Tunnel pavement concrete early-age strength prediction method based on multi-network fusion modeling

Abstract

The invention discloses a tunnel pavement concrete early-age strength prediction method based on multi-network fusion modeling. The method comprises the following steps: S1, continuously and equally dividing a tunnel center line according to a certain interval; s2, only keeping a left-right adjacent segment relation for the segment, and calculating a direction weighted thermal gradient, a direction weighted humidity gradient and a direction weighted traffic disturbance gradient based on the state vector; s3, writing the state vector of each segment into a sliding window with a set time length at a fixed sampling period; s4, inputting the dynamic graph input tensor into a space-time diagram neural network; s5, comparing the prediction sequence with a preset strength threshold value, identifying potential weak segments, and calculating risk confidence in combination with the gating coefficient; and S6, writing the maintenance instruction packet into a corresponding PLC channel. Potential weak sections can be recognized, spraying, heat preservation curtains and ventilation equipment can be linked, and local precise maintenance and batch open traffic unified management are achieved.

Description

Technical Field

[0001] This invention relates to the intersection of early-age concrete monitoring, intelligent maintenance and traffic control in tunnel engineering, and particularly to a method for predicting the early-age strength of tunnel pavement concrete based on multi-network fusion modeling. Background Technology

[0002] Current technical approaches for early-age management of tunnel lining concrete generally involve pre-embedding thermocouples, humidity sensors, or fiber Bragg gratings within the lining during the pouring process. Temperature, humidity, and strain data are then transmitted to a field acquisition box via wired or short-range wireless methods, and subsequently back to the central SCADA platform via fiber optic or 4G networks. Most systems use a fixed mileage or ring number as the primary key for retrieval, but the segment length is generally no less than one ring or one compartment, and the mapping between the numbering of each measuring point and the construction and maintenance systems is only achieved through manual maintenance.

[0003] In terms of data processing, publicly available literature generally employs maturity methods or exponential hydration equations, correlating temperature integral values ​​with previous experimental curves to calculate compressive strength at the target age. Temperature integrals are typically accumulated at 10-30 minute intervals, with humidity used only as a correction factor. Environmental disturbances and traffic loads are mostly considered external boundary conditions and not explicitly included in the calculation. Another approach introduces a simplified finite difference model, establishing a one-dimensional heat transfer equation for the lining section. Using measured temperature as the boundary, the temperature rise in the core area is iteratively calculated, indirectly obtaining the hydration heat release rate and inferring the strength. The model typically updates parameters daily; if the wind speed or humidity at the tunnel entrance fluctuates significantly in a short period, the convective heat transfer coefficient needs to be recalibrated by maintenance personnel.

[0004] Machine learning has also been used in recent years for early-age strength estimation. A common implementation process involves first collecting maturity and strength data of typical mix proportions under different temperature and humidity conditions offline, and then using multiple linear regression, shallow neural networks, or random forests to build a mapping model. In the online phase, after the system detects the latest temperature and humidity data, it directly calls the trained model to derive the strength. The advantages of this method are its simplicity and fast inference speed, but the model parameters are basically fixed after deployment. If different sections of the tunnel encounter differentiated ventilation or sprinkler interventions, manual retraining or weight adjustment is required to adapt the model.

[0005] However, current early-age management of tunnel linings often uses a single compartment or ring as the monitoring unit, resulting in a large spatial scale. The averaged temperature and humidity differences between the tunnel entrance and deeper layers make it difficult to expose local weaknesses. Data processing generally relies on fixed maturity curves or static empirical models, lacking an adaptive update mechanism for model parameters after deployment. When significant fluctuations occur in on-site ventilation, sprinkler systems, or traffic loads, manual recalibration or retraining is required to maintain accuracy. Multiple disturbances such as temperature, humidity, and load are often treated as independent boundary conditions, with coupling relationships not explicitly described, leading to insufficient sensitivity of inference results to complex environmental changes. The automatic control layer drives sprinklers, insulation curtains, and fans using single-factor thresholds, reflecting only surface temperature and humidity conditions and failing to consider the internal strength development rate. There is a lack of joint scheduling between devices, making it difficult to meet the needs of multiple sections under limited resources. Traffic management is usually based on unified age or construction schedule decisions, failing to implement segmented traffic opening according to real-time strength distribution, thus prolonging closure time or increasing safety hazards.

[0006] Therefore, in construction scenarios where temperature and humidity fluctuations are frequent within tunnels and the environment differs significantly between the tunnel entrance and the deep interior, it is urgent to construct a segmented strength inference mechanism that can infer the early-age strength of concrete in real time at a short spatial scale, and which can be adaptively updated over time and has physical interpretability. This mechanism would enable the system to promptly identify potential weak sections and coordinate with spraying, insulation curtains, and ventilation equipment to achieve precise local maintenance and unified management of phased traffic opening. Summary of the Invention

[0007] This invention provides a method for predicting the early-age strength of tunnel pavement concrete based on multi-network fusion modeling. This invention can identify potential weak sections and link spraying, insulation curtains and ventilation equipment to achieve localized precise maintenance and unified management of phased traffic opening.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A method for predicting the early-age strength of tunnel pavement concrete based on multi-network fusion modeling includes:

[0010] S1. Divide the tunnel centerline into equal parts at a certain interval, obtain a unique index for each segment and the corresponding start and end coordinates and mileage information; configure a unique real-time acquisition node for each segment, collect and splice temperature, relative humidity, hydration temperature rise, ventilation fan speed and cross-sectional traffic flow in a fixed field order to form an original state vector of fixed length.

[0011] S2. For the segmentation, only the left and right adjacent segment relationships are retained. Based on the state vector, the directional weighted thermal gradient, directional weighted humidity gradient and directional weighted traffic disturbance gradient are calculated. After being multiplied by the corresponding physical weights, they are linearly superimposed and interval compressed to obtain the normalized inter-segment edge weights, which are then filled into the weighted adjacency matrix.

[0012] S3. Write the state vector of each segment into a sliding window of a set duration with a fixed sampling period, stack the node feature matrix synchronously with time and concatenate it with the weighted adjacency matrix with the same timestamp to generate a three-dimensional dynamic graph input tensor.

[0013] S4. Input the dynamic graph into a tensor and then into a spatiotemporal graph neural network. First, perform graph convolution in the spatial dimension to aggregate information of the current segment and its left and right adjacent segments. Then, perform causal convolution in the temporal dimension to extract stage features. Calculate the hydration rate index based on the slope of the hydration temperature rise curve and generate a gating coefficient to adjust the cross-segment information throughput. Output a segmented compressive strength prediction sequence for a future set duration.

[0014] S5. Compare the predicted sequence with the preset intensity threshold to identify potential weak segments, and calculate the risk confidence level in combination with the gating coefficient. Then, package the target control parameters of the sprinkler, heat preservation curtain and ventilation fan according to priority to generate a maintenance instruction package.

[0015] S6. Write the maintenance instruction package into the corresponding PLC channel, execute the spraying, heat preservation and ventilation operations according to priority, calculate the earliest time when the road can be opened to traffic based on the maintenance completion time and intensity prediction results, and output the timetable for opening the road to traffic in batches.

[0016] Optionally, step S1 includes the following sub-steps:

[0017] S11. Divide the tunnel centerline into 5m units continuously and equally according to the construction mileage, and denote the unique index of the i-th segment as α. i For each α i Calculate the starting point coordinates (x) i y i , z i ), endpoint coordinates (x) i+1 y i+1 , z i+1 ) and the measured distance d at the entrance i , and then α i Mapped to site station β i With maintenance responsibility area γ i Complete the mapping, and set <α> i , (x i y i , z i ), (x i+1 y i+1 , z i+1 ), d i ,β i γ i The data is uploaded to the main monitoring station and simultaneously distributed to the construction, monitoring, and maintenance systems.

[0018] S12. Set a unique real-time acquisition node for each 5m unit, and record the node number as δ. i The internal monitoring quantity, temperature T i (t), relative humidity H i (t), hydration temperature rise R i (t) and external disturbance amount, fan speed F i (t), Cross-sectional traffic flow V i (t) Bind to δ according to a fixed field order i and with α i Establish a one-to-one mapping to form a consistent entity of spatial segments and data nodes.

[0019] Put all δ i Access to the tunnel monitoring network, at the network side for each δ i Set up an independent data channel κ i To ensure single-node failure, neighboring segment κ i+1 Data flow is unaffected.

[0020] S13, when δ i When data collection is triggered, the latest measured values ​​are sent in a fixed field order. Write to the cache and immediately concatenate them into a fixed-length original state vector: Vector s i Composed solely of physical measurements and on-site control quantities, without introducing any secondary statistical features, ensuring that each dimension can be directly traced back to the actual tunnel scene. After splicing, s i Through data channel κ i Push to the main monitoring station.

[0021] Optionally, step S2 includes the following sub-steps:

[0022] S21. Index the tunnel segments α i Negotiate and resolve the issue regarding number α. i The monitoring unit being pointed to retains only the adjacent number α on the left. i-1 Adjacent to the right, number α i+1 And the three are combined into a fixed neighborhood set. All subsequent discussions of coupling are limited to three physically continuous segments.

[0023] S22. Under the progressive logic of instantaneous temperature difference extraction, convection correction, exothermic amplification, and normalization, this sub-step connects the local driving force of the temperature field with construction disturbances into a two-way coupled thermal gradient:

[0024] Monitoring unit δ i The temperature at the current moment is denoted as T. i The temperatures of the left and right adjacent segments are denoted as T. i-1 T i+1Performing a difference operation on the three temperatures yields the original temperature difference vector: in

[0025] Read the fan speed F of this segment and the adjacent segment i F i-1 F i+1 With linear decay function φ i，L =1-k F |F i -F i-1 |、φ i，R =1-k F |F i -F i+1 Calculate the convection factor, constant k F ∈(0,1) is calibrated by ventilation parameters during construction; then the hydration temperature rise R is recorded. i (t) is obtained by performing a first-order difference. and use ψ i =r i / r max It is converted into an exothermic factor, where r max The measured peak value is the same as the mix proportion.

[0026] The vectorized operator is used to perform serial adjustment of the two factors and the temperature difference, and the directional weighted thermal gradient is obtained from the above equation:

[0027]

[0028] Among them, T i T i+1 The instantaneous temperature of the i-th segment and its adjacent segments; The left and right original temperature differences; sgn((·) is the element-wise sign function, outputting ±1; ⊙ is the Hadamard element-wise multiplication; F i F i+1 For fan speed setting; k F φ is the reduction factor for gear position difference. i，L , φ i，R R represents the left and right convection factors; i (t) represents the hydration temperature rise curve for this segment; r i r is the current heat release slope; max The reference peak slope for the same mix proportion; ψ i τ is the exothermic factor, with a value of [0.1]; T The normalized constant for temperature difference is set by the empirical threshold of 10℃ measured over a 5m segment; 1 represents a two-dimensional vector of all 1s; θ i，L θ i，R The direction-weighted thermal gradient is represented by

[01] , the magnitude of which represents the intensity of thermal diffusion, and the symbol indicates the direction of heat flow.

[0029] First use φi，* To suppress redundant temperature differences in the direction of strong convection of the fan, and then with ψ i Amplifying the thermal driving force during the active hydration period, ultimately through τ T The gradients of the min operation are limited to a comparable interval.

[0030] S23. Further, this sub-step follows the progressive logic of moisture difference extraction, water replenishment suppression, evaporation amplification, and interval regularization, embedding both internal and external water replenishment and evaporation drive into the humidity difference, and outputting a signed moisture migration gradient, providing a traceable criterion consistent with sprinkler and ventilation measures for subsequent inter-segment weights.

[0031] For monitoring unit δ i Reading real-time relative humidity H i The humidity of the left and right adjacent segments is denoted as H. i-1 H i+1 Performing a difference operation yields the original moisture difference vector. in

[0032] The water replenishment inhibition factor and evaporation amplification factor are embedded sequentially in two directions:

[0033] Hydration inhibitor σ i，L =1-k p |P i -P i-1 |、σ i，R =1-k p |P i -P i+1 |, where P i k represents the opening degree of the sprinkler valve in this section. p ∈(0,1) is determined by the nozzle flow rate; the larger the opening difference, the more sufficient the water replenishment, and the more the unidirectional wet difference is weakened.

[0034] Evaporation Amplification Factor Where T i D represents the air temperature. i U is the dew point temperature. i For ventilation speed, The calibration constant is used to normalize the typical evaporation potential of 10℃×2m / s to 1; the higher the temperature-dew point difference and wind speed, the stronger the evaporation driving force, thus amplifying the overall humidity migration tendency.

[0035] Finally, a vectorized operator is used to sequentially adjust the two factors and the humidity difference, outputting a direction-weighted humidity gradient:

[0036]

[0037] Wherein: H i H i+1 The instantaneous relative humidity of the i-th segment and its adjacent segments; The original moisture difference between the left and right sides; P i P i+1 For the spray valve opening (0-1 value); k p σ is the reduction factor for the opening difference; i，L , σ i，R It is a hydration inhibitor; T i D represents air temperature. i U is the dew point temperature. i Ventilation speed; The normalized constant of evaporation potential; ∈ i τ is the evaporation amplification factor, taking values ​​[0, 1]; sgn(·) is the element-wise sign function; ⊙ is the Hadamard element-wise multiplication; τ H The humidity difference normalization constant is set by a typical 15% RH threshold; 1 represents a two-dimensional all-1 vector; Λ i，L Λ i，R The direction-weighted humidity gradient is represented by the amplitude [0, 1], which characterizes the intensity of moisture migration, and the symbol indicates the direction of moisture flow.

[0038] First use σ i，* Reduce the moisture difference in the direction of high water replenishment, and then use ∈ i The migration driving force that amplifies high evaporation potential is finally activated via τ. H Normalization to a unified range achieves external water replenishment inhibition, internal evaporation promotion, and amplitude regularization.

[0039] S24. This step follows the progressive logic of load conversion, thermal gain coupling, adjacent difference, and amplitude regularization, and integrates vehicle vibration and shock with exhaust gas / wind heat effects into a signed traffic disturbance gradient within the segment level, providing a traceable criterion directly corresponding to operational traffic for coupling weight allocation.

[0040] For monitoring unit δ i Read three real-time traffic indicators: cross-sectional traffic flow V i (vehicles·min) -1 The proportion of heavy-duty vehicles (Q) i (0-1), average vehicle speed S i (km·h -1 The traffic flow-heavy load combination is used to form the equivalent dynamic load index. Where ω V To make 1 vehicle·min -1 The calibration factor is calculated by multiplying the 100% heavy load to the baseline impact energy; then the vehicle speed is converted into a heat / wind gain factor. Where ω S S is the amplification factor of vehicle speed on the exhaust gas thermal-wind field gain. ref The design reference speed is used. The product of these two values ​​yields the overall traffic disturbance value for this section. Next, for χ iWith the left and right adjacent segments χ i-1 χ i+1 Difference is calculated to form adjacent perturbation difference. To prevent numerical overflow caused by extreme traffic flow, a threshold constant τ is used. X Saturation compression is applied to the difference magnitude while preserving the direction sign to obtain the normalized traffic disturbance gradient:

[0041]

[0042] Among them, V i Let Q be the instantaneous traffic flow of the i-th segment; i The percentage of heavy-load vehicles in the same segment; ω V Traffic flow-heavy load conversion factor; S is the equivalent dynamic load index; i ω represents the average vehicle speed. S S is the thermal gain coefficient based on vehicle speed; ref The reference value for vehicle speed; ξ i χ represents the heat / air gain coefficient. i This represents the overall traffic disturbance value. The difference between left and right disturbances; τ X The normalization threshold for the disturbance difference is given by "2kN·min -1 "Empirical shock-thermal energy difference conversion; sgn(·) is the element-wise sign function; ⊙ is the Hadamard element-wise multiplication; 1 is a two-dimensional all-1 vector; γ" i，L γ i，R Let be the direction-weighted traffic disturbance gradient, with amplitude [0, 1] representing the vibration-thermal coupling energy level, and the sign indicating the energy flow direction.

[0043] The above formula captures wheel pressure impact through dynamic load conversion, maps vehicle speed and exhaust gas heat through thermal gain coupling, extracts energy conduction direction through adjacent differential calculation, and regularizes the amplitude to unify dimensions, ultimately outputting Υ. i，* It directly participates in subsequent linear superposition, ensuring that the potential impact of traffic disturbance on early-age concrete is fully, interpretably, and quantitatively reflected in the inter-segment coupling weights.

[0044] S25. Assign weights w to the same inter-segment edges according to their physical importance. T (Hot), w H (wet), w X (Transportation). Θ, Λ, Δ x The three gradient differences are multiplied by their corresponding weights in sequence and then summed to obtain the one-sided coupling score Ω. i，L or Ω i，R Then, perform interval compression on Ω to generate normalized edge weights g. i，L With g i，R This will cause it to fall into [0, 1].

[0045] S26. Simultaneously generate weights for all inter-segment edges, and then... m，n Fill the weighted adjacency matrix G according to the row and column correspondence of the segment numbers. Use matrix G as the instantaneous topological input of the graph neural network.

[0046] Optionally, step S3 includes the following sub-steps:

[0047] S31, will receive data from each 5m monitoring unit δ i The original state vector s i (t k Write the data into the sliding window of this segment in a fixed sampling rhythm τ. The window span is set to Θ = 3h, corresponding to a length L = Θ / τ. During writing, for s... i (t k Preserve field order [T] i H i R i F i V i [Unchanged, and time stamp t is recorded synchronously] k When k > L, the earliest frame at the beginning of the window is automatically removed to achieve constant-duration scrolling. Based on this, It covers the actual trajectory of temperature, humidity, hydration temperature rise, ventilation, and traffic disturbance over the past three hours at any given time, and each entry can be viewed... <t k α i The dual indexes trace back to both the moment of data collection and the spatial location.

[0048] S32. Index t for each time period within the sliding window. j ∈{t k-L+1 , ..., t k Perform the following two operations in sequence:

[0049] For all δ i Extract synchronization frame s i (t j And stack them in ascending order of segment number to form a node feature matrix S(t) j );

[0050] Retrieve the weighted adjacency matrix G(t) generated by S26 at the same timestamp j ). In obtaining <S(t j ),G(t j After that, the frames are concatenated in a time-first, then space-second order to construct a 3D dynamic graph input tensor:

[0051]

[0052] tensor The first dimension is sorted in ascending order of time, the second dimension corresponds one-to-one with the segment number, and the last dimension carries both real physical quantity fields and real-time coupled weights. This allows the network to capture both the temporal continuity of early-age heat release processes and the cross-segment transmission direction of temperature and humidity diffusion and traffic disturbances. After splicing, [the network will...] By directly inputting the backbone of the spatiotemporal graph neural network, the original measurement and control quantities and the dynamic topology structure are seamlessly integrated.

[0053] Optionally, step S4 includes the following sub-steps:

[0054] S41. Input the dynamic graph output from S3 into a tensor. Input into the spatiotemporal neural trunk, for Each time segment (t) j First, a first-order graph convolution kernel and the corresponding weighted adjacency matrix G(t) are used. j Convolutional operations are performed to aggregate only the information of the current segment and its left and right adjacent segments to match a 5m physical connectivity scale. The resulting spatial feature sequence is then input along the time axis into a causal temporal convolutional stack with an expansion coefficient of 2 to capture the phased coupling pattern between hydration heat release and external disturbances within a 3h window. After the convolutional layers and temporal convolutional layers are concatenated, a unified fusion feature vector h is output for each segment. i (t j ).

[0055] S42, Regarding the monitoring unit δ i The hydration temperature rise curve R within the same 3-hour window i (t) Perform piecewise linear fitting, dividing the window into equal-length 30-minute sub-intervals; calculate the slope of each sub-interval and then take the latest slope r. i As the current heat release rate, then according to equation ψ i =r i / r peak Normalization yields the hydration rate exponent ψ i ∈[0,1], where r peak This represents the slope of the historical peak value for the same mix proportion. This index monotonically maps the progress of cement hydration, revealing the activity level of strength development.

[0056] S43. Based on stage difference assessment, rate compatibility fusion, and nonlinear emphasis processing logic, this sub-step maps the differences in heat release progress between stages into edge vectorization gating coefficients in real time, and directly applies them to the message channel of the convolutional layer in the previous graph, ensuring that cross-stage information flow is synchronized with the hydration stage.

[0057] 1. For any edge (i,j) between segments, read the hydration rate exponent ψ at both ends. i , ψ j The absolute value of the difference between the two is taken to obtain the stage difference Δ. ij =|ψ i -ψj The larger the stage difference, the more inconsistent the stages of the two heat release processes. A subtraction operation is then performed on the stage difference to obtain the compatibility ρ. ij =1-Δ ij , ρ ij ∈[0,1] reflects the degree of matching in the exothermic phase.

[0058] 2. A rate lower bound-stage compatible linear fusion strategy is adopted, first selecting the smaller of the two rate exponents. This represents the commonly available exothermic activity, then weighted by λ∈(0,1). With ρ ij We perform a weighted summation; to further highlight the coupling sensitivity during the exothermic acceleration period, we introduce a power exponent γ>1 for nonlinear emphasis, and the gating coefficient is calculated as follows:

[0059]

[0060] Where: ψ i , ψ j Δ is the hydration rate exponent for segment i and j; ij Difference in hydration stages; ρ ij For stage compatibility; λ is the lower limit of the rate; λ is the linear fusion weight, taken as 0.6 based on field experience; γ is the nonlinear emphasis exponent, recommended to be 1.5 for construction purposes; g ij is the gating coefficient, ranging from [0, 1].

[0061] 3. The feature vector h from neighboring segment j in the convolutional layer of the previous image... j (t) according to element-wise product g ij h j (t) Inject into node i. If or ρ ij If it is lower, then g ij Compressed, cross-segment information injection is suppressed; when both ends enter the exothermic acceleration period and the stages are compatible, g ij The power-law amplification value is higher than the linear superposition value, thus enhancing the cross-segment information throughput. For node self-loops, directly set... This ensures that the contribution of self-generated heat in this segment amplifies synchronously with the heat release rate. The gating mechanism, through real-time tuning of message weights, prevents the slowed-down segment at the tunnel entrance from being over-assimilated by the accelerated segment in the deep interior, while ensuring sensitive coupling during the heat release peak, providing a physically consistent information flow allocation for weak segment identification.

[0062] S44, regarding the gating coefficient g ij Set the linear interpolation threshold: when ψ i ≥ψ high Scaled up proportionally (g) ij To η up When ψ i ≤ψ lowTime compression g ij To η down ψ high With ψ low The information flux is taken from the inflection point of the hydration curve with the same mix ratio. It can automatically change with the hydration stage without manual switching, making it sensitive during the acceleration period and stable during the deceleration period, keeping it in sync with the actual exothermic trend.

[0063] S45, Based on historical measured compressive equivalent strength of the same period Supervised training was performed on the entire network. During training, only the parameters of the backbone and prediction head were updated, while the hydration rate exponent ψ was kept constant. i As a fixed gating coefficient, it does not participate in gradient backpropagation, ensuring that its one-way mapping with the physical exothermic slope is not altered by model optimization.

[0064] S46. During the online phase, the real-time input tensor is processed by the backbone and gating layer to obtain the latest multi-layer feature z. i (t k Then, the data is fed into a fully connected prediction head to generate a segmented compressive strength prediction sequence with several steps in the time domain for the next 6 hours. Immediately The data is written into the tunnel monitoring master station to provide real-time decision-making basis for subsequent screening of weak sections, linkage of sprinkler and ventilation, and phased opening of traffic, thus completing the closed loop of spatiotemporal neural intensity prediction driven by hydration rate.

[0065] Optionally, step S5 includes the following sub-steps:

[0066] S51. Develop the intensity sequence of the next six hours output by the spatiotemporal graph neural prediction head. Perform threshold comparison for each segment number α i Read the corresponding threshold τ set by the construction technology department. i The two are encapsulated in key-value pair format as follows: And write it to the weak segment detection cache; synchronously record mileage l i With three-dimensional physical coordinates (x) i y i , z i This forms a complete index row.

[0067] S52. Perform a single-point check on each key-value pair in the cache: when Immediately change segment number α i Mark as lagging in intensity; otherwise mark as compliant. Write the judgment result and segment number into the risk bitmap. And add the current timestamp t for each lagging segment. flag .

[0068] S53, the inhibition factor g output by the hydration rate gating layer in this cycle. iiAlign with the risk bitmap by segment number, and extract g for each one-dimensional element marked as hysteretic. ii As a credibility benchmark, a 1-minus mapping strategy is used to generate the risk confidence level κ. i =1-g ii Stronger suppression means more difficulty in transmitting information across segments, resulting in a more conservative prediction of the segment's intensity and higher confidence level. (The last part, "<α"), appears to be a separate, unrelated statement and is left untranslated. i κ i >Write to the weak segment temporary table, and press κ i Set priority weights in descending order.

[0069] S54. Sort the temporary table first by priority weight, then by mileage. i Arrange the segments in ascending order for stability and output the final list of weak segments. The four columns of information in the list—mileage, predicted intensity, threshold, and confidence level—are then packaged and pushed to the tunnel monitoring master station, triggering a control logic traversal. For each α i Adjust the spray nozzles. i Insulation curtain i Ventilation fan φ i The predefined parameter templates are automatically written into the corresponding PLC channels.

[0070] Optionally, step S6 includes the following sub-steps:

[0071] S61. List of weak segments generated in S5 and its confidence level κ i By mileage i Import the tunnel monitoring master station in ascending order. Call a fixed risk classification rule for each segment: if κ i ≤θ1 is denoted as lower level, θ1<κ i ≤θ2 is denoted as intermediate, κ i >θ2 is designated as high-level, where θ1 and θ2 are the uniformly set grading thresholds during the construction period. Sub-queues Q are then generated within each of the three levels while maintaining mileage order. low、 Q mid Q high The data is then concatenated into a global priority queue Q, ordered by high, medium, and low priority. The status of each segment of Q is checked against the PLC channel status, and segment number α is determined. i With spray Π i Insulation curtain i Ventilation fan φ i Control interface binding.

[0072] S62. Traverse the priority queue Q using the parameter template library to call the function Lookup(α). i , level, d i Retrieval and segment priority as well as the distance d from the opening iThe three target parameters are: spray duration, insulation curtain closing time, and fan speed. Compare the target parameters with real-time measurements Comparison: If Then the delay factor δ is automatically added. Πi ;like Then for α i Mark for manual verification. After comparison is completed, package the segment-level maintenance instruction package. It is then written into the monitoring master station's execution queue to await system scheduling.

[0073] S63, Read all instruction packets Expected completion time Align it with the maintenance threshold decay curve of the same section, and calculate the earliest time when it can be opened to traffic using the rule of unlocking the highest priority first. Subsequently, each adjacent segment α not included in the maintenance was... i+1 Check the humidity gradient Λ i，±1 When |Λ i，±1 |≤∈ H Then it is considered stable and can be inserted into α. i Within the same release batch. The final generation will be a staggered traffic opening schedule arranged by mileage. The information is simultaneously pushed to on-site display screens and mobile terminals, guiding workers to perform only partial delays or enhanced maintenance on the sections listed, thus achieving precise coordination between early-age tunnel maintenance and phased opening to traffic.

[0074] Beneficial effects:

[0075] Existing methods use ring sections or compartments as monitoring units, resulting in low spatial resolution and fixed topology, making it difficult to characterize the differences between the tunnel entrance and the deeper parts. This scheme first divides the tunnel centerline into 5m segments and assigns each segment a unique number. Each segment is configured with an independent acquisition node. Temperature, relative humidity, hydration temperature rise, ventilation fan speed, and cross-sectional traffic flow are concatenated into a real-time state vector using fixed fields. Only the connections between left and right adjacent segments are retained. By differentially superimposing convection attenuation, evaporation amplification, and dynamic load correction coefficients based on temperature difference, humidity difference, and traffic load, a signed directional weighted gradient is obtained. Then, normalized edge weights are generated by linearly fusing the three physical weights of heat, humidity, and load. The weighted adjacency matrix is ​​dynamically refreshed with the sampling period, which differs from traditional graph construction methods that rely on static distance or pure similarity.

[0076] At the model level, this application employs a spatiotemporal graph neural network cascaded with first-order graph convolution and dilated causal temporal convolution, and proposes a rate exponent calculated based on the hydration heat slope as a gating mechanism: the rate difference between adjacent segments is used to obtain the stage compatibility, which is then fused with the rate lower limit through a power law to generate gating coefficients, which directly affect the side message channels to adjust the cross-segment information throughput in real time. The gating parameters do not participate in the training gradient backpropagation, ensuring the physical interpretability of the exothermic process and preventing the slowing segment at the cave entrance from being assimilated by the features of the deep high-temperature segment; compared with maturity integral, static machine learning, or ungated GNNs, this approach achieves adaptive coupling of the hydration stage within the model.

[0077] In terms of decision-making, this application compares the six-hour intensity prediction sequence with segment-level thresholds, combines gating inhibition factors to inversely map them into risk confidence levels and sorts them, automatically packages parameter packages for sprinklers, insulation curtains, and ventilators, and distributes them through PLC channels mapped one-to-one with segment numbers to achieve multi-device linkage; at the same time, it uses the prediction curve and maintenance completion time to calculate the earliest time when traffic can be opened, and outputs a segmented open list. The decision-making link is directly derived from the heat-humidity-load coupling diagram and hydration rate gating characteristics, which is fundamentally different from traditional single-factor thresholds or manual recalibration processes in terms of structure and implementation. Attached Figure Description

[0078] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0079] Figure 1 This is a flowchart of a method for predicting the early-age strength of tunnel pavement concrete based on multi-network fusion modeling proposed in this invention.

[0080] Figure 2 This is a schematic diagram comparing the early-age strength prediction method for tunnel pavement concrete based on multi-network fusion modeling proposed in this invention with existing technologies.

[0081] Figure 3 This is a flowchart of the three-dimensional dynamic graph input tensor construction process for the method of predicting the early-age strength of tunnel pavement concrete based on multi-network fusion modeling proposed in this invention.

[0082] Figure 4 This is a flowchart of the spatiotemporal graph neural network prediction method for predicting the early-age strength of tunnel pavement concrete based on multi-network fusion modeling proposed in this invention.

[0083] Figure 5 This is a flowchart illustrating the threshold comparison process for a method for predicting the early-age strength of tunnel pavement concrete based on multi-network fusion modeling proposed in this invention. Detailed Implementation

[0084] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0085] Example 1:

[0086] refer to Figure 1 A method for predicting the early-age strength of tunnel pavement concrete based on multi-network fusion modeling is presented. A comparative diagram of this invention's method with existing technologies is shown below. Figure 2 As shown, the flowchart for constructing the three-dimensional dynamic graph input tensor of the method for predicting the early-age strength of tunnel pavement concrete proposed in this invention is as follows: Figure 3 As shown, the spatiotemporal graph neural network prediction flowchart of the method for predicting the early-age strength of tunnel pavement concrete in this invention is as follows: Figure 4 As shown in the figure, the threshold comparison flowchart of the method for predicting the early-age strength of tunnel pavement concrete is as follows: Figure 5 As shown, it includes:

[0087] S1. Divide the tunnel centerline into equal parts at a certain interval, obtain a unique index for each segment and the corresponding start and end coordinates and mileage information; configure a unique real-time acquisition node for each segment, collect and splice temperature, relative humidity, hydration temperature rise, ventilation fan speed and cross-sectional traffic flow in a fixed field order to form an original state vector of fixed length.

[0088] S2. For the segmentation, only the left and right adjacent segment relationships are retained. Based on the state vector, the directional weighted thermal gradient, directional weighted humidity gradient and directional weighted traffic disturbance gradient are calculated. After being multiplied by the corresponding physical weights, they are linearly superimposed and interval compressed to obtain the normalized inter-segment edge weights, which are then filled into the weighted adjacency matrix.

[0089] S3. Write the state vector of each segment into a sliding window of a set duration with a fixed sampling period, stack the node feature matrix synchronously with time and concatenate it with the weighted adjacency matrix with the same timestamp to generate a three-dimensional dynamic graph input tensor.

[0090] S4. Input the dynamic graph into a tensor and then into a spatiotemporal graph neural network. First, perform graph convolution in the spatial dimension to aggregate information of the current segment and its left and right adjacent segments. Then, perform causal convolution in the temporal dimension to extract stage features. Calculate the hydration rate index based on the slope of the hydration temperature rise curve and generate a gating coefficient to adjust the cross-segment information throughput. Output a segmented compressive strength prediction sequence for a future set duration.

[0091] S5. Compare the predicted sequence with the preset intensity threshold to identify potential weak segments, and calculate the risk confidence level in combination with the gating coefficient. Then, package the target control parameters of the sprinkler, heat preservation curtain and ventilation fan according to priority to generate a maintenance instruction package.

[0092] S6. Write the maintenance instruction package into the corresponding PLC channel, execute the spraying, heat preservation and ventilation operations according to priority, calculate the earliest time when the road can be opened to traffic based on the maintenance completion time and intensity prediction results, and output the timetable for opening the road to traffic in batches.

[0093] Step S1 includes the following sub-steps:

[0094] S11. Divide the tunnel centerline into 5m units continuously and equally according to the construction mileage, and denote the unique index of the i-th segment as α. i For each α i Calculate the starting point coordinates (x) i y i , z i ), endpoint coordinates (x) i+1 y i+1 , z i+1 ) and the measured distance d at the entrance i , and then α i Mapped to site station β i With maintenance responsibility area γ i Complete the mapping, and set <α> i , (x i y i , z i ), (x i+1 y i+1 , z i+1 ),d i ,β i γ i The data is uploaded to the main monitoring station and simultaneously distributed to the construction, monitoring, and maintenance systems.

[0095] S12. Set a unique real-time acquisition node for each 5m unit, and record the node number as δ. i The internal monitoring quantity, temperature T i (t), relative humidity H i (t), hydration temperature rise R i (t) and external disturbance amount, fan speed F i (t), Cross-sectional traffic flow V i (t) Bind to δ according to a fixed field order i and with α i Establish a one-to-one mapping to form a consistent entity of spatial segments and data nodes.

[0096] Put all δ i Access to the tunnel monitoring network, at the network side for each δ i Set up an independent data channel κ i To ensure single-node failure, neighboring segment k i+1 Data flow is unaffected.

[0097] S13, when δi When data collection is triggered, the latest measured values ​​are sent in a fixed field order. Write to the cache and immediately concatenate them into a fixed-length original state vector: Vector s i Composed solely of physical measurements and on-site control quantities, without introducing any secondary statistical features, ensuring that each dimension can be directly traced back to the actual tunnel scene. After splicing, s i Through data channel κ i Push to the main monitoring station.

[0098] Step S2 includes the following sub-steps:

[0099] S21. Index the tunnel segments α i Negotiate and resolve the issue regarding number α. i The monitoring unit being pointed to retains only the adjacent number α on the left. i-1 Adjacent to the right, number α i+1 And the three are combined into a fixed neighborhood set. All subsequent discussions of coupling are limited to three physically continuous segments.

[0100] S22. Under the progressive logic of instantaneous temperature difference extraction, convection correction, exothermic amplification, and normalization, this sub-step connects the local driving force of the temperature field with construction disturbances into a two-way coupled thermal gradient:

[0101] Monitoring unit δ i Let the temperature at the current moment be denoted as Ti, and the temperatures of the left and right adjacent segments be denoted as T, respectively. i-1 T i+1 Performing a difference operation on the three temperatures yields the original temperature difference vector: in

[0102] Read the fan speed F of this segment and the adjacent segment i F i-1 F i+1 With linear decay function φ i，L =1-k F |F i -F i-1 |、φ i，R =1-k F |F i -F i+1| Calculate the convection factor, constant k F ∈(0,1) is calibrated by ventilation parameters during construction; then the hydration temperature rise R is recorded. i (t) is obtained by performing a first-order difference. and use ψ i =r i / r max It is converted into an exothermic factor, where r maxThe measured peak value is the same as the mix proportion.

[0103] The vectorized operator is used to perform serial adjustment of the two factors and the temperature difference, and the directional weighted thermal gradient is obtained from the above equation:

[0104]

[0105] Among them, T i T i+1 The instantaneous temperature of the i-th segment and its adjacent segments; The left and right initial temperature differences; sgn(·) is the element-wise sign function, outputting ±1; ⊙ is the Hadamard element-wise multiplication; F i F i+1 For fan speed setting; k F φ is the reduction factor for gear position difference. i，L , φ i，R R represents the left and right convection factors; i (t) represents the hydration temperature rise curve for this segment; r i r is the current heat release slope; max The reference peak slope for the same mix proportion; ψ i τ is the exothermic factor, taking values ​​[0, 1]; T The normalized constant for temperature difference is set based on the empirical threshold of 10℃ measured over a 5m segment; 1 represents a two-dimensional vector of all 1s; Θ i，L Θ i，R The direction-weighted thermal gradient is represented by the amplitude [0,1], which characterizes the intensity of thermal diffusion, and the symbol indicates the direction of heat flow.

[0106] First use φ i，* To suppress redundant temperature differences in the direction of strong convection of the fan, and then with ψ i Amplifying the thermal driving force during the active hydration period, ultimately through τ T The gradients of the min operation are limited to a comparable interval.

[0107] S23. Further, this sub-step follows the progressive logic of moisture difference extraction, water replenishment suppression, evaporation amplification, and interval regularization, embedding both internal and external water replenishment and evaporation drive into the humidity difference, and outputting a signed moisture migration gradient, providing a traceable criterion consistent with sprinkler and ventilation measures for subsequent inter-segment weights.

[0108] For monitoring unit δ i Reading real-time relative humidity H i The humidity of the left and right adjacent segments is denoted as H. i-1 H i+ Performing a difference operation yields the original moisture difference vector. in

[0109] The water replenishment inhibition factor and evaporation amplification factor are embedded sequentially in two directions:

[0110] Hydration inhibitor σ i，L =1-k p |P i -P i-1 |、σ i，R =1-k p |P i -P i+1 |, where P i k represents the opening degree of the sprinkler valve in this section. p ∈(0,1)) is calibrated by the nozzle flow rate; the larger the opening difference, the more sufficient the water replenishment, and the more the same-direction wet difference is weakened.

[0111] Evaporation Amplification Factor Where T i D represents the air temperature. i U is the dew point temperature. i For ventilation speed, The calibration constant is used to normalize the typical evaporation potential of 10℃×2m / s to 1; the higher the temperature-dew point difference and wind speed, the stronger the evaporation driving force, thus amplifying the overall humidity migration tendency.

[0112] Finally, a vectorized operator is used to sequentially adjust the two factors and the humidity difference, outputting a direction-weighted humidity gradient:

[0113]

[0114] Wherein: H i H i+1 The instantaneous relative humidity of the i-th segment and its adjacent segments; The original moisture difference between the left and right sides; P i P i+1 For the spray valve opening (0-1 value); k p σ is the reduction factor for the opening difference; i，L , σ i，R It is a hydration inhibitor; T i D represents air temperature. i U is the dew point temperature. i Ventilation speed; The normalized constant of evaporation potential; ∈ i τ is the evaporation amplification factor, taking values ​​[0, 1]; sgn(·) is the element-wise sign function; ⊙ is the Hadamard element-wise multiplication; τ H The humidity difference normalization constant is set by a typical 15% RH threshold; 1 represents a two-dimensional all-1 vector; Λ i，L Λ i，R The direction-weighted humidity gradient is represented by the amplitude [0, 1], which characterizes the intensity of moisture migration, and the symbol indicates the direction of moisture flow.

[0115] First use σ i，* Reduce the moisture difference in the direction of high water replenishment, and then use ∈i The migration driving force that amplifies high evaporation potential is finally activated via τ. H Normalization to a unified range achieves external water replenishment inhibition, internal evaporation promotion, and amplitude regularization.

[0116] S24. This step follows the progressive logic of load conversion, thermal gain coupling, adjacent difference, and amplitude regularization, and integrates vehicle vibration and shock with exhaust gas / wind heat effects into a signed traffic disturbance gradient within the segment level, providing a traceable criterion directly corresponding to operational traffic for coupling weight allocation.

[0117] For monitoring unit δ i Read three real-time traffic indicators: cross-sectional traffic flow V i (vehicles·min) -1 The proportion of heavy-duty vehicles (Q) i (0-1), average vehicle speed S i (km·h -1 The traffic flow-heavy load combination is converted into an equivalent dynamic load index θ. i =ω V V i Q i , where ω V To make 1 vehicle·min -1 The calibration factor is calculated by multiplying the 100% heavy load to the baseline impact energy; then the vehicle speed is converted into a heat / wind gain factor. Where ω S S is the amplification factor of vehicle speed on the exhaust gas thermal-wind field gain. ref The design reference speed is used. The product of these two speeds yields the comprehensive traffic disturbance value χ for this section. i =θ i ξ i Then for χ i With the left and right adjacent segments χ i-1 , χ i+1 Difference is calculated to form adjacent perturbation difference. To prevent numerical overflow caused by extreme traffic flow, a threshold constant τ is used. X Saturation compression is applied to the difference magnitude while preserving the direction sign to obtain the normalized traffic disturbance gradient:

[0118]

[0119] Among them, V i Let Q be the instantaneous traffic flow of the i-th segment; i The percentage of heavy-load vehicles in the same segment; ω V θ is the traffic flow-heavy load conversion factor; i S is the equivalent dynamic load index; i ω represents the average vehicle speed. S S is the thermal gain coefficient based on vehicle speed; ref The reference value for vehicle speed; ξi χ represents the heat / air gain coefficient. i This represents the overall traffic disturbance value. The difference between left and right disturbances; τ X The normalization threshold for the disturbance difference is given by "2kN·min -1 "Empirical shock - thermal energy difference conversion; sgn(·) is the element-wise sign function; ⊙ is the Hadamard element-wise multiplication; 1 is a two-dimensional all-1 vector; Υ" i，L γ i，R The gradient is a direction-weighted traffic disturbance gradient, with amplitude [0,1] representing the vibration-thermal coupling energy level and the sign indicating the energy flow direction.

[0120] The above formula captures wheel pressure impact through dynamic load conversion, maps vehicle speed and exhaust gas heat through thermal gain coupling, extracts energy conduction direction through adjacent differential calculation, and regularizes the amplitude to unify dimensions, ultimately outputting Υ. i，* It directly participates in subsequent linear superposition, ensuring that the potential impact of traffic disturbance on early-age concrete is fully, interpretably, and quantitatively reflected in the inter-segment coupling weights.

[0121] S25. Assign weights w to the same inter-segment edges according to their physical importance. T (Hot), w H (wet), w X (Transportation). The three gradient differences Θ, Λ, and Δx are multiplied sequentially by their corresponding weights and summed to obtain the one-sided coupling score Ω. i，L or Ω i，R Then, perform interval compression on Ω to generate normalized edge weights g. i，L With g i，R This will cause it to fall into [0, 1].

[0122] S26. Simultaneously generate weights for all inter-segment edges, and then... m，n Fill the weighted adjacency matrix G according to the row and column correspondence of the segment numbers. Use matrix G as the instantaneous topological input of the graph neural network.

[0123] Step S3 includes the following sub-steps:

[0124] S31, will receive data from each 5m monitoring unit δ i The original state vector s i (t k Write the data into the sliding window of this segment in a fixed sampling rhythm τ. The window span is set to Θ = 3h, corresponding to a length L = Θ / τ. During writing, for s... i (t k Preserve field order [T] i H i R i F i V i[Unchanged, and time stamp t is recorded synchronously] k When k > L, the earliest frame at the beginning of the window is automatically removed to achieve constant-duration scrolling. Based on this, It covers the actual trajectory of temperature, humidity, hydration temperature rise, ventilation, and traffic disturbance over the past three hours at any given time, and each entry can be viewed... <t k α i The dual indexes trace back to both the moment of data collection and the spatial location.

[0125] S32. Index t for each time period within the sliding window. j ∈{t k-L+1 ,...,t k Perform the following two operations in sequence:

[0126] For all δ i Extract synchronization frame s i (t j And stack them in ascending order of segment number to form a node feature matrix S(t) j );

[0127] Retrieve the weighted adjacency matrix G(t) generated by S26 at the same timestamp j ). In obtaining <S(t i ), G(t) j Then, the frames are concatenated in a time-first, then space-second order to construct a 3D dynamic graph input tensor:

[0128]

[0129] tensor The first dimension is sorted in ascending order of time, the second dimension corresponds one-to-one with the segment number, and the last dimension carries both real physical quantity fields and real-time coupled weights. This allows the network to capture both the temporal continuity of early-age heat release processes and the cross-segment transmission direction of temperature and humidity diffusion and traffic disturbances. After splicing, [the network will...] By directly inputting the backbone of the spatiotemporal graph neural network, the original measurement and control quantities and the dynamic topology structure are seamlessly integrated.

[0130] Step S4 includes the following sub-steps:

[0131] S41. Input the dynamic graph output from S3 into a tensor. Input into the spatiotemporal neural trunk, for Each time segment (t) j First, a first-order graph convolution kernel and the corresponding weighted adjacency matrix G(t) are used. jConvolutional operations are performed to aggregate only the information of the current segment and its left and right adjacent segments to match a 5m physical connectivity scale. The resulting spatial feature sequence is then input along the time axis into a causal temporal convolutional stack with an expansion coefficient of 2 to capture the phased coupling pattern between hydration heat release and external disturbances within a 3h window. After the convolutional layers and temporal convolutional layers are concatenated, a unified fusion feature vector h is output for each segment. i (t j ).

[0132] S42, Regarding the monitoring unit δ i The hydration temperature rise curve R within the same 3-hour window i (t) Perform piecewise linear fitting, dividing the window into equal-length 30-minute sub-intervals; calculate the slope of each sub-interval and then take the latest slope r. i As the current heat release rate, then according to equation ψ i =r i / r peak Normalization yields the hydration rate exponent ψ i ∈[0,1], where r peak This represents the slope of the historical peak value for the same mix proportion. This index monotonically maps the progress of cement hydration, revealing the activity level of strength development.

[0133] S43. Based on stage difference assessment, rate compatibility fusion, and nonlinear emphasis processing logic, this sub-step maps the differences in heat release progress between stages into edge vectorization gating coefficients in real time, and directly applies them to the message channel of the convolutional layer in the previous graph, ensuring that cross-stage information flow is synchronized with the hydration stage.

[0134] 3. For any edge (i, j) between segments, read the hydration rate exponent ψ at both ends. i , ψ j The absolute value of the difference between the two is taken to obtain the stage difference Δ. ij =|ψ i -ψ j The larger the stage difference, the more inconsistent the stages of the two heat release processes. A subtraction operation is then performed on the stage difference to obtain the compatibility ρ. ij =1-Δ ij , ρ ij ∈[0,1] reflects the degree of matching in the exothermic phase.

[0135] 4. Adopt a rate lower limit one-stage compatible linear fusion strategy, first taking the smaller of the two rate exponents. This represents the commonly available exothermic activity, then weighted by λ∈(0,1). With ρ ij We perform a weighted summation; to further highlight the coupling sensitivity during the exothermic acceleration period, we introduce a power exponent γ>1 for nonlinear emphasis, and the gating coefficient is calculated as follows:

[0136]

[0137] Where: ψ i , ψ j Δ is the hydration rate exponent for segment i and j; ij Difference in hydration stages; ρ ij For stage compatibility; λ is the lower limit of the rate; λ is the linear fusion weight, taken as 0.6 based on field experience; γ is the nonlinear emphasis exponent, recommended to be 1.5 for construction purposes; g ij is the gating coefficient, ranging from [0, 1].

[0138] 3. The feature vector h from neighboring segment j in the convolutional layer of the previous image... i (t) according to element-wise product g ij h j (t) Inject into node i. If or ρ ij If it is lower, then g ij Compressed, cross-segment information injection is suppressed; when both ends enter the exothermic acceleration period and the stages are compatible, g ij The power-law amplification value is higher than the linear superposition value, thus enhancing the cross-segment information throughput. For node self-loops, directly set... This ensures that the contribution of self-generated heat in this segment amplifies synchronously with the heat release rate. The gating mechanism, through real-time tuning of message weights, prevents the slowed-down segment at the tunnel entrance from being over-assimilated by the accelerated segment in the deep interior, while ensuring sensitive coupling during the heat release peak, providing a physically consistent information flow allocation for weak segment identification.

[0139] S44, regarding the gating coefficient g ij Set the linear interpolation threshold: when ψ i ≥ψ high Scaled up proportionally (g) ij To η up When ψ i ≤ψ low Time compression g ij To η down ψ high With ψ low The information flux is taken from the inflection point of the hydration curve with the same mix ratio. It can automatically change with the hydration stage without manual switching, making it sensitive during the acceleration period and stable during the deceleration period, keeping it in sync with the actual exothermic trend.

[0140] S45, Based on historical measured compressive equivalent strength of the same period Supervised training was performed on the entire network. During training, only the parameters of the backbone and prediction head were updated, while the hydration rate exponent ψ was kept constant. i As a fixed gating coefficient, it does not participate in gradient backpropagation, ensuring that its one-way mapping with the physical exothermic slope is not altered by model optimization.

[0141] S46. During the online phase, the real-time input tensor is processed by the backbone and gating layer to obtain the latest multi-layer feature z. i (t k Then, the data is fed into a fully connected prediction head to generate a segmented compressive strength prediction sequence with several steps in the time domain for the next 6 hours. Immediately The data is written into the tunnel monitoring master station to provide real-time decision-making basis for subsequent screening of weak sections, linkage of sprinkler and ventilation, and phased opening of traffic, thus completing the closed loop of spatiotemporal neural intensity prediction driven by hydration rate.

[0142] Step S5 includes the following sub-steps:

[0143] S51. Develop the intensity sequence of the next six hours output by the spatiotemporal graph neural prediction head. Perform threshold comparison for each segment number α i Read the corresponding threshold τ set by the construction technology department. i The two are encapsulated in key-value pair format as follows: And write it to the weak segment detection cache; synchronously record mileage l i With three-dimensional physical coordinates (x) i y i , z i This forms a complete index row.

[0144] S52. Perform a single-point check on each key-value pair in the cache: when Immediately change segment number α i Mark as lagging in intensity; otherwise mark as compliant. Write the judgment result and segment number into the risk bitmap. And add the current timestamp t for each lagging segment. flag .

[0145] S53, the inhibition factor g output by the hydration rate gating layer in this cycle. ii Align with the risk bitmap by segment number, and extract g for each one-dimensional element marked as hysteretic. ii As a credibility benchmark, a 1-minus mapping strategy is used to generate the risk confidence level κ. i =1-g ii Stronger suppression means more difficulty in transmitting information across segments, resulting in a more conservative prediction of the segment's intensity and higher confidence level. (The last part, "<α"), appears to be a separate, unrelated statement and is left untranslated. i κ i Write to the weak segment temporary table and press κ. i Set priority weights in descending order.

[0146] S54. Sort the temporary table first by priority weight, then by mileage. i Arrange the segments in ascending order for stability and output the final list of weak segments. The four columns of information in the list—mileage, predicted intensity, threshold, and confidence level—are then packaged and pushed to the tunnel monitoring master station, triggering a control logic traversal. For each α i Adjust the spray nozzles. i Insulation curtain i Ventilation fan φ i The predefined parameter templates are automatically written into the corresponding PLC channels.

[0147] S6 includes the following sub-steps:

[0148] S61. List of weak segments generated in S5 and its confidence level κ i By mileage i Import the tunnel monitoring master station in ascending order. Call a fixed risk classification rule for each segment: if κ i ≤θ1 is denoted as lower level, θ1<κ i ≤θ2 is denoted as intermediate, κ i >θ2 is designated as high-level, where θ1 and θ2 are the uniformly set grading thresholds during the construction period. Sub-queues Q are then generated within each of the three levels while maintaining mileage order. low Q mid Q high The data is then concatenated into a global priority queue Q, ordered by high, medium, and low priority. The status of each segment of Q is checked against the PLC channel status, and segment number α is determined. i With spray Π i Insulation curtain i Ventilation fan Φ i Control interface binding.

[0149] S62. Traverse the priority queue Q using the parameter template library to call the function Lookup(α). i level, d i Retrieval and segment priority as well as the distance d from the opening i The three target parameters are: spray duration, insulation curtain closing time, and fan speed. Compare the target parameters with real-time measurements Comparison: If Then the delay factor δ is automatically added. Πi ;like Then for α i Mark for manual verification. After comparison is completed, package the segment-level maintenance instruction package. It is then written into the monitoring master station's execution queue to await system scheduling.

[0150] S63, Read all instruction packets Expected completion time Align it with the maintenance threshold decay curve of the same section, and calculate the earliest time when it can be opened to traffic using the rule of unlocking the highest priority first. Subsequently, each adjacent segment α not included in the maintenance was... i+1 Check the humidity gradient Λ i，±1 When |Λ i，±1 |≤∈ H Then it is considered stable and can be inserted into α. i Within the same release batch. The final generation will be a staggered traffic opening schedule arranged by mileage. The information is simultaneously pushed to on-site display screens and mobile terminals, guiding workers to perform only partial delays or enhanced maintenance on the sections listed, thus achieving precise coordination between early-age tunnel maintenance and phased opening to traffic.

[0151] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the early-age strength of tunnel pavement concrete based on multi-network fusion modeling, characterized in that, The steps include: S1. Divide the tunnel centerline into equal parts at a certain interval, obtain a unique index for each segment and the corresponding start and end coordinates and mileage information; configure a unique real-time acquisition node for each segment, collect and splice temperature, relative humidity, hydration temperature rise, ventilation fan speed and cross-sectional traffic flow in a fixed field order to form an original state vector of fixed length. S2. For the segmentation, only the left and right adjacent segment relationships are retained. Based on the state vector, the directional weighted thermal gradient, directional weighted humidity gradient and directional weighted traffic disturbance gradient are calculated. After being multiplied by the corresponding physical weights, they are linearly superimposed and interval compressed to obtain the normalized inter-segment edge weights, which are then filled into the weighted adjacency matrix. S3. Write the state vector of each segment into a sliding window of a set duration with a fixed sampling period, stack the node feature matrix synchronously with time and concatenate it with the weighted adjacency matrix with the same timestamp to generate a three-dimensional dynamic graph input tensor. S4. Input the dynamic graph into a tensor and then into a spatiotemporal graph neural network. First, perform graph convolution in the spatial dimension to aggregate information of the current segment and its left and right adjacent segments. Then, perform causal convolution in the temporal dimension to extract stage features. The hydration rate index is calculated based on the slope of the hydration temperature rise curve, and a gating coefficient is generated to adjust the cross-segment information throughput, outputting a segmented compressive equivalent strength prediction sequence for a future set duration. S5. Compare the predicted sequence with the preset intensity threshold to identify potential weak segments, and calculate the risk confidence level in combination with the gating coefficient. Then, package the target control parameters of the sprinkler, heat preservation curtain and ventilation fan according to priority to generate a maintenance instruction package. S6. Write the maintenance instruction package into the corresponding PLC channel, execute the spraying, heat preservation and ventilation operations according to priority, calculate the earliest time when the road can be opened to traffic based on the maintenance completion time and intensity prediction results, and output the timetable for opening the road to traffic in batches.

2. The method for predicting the early-age strength of tunnel pavement concrete based on multi-network fusion modeling according to claim 1, characterized in that, Step S1 includes the following sub-steps: S11. Divide the tunnel centerline into 5m units continuously and equally according to the construction mileage, and denote the unique index of the i-th segment as α. i For each α i Calculate the starting point coordinates (x) i y i , z i ), endpoint coordinates (x) i+1 y i+1 , z i+1 ) and the measured distance d at the entrance i , and then α i Mapped to site station β i With maintenance responsibility area γ i Complete the mapping, and set <α> i ,(x i y i , z i ), (x i+1 y i+1 , z i+1 ), d i ,β i γ i The data is uploaded to the main monitoring station and simultaneously distributed to the construction, monitoring, and maintenance systems. S12. Set a unique real-time acquisition node for each 5m unit, and record the node number as δ. i The internal monitoring quantity, temperature T i (t), relative humidity H i (t), hydration temperature rise R i (t) and external disturbance amount, fan speed F i (t), Cross-sectional traffic flow V i (t) Bind to δ according to a fixed field order i and with α i Establish a one-to-one mapping to form a consistent entity of spatial segments and data nodes; Put all δ i Access to the tunnel monitoring network, at the network side for each δ i Set up an independent data channel κ i To ensure single-node failure, neighboring segment κ i±1 Data flow is unaffected; S13, when δ i When data collection is triggered, the latest measured values ​​are sent in a fixed field order. Write to the cache and immediately concatenate them into a fixed-length original state vector: Vector s i Composed solely of physical measurements and on-site control quantities, without introducing any secondary statistical features, ensuring that each dimension can be directly traced back to the actual tunnel scene. After splicing, s i Through data channel κ i Push to the main monitoring station.

3. The method for predicting the early-age strength of tunnel pavement concrete based on multi-network fusion modeling according to claim 1, characterized in that, Step S2 includes the following sub-steps: S21. Index the tunnel segments α i Negotiate and resolve the issue regarding number α. i The monitoring unit being pointed to retains only the adjacent number α on the left. i-1 Adjacent to the right, number α i+1 And the three are combined into a fixed neighborhood set. All subsequent discussions of coupling are limited to three physically continuous segments; S22. Under the progressive logic of instantaneous temperature difference extraction, convection correction, exothermic amplification, and normalization, this sub-step connects the local driving force of the temperature field with construction disturbances into a two-way coupled thermal gradient: Monitoring unit δ i The temperature at the current moment is denoted as T. i The temperatures of the left and right adjacent segments are denoted as T. i-1 T i+1 Performing a difference operation on the three temperatures yields the original temperature difference vector: in Read the fan speed F of this segment and the adjacent segment i F i-1 F i+1 With linear decay function φ i，L =1-k F |F i -F i-1 |、φ i，R =1-k F |F i -F i+1 | Calculate the convection factor, constant k F ∈(0,1) is calibrated by ventilation parameters during construction; then the hydration temperature rise R is recorded. i (t) is obtained by performing a first-order difference. and use ψ i =r i / r max It is converted into an exothermic factor, where r max This represents the measured peak value with the same mix proportion; The vectorized operator is used to perform serial adjustment of the two factors and the temperature difference, and the directional weighted thermal gradient is obtained from the above equation: Among them, T i T i±1 The instantaneous temperature of the i-th segment and its adjacent segments; The left and right initial temperature differences; sgn(·) is the element-wise sign function, outputting ±1; ⊙ is the Hadamard element-wise multiplication; F i F i±1 For fan speed setting; k F φ is the reduction factor for gear position difference. i，L , φ i，R R represents the left and right convection factors. i (t) represents the hydration temperature rise curve for this segment; r i r is the current heat release slope; max The reference peak slope for the same mix proportion; ψ i τ is the exothermic factor, with a value of [0.1]; T The normalized constant for temperature difference is set based on the empirical threshold of 10℃ measured over a 5m segment; 1 represents a two-dimensional vector of all 1s; Θ i，L Θ i,R For the directional weighted thermal gradient, the amplitude [0, 1] characterizes the intensity of thermal diffusion, and the sign indicates the direction of heat flow; First use φ i，* To suppress redundant temperature differences in the direction of strong convection of the fan, and then with ψ i Amplifying the thermal driving force during the active hydration period, ultimately through τ T The gradients of the min operation are limited to a comparable interval; S23. Further, this sub-step follows the progressive logic of moisture difference extraction, water replenishment suppression, evaporation amplification, and interval regularization, embedding both the internal and external water replenishment and evaporation drive into the humidity difference, and outputting a signed moisture migration gradient, providing a traceable criterion consistent with spraying and ventilation measures for subsequent inter-segment weights. For monitoring unit δ i Reading real-time relative humidity H i The humidity of the left and right adjacent segments is denoted as H. i-1 H i+1 Performing a difference operation yields the original moisture difference vector. in The water replenishment inhibition factor and the evaporation amplification factor are sequentially embedded in two directions; Hydration inhibitor σ i，L =1-k p |P i -P i-1 |、σ i，R =1-k p |P i -P i+1 |, where P i k represents the opening degree of the sprinkler valve in this section. p ∈(0,1) is determined by the nozzle flow rate; the larger the opening difference, the more sufficient the water replenishment, and the more the unidirectional wet difference is weakened; Evaporation Amplification Factor Where T i D represents the air temperature. i U is the dew point temperature. i For ventilation speed, The calibration constant is used to normalize the typical evaporation potential of 10℃×2m / s to 1; the higher the temperature-dew point difference and wind speed, the stronger the evaporation driving force, thus amplifying the overall humidity migration tendency. Finally, a vectorized operator is used to sequentially adjust the two factors and the humidity difference, outputting a direction-weighted humidity gradient: Wherein: H i H i±1 The instantaneous relative humidity of the i-th segment and its adjacent segments; The original moisture difference between the left and right sides; P i P i±1 For the spray valve opening, the value is 0-1; k p σ is the reduction factor for the opening difference; i，L , σ i，R It is a hydration inhibitor; T i D represents air temperature. i U is the dew point temperature. i Ventilation speed; The normalized constant of evaporation potential; ∈ i τ is the evaporation amplification factor, taking values ​​[0, 1]; sgn(·) is the element-wise sign function; ⊙ is the Hadamard element-wise multiplication; τ H The humidity difference normalization constant is set by a typical 15% RH threshold; 1 represents a two-dimensional all-1 vector; Δ i，L Λ i，R The direction-weighted humidity gradient is represented by the amplitude [0.1], which characterizes the intensity of moisture migration, and the sign indicates the direction of moisture flow. First use σ i，* Reduce the moisture difference in the direction of high water replenishment, and then use ∈ i The migration driving force that amplifies high evaporation potential is finally activated via τ. H Normalization to a uniform range achieves external water replenishment suppression, internal evaporation promotion, and amplitude regularization; S24. This step follows the progressive logic of load conversion, thermal gain coupling, adjacent difference, and amplitude regularization, and integrates vehicle vibration and shock with exhaust gas / wind heat effects into a signed traffic disturbance gradient within the segment level, providing a traceable criterion directly corresponding to operational traffic for coupling weight allocation. For monitoring unit δ i Read three real-time traffic indicators: cross-sectional traffic flow V i Unit vehicle·min -1 The proportion of heavy-duty vehicles Q i 0-1, average vehicle speed S i Unit: km·h -1 The traffic flow-heavy load combination is converted into an equivalent dynamic load index. Where ω v To make 1 vehicle·min -1 The calibration factor is calculated by multiplying the 100% heavy load to the baseline impact energy; then the vehicle speed is converted into a heat / wind gain factor. Where ω S S is the amplification factor of vehicle speed on the exhaust gas thermal-wind field gain. ref The product of the two values, taken as the design reference speed, yields the comprehensive traffic disturbance value χ for this segment. i =θ i ξ i Then for χ i With the left and right adjacent segments χ i-1 , χ i+1 Difference is calculated to form adjacent perturbation difference. To prevent numerical overflow caused by extreme traffic flow, a threshold constant τ is used. x Saturation compression is applied to the difference magnitude while preserving the direction sign to obtain the normalized traffic disturbance gradient: Among them, V i Let Q be the instantaneous traffic flow of the i-th segment; i The percentage of heavy-load vehicles in the same segment; ω V θ is the traffic flow-heavy load conversion factor; i S is the equivalent dynamic load index; i ω represents the average vehicle speed. S S is the thermal gain coefficient based on vehicle speed; ref The reference value for vehicle speed; ξ i χ represents the heat / air gain coefficient. i This represents the overall traffic disturbance value. The difference between left and right disturbances; τ x The normalization threshold for the disturbance difference is given by "2kN·min -1 "Empirical shock-thermal energy difference conversion; sgn(·) is the element-wise sign function; ⊙ is the Hadamard element-wise multiplication; 1 is a two-dimensional all-1 vector; γ" i，L γ i，R The direction-weighted traffic disturbance gradient is represented by the amplitude [0.1], which characterizes the vibration-thermal coupling energy level, and the sign indicates the energy flow direction. The above formula captures wheel pressure impact through dynamic load conversion, maps vehicle speed and exhaust gas heat through thermal gain coupling, extracts the energy conduction direction through adjacent differential calculation, and regularizes the amplitude to unify dimensions, ultimately outputting γ. i，* It directly participates in subsequent linear superposition to ensure that the potential impact of traffic disturbance on early-age concrete is fully, interpretably and quantitatively reflected in the inter-segment coupling weights; S25. Assign weights w to the same inter-segment edges according to their physical importance. T (Hot), w H (wet), w χ (Transportation). The three gradient differences Θ, Λ, and Δx are multiplied sequentially by their corresponding weights and summed to obtain the one-sided coupling score Ω. i，L or Ω i，R Then, perform interval compression on Ω to generate normalized edge weights g. i，L With g i，R This will cause it to fall into [0, 1]; S26. Perform weight generation synchronously on all inter-segment edges, and then... n Fill the weighted adjacency matrix G according to the row and column correspondence of the segment numbers. Use matrix G as the instantaneous topological input of the graph neural network.

4. The method for predicting the early-age strength of tunnel pavement concrete based on multi-network fusion modeling according to claim 1, characterized in that, Step S3 includes the following sub-steps: S31, will receive data from each 5m monitoring unit δ i The original state vector s i (t k Write the data into the sliding window of this segment in a fixed sampling rhythm τ. The window span is set to Θ = 3h, corresponding to a length L = Θ / τ. During writing, for s... i (t k Preserve field order [T] i H i ,R i F i V i [Unchanged, and time stamp t is recorded synchronously] k When k > L, the earliest frame at the beginning of the window is automatically removed to achieve constant-duration scrolling. Based on this, It covers the actual trajectory of temperature, humidity, hydration temperature rise, ventilation, and traffic disturbance over the past three hours at any given time, and each entry can be viewed... <t k α i The dual indexes trace back to the moment of collection and the spatial location. S32. Index t for each time period within the sliding window. j ∈{t k-L+1 , ..., t k Perform the following two operations in sequence: For all δ i Extract synchronization frame s i (t j And stack them in ascending order of segment number to form a node feature matrix S(t) j ); Retrieve the weighted adjacency matrix G(t) generated by S26 at the same timestamp j ). In obtaining <S(t j ), G(t) j Then, the frames are concatenated in a time-first, then space-second order to construct a 3D dynamic graph input tensor: tensor The first dimension is sorted in ascending order of time, the second dimension corresponds one-to-one with the segment number, and the last dimension carries both real physical quantity fields and real-time coupling weights. This allows the network to capture the temporal continuity of early-age heat release processes and to understand the cross-segment transmission direction of temperature and humidity diffusion and traffic disturbances, thus completing the splicing process. By directly inputting the backbone of the spatiotemporal graph neural network, the original measurement and control quantities and the dynamic topology structure are seamlessly integrated.

5. The method for predicting the early-age strength of tunnel pavement concrete based on multi-network fusion modeling according to claim 1, characterized in that, Step S4 includes the following sub-steps: S41. Input the dynamic graph output from S3 into a tensor. Input into the spatiotemporal neural trunk, for Each time segment (t) j First, a first-order graph convolution kernel and the corresponding weighted adjacency matrix G(t) are used. j Convolution operations are performed to aggregate only the information of the current segment and its left and right adjacent segments to match a 5m physical connectivity scale. The resulting spatial feature sequence is then input along the time axis into a causal temporal convolution stack with an expansion coefficient of 2 to capture the phased coupling pattern between hydration heat release and external disturbances within a 3h window. After the graph convolutional layer and the temporal convolutional layer are concatenated, a unified fusion feature vector h is output for each segment. i (t j ); S42, Regarding the monitoring unit δ i The hydration temperature rise curve R within the same 3-hour window i (t) Perform piecewise linear fitting, dividing the window into equal-length 30-minute sub-intervals; calculate the slope of each sub-interval and then take the latest slope r. i As the current heat release rate, then according to equation ψ i =r i / r peak Normalization yields the hydration rate exponent ψ i ∈[0,1], where r peak This represents the slope of the historical peak value for the same mix proportion. This index monotonically maps the progress of cement hydration, revealing the activity level of strength development; S43. Based on stage difference evaluation, rate compatibility fusion, and nonlinear emphasis processing logic, this sub-step maps the differences in heat release progress between stages into edge vectorization gating coefficients in real time, and directly applies them to the message channel of the convolutional layer in the previous graph, ensuring that cross-stage information flow is synchronized with the hydration stage. S44, regarding the gating coefficient g ij Set the linear interpolation threshold: when ψ i ≥ψ high Scaled up proportionally (g) ij To η up When ψ i ≤ψ low Time compression g ij To η down ψ