the digital twin virtual model in real time through the established mapping relationship, updating the model boundary conditions and parameter set, and triggering a new round of optimization process execution;

[0223] For intervals marked as low confidence during the optimization process, their corresponding prediction results are simultaneously displayed in the early warning system with a low confidence mark, prompting maintenance personnel to make a careful judgment based on the actual situation on site.

[0224] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of protection claimed by the present invention.

Claims

1. A method for predicting gas diffusion in shield tunnels based on digital twins, characterized in that: Includes the following steps: A prediction model for gas diffusion in shield tunnels was constructed based on the digital twin algorithm. The constructed shield tunnel gas diffusion prediction model is deeply optimized through a multi-layer data processing mechanism, specifically as follows: Multi-source sensing data is time-aligned using a unified time reference. Outlier detection is performed on the aligned sensing data. Abnormal data points that exceed the physical reasonable range are marked and isolated to an abnormal buffer. Missing data is interpolated according to the time sequence of adjacent sensors in the same area. The model is divided into multiple sub-model units. Independent optimization boundary conditions are set for each sub-model unit, and the internal parameter configuration of each sub-model unit is adjusted accordingly. After the partition optimization is completed, cross-regional physical continuity constraints are injected at the interface between adjacent sub-model units. The numerical solution of the global convection-diffusion equation is re-executed with the updated parameter set of each sub-model unit as input. The corresponding working condition label is bound to each sub-model unit. The collected multi-source sensing data is divided into two categories according to the update frequency: high-frequency disturbance data and low-frequency trend data. The graded data is then input into the model through an asynchronous parallel mechanism. It is determined whether the high-frequency injection result and the low-frequency correction result cause state conflicts at the same model node. A physical constraint judgment function is constructed to uniformly revise the conflicting state. The unified state value is written into the corresponding grid node in the model through the established mapping relationship. The state synchronization deviation value of the node is calculated. The injection weight ratio of high-frequency and low-frequency data is dynamically adjusted according to the extracted state synchronization deviation value. The optimized parameter set and the revised state set are fused according to the fusion weight. Consistency verification is performed on the fused model. After the consistency verification is completed, the fusion confidence score is calculated for each functional area of ​​the fused model. Targeted supplementary optimization is performed based on the calculation results. The gas diffusion prediction for shield tunnels is performed based on the optimized shield tunnel gas diffusion prediction model.

2. The method for predicting gas diffusion in shield tunnels based on digital twins as described in claim 1, characterized in that: The specific details of the shield tunnel gas diffusion prediction model constructed based on the digital twin algorithm are as follows: Based on the actual construction drawings of the shield tunnel, establish a mapping relationship between each roadway segment, functional area division node, and sensor installation location in the physical entity of the tunnel and the corresponding grid node in the virtual model; Based on the gas emission intensity parameters of each functional zone diffusion transport coefficient parameters and ventilation rate parameters Using the core parameters, the convection-diffusion equations are used for coupling, specifically: ； in, Indicates gas concentration. Indicates time, Represents the concentration gradient. Represents the concentration Laplace operator. This represents the wind speed field corresponding to the ventilation rate. Represents the diffusion transport coefficient. This indicates the source term of gas emission intensity.

3. The method for predicting gas diffusion in shield tunnels based on digital twins as described in claim 1, characterized in that: The outlier detection process for the aligned sensor data is as follows: Setting the first The original sampling time sequence of the sensor is The target sampling time sequence with a unified time reference is The aligned first digit is determined by linear interpolation. Sensor-like sensors at the target time Data values ​​at ; After setting alignment, the first Sensor-like sensors at the target time The upper and lower bounds of the physical reasonable range of the data values ​​at that location are respectively and If abnormal data is isolated based on the defined upper and lower bounds of a reasonable physical range, then: If the first Sensor-like sensors at the target time The aligned data values, when compared with the set upper and lower limits of the physical reasonable range, satisfy the formula. Then Data points marked as abnormal are isolated to an abnormal buffer and not included in subsequent channel inputs; otherwise, they are considered normal data.

4. The method for predicting gas diffusion in shield tunnels based on digital twins as described in claim 1, characterized in that: The injection of cross-regional physical continuity constraints at the interface of adjacent sub-model units is as follows: Setting the first The and the first At the interface of each sub-model unit, the calculated gas concentration values ​​on both sides are as follows: and At the same time, a maximum concentration continuity deviation threshold is set at the interface. ; The maximum concentration continuity deviation threshold at the interface is compared with the gas concentration on both sides of the interface. Based on the comparison results, it is determined whether the overall model maintains state consistency at the interval boundary. Specifically: If the comparison results satisfy the formula If we perform a coordinated correction on the parameters on both sides of the interface, then we have: Prioritize adjusting the diffusion transport coefficient parameters and The concentrations on both sides converge toward the mean by shrinking or expanding in proportion to the diffusion coefficient. If adjusting the diffusion coefficient alone still fails to meet the constraints, then the ventilation rate parameter at the boundary should be adjusted jointly. and The coordination correction is performed iteratively, with recalculation after each iteration. When the continuity constraint is met or the number of iterations reaches the preset maximum number of iterations, Stop when, if reached If the constraints are still not met, the current optimal correction result is retained and the current interface is recorded as the interval to be manually reviewed.

5. The method for predicting gas diffusion in shield tunnels based on digital twins as described in claim 1, characterized in that: The collected on-site multi-source sensing data is divided into two categories according to the update frequency: high-frequency disturbance data and low-frequency trend data, as detailed below: The shortest time required for the ventilation status to change during tunnel boring machine excavation. The reciprocal of the value is used as a reference to set the frequency threshold. ; Based on the set frequency threshold, the collected multi-source sensing data is divided into two categories according to the update frequency: high-frequency disturbance data and low-frequency trend data. Then: Setting the first The data update frequency of this type of sensor is , will the The data update frequency of the sensor, relative to the set frequency threshold. By comparing the data and classifying it according to the comparison results, we have: If the comparison results satisfy the formula ,but Then the first Sensor-like data Data classified as high-frequency disturbances ; If the comparison results satisfy the formula ,but Then the first Sensor-like data Classified as low-frequency trend data .

6. The method for predicting gas diffusion in shield tunnels based on digital twins as described in claim 1, characterized in that: The determination of whether the high-frequency injection result and the low-frequency correction result cause a state conflict at the same model node is as follows: The physical constraints constructed based on the principle of mass conservation are as follows: node The state value at the point must satisfy the condition that the difference between the inflow gas mass and the outflow gas mass is equal to the change in gas concentration at the corresponding node over time. Simultaneously set the allowable deviation threshold for mass conservation. and physical constraint determination function If the physical constraint determination function satisfies the formula , then it represents a node State value at Satisfy the physical constraints of the construction; Set Node The high-frequency injection result is The low-frequency correction result is and conflict judgment tolerance threshold The set conflict judgment tolerance threshold is compared with the high-frequency injection results and the low-frequency correction results. Based on the comparison results, it is determined whether to perform conflict arbitration. Specifically: If the comparison results satisfy the formula This indicates that the high-frequency injection result and the low-frequency correction result cause a state conflict at the same model node.

7. The method for predicting gas diffusion in shield tunnels based on digital twins as described in claim 1, characterized in that: The physical constraint determination function is constructed to uniformly revise conflict states, as follows: The high-frequency injection results and low-frequency correction results are input into the constructed physical constraint determination function. The unified state value after arbitration is determined based on the function's output. Specifically: If the output of the physical constraint determination function satisfies the formula The unified state value after arbitration will be revised to... ; If the output of the physical constraint determination function satisfies the formula The unified state value after arbitration will be revised to... ; If the output of the physical constraint determination function satisfies the formula The unified state value after arbitration will be revised to... .

8. The method for predicting gas diffusion in shield tunnels based on digital twins as described in claim 1, characterized in that: The consistency check performed on the fused model is as follows: Set the set of reconstruction parameters for the structure optimization channel output as follows: The real-time state set of the nodes output by the state synchronization channel is The fusion weights are determined based on the mean of the synchronization deviations of all nodes in the sub-model unit. The fused weights determine the fused digits. The integrated parameter set of each sub-model unit ; According to the node fusion parameter set The model predicts the output value. Concentration value measured by sensor Calculate the consistency deviation based on the relative error between them. ; Set a reasonable deviation threshold The set reasonableness deviation threshold is compared with the calculated consistency deviation, and a local callback correction is performed based on the comparison result. Specifically: If the comparison results satisfy the formula Then a local callback correction is triggered for the current node, specifically: Setting the first The step size of the next callback iteration is Initial step size After each iteration, check Check if the condition is met; if it is, retain the current step size and continue iterating. If the condition is not met, the step size will be halved. The callback is then re-executed to ensure that after each valid iteration... Monotonically decreasing, iterative execution until satisfied. Or the number of callback iterations reaches the preset limit. Until then, if it reaches If the consistency check condition is still not met, the prediction result of the current node will be marked as a low confidence output.

9. The method for predicting gas diffusion in shield tunnels based on digital twins as described in claim 1, characterized in that: The targeted supplementation and optimization based on the calculation results are as follows: Computational structure optimization depth scoring and status synchronization data coverage score Based on the calculated structure optimization depth score and state synchronization data coverage score, a fusion confidence score is calculated. ; Set fusion confidence threshold By comparing the set fusion confidence threshold with the calculated fusion confidence score, and then making targeted supplementary optimizations based on the comparison results, we have: If the comparison results satisfy the formula This will trigger targeted supplementary optimization, specifically: in accordance with and Those with lower scores are identified as potential candidates for further optimization, and thus: like Then the drive structure optimization channel is for the first Each sub-unit will re-execute partition reconstruction; like To determine the cause of the missing covered node, we have: If the missing coverage node is caused by a data transmission interruption, the driver state synchronization channel will inject supplementary data. If the missing coverage node is due to objective limitations, the spatial interpolation estimate of the adjacent covered node is used as a substitute, and a data coverage insufficiency warning is generated. In addition, after each supplementary optimization is completed, channel fusion embedding, consistency verification, and fusion confidence assessment are re-executed, with the supplementary optimization being performed a maximum of [number] times. Wheel, if reached The comparison results still satisfy the formula. If the current optimal fusion result is retained, the current interval is marked as an interval with insufficient confidence and an early warning is generated.

10. The method for predicting gas diffusion in shield tunnels based on digital twins as described in claim 1, characterized in that: The process of performing shield tunnel gas diffusion prediction based on the optimized shield tunnel gas diffusion prediction model is as follows: Before the prediction is executed, the optimized prediction model is verified using reserved historical monitoring data. Based on the verified shield tunnel gas diffusion prediction model, the current real-time sensing data is used as the driving input to perform the full-area gas diffusion prediction of the shield tunnel, output the spatiotemporal distribution prediction results of gas concentration within the future time window, and issue gas disaster early warning based on the prediction results.

Images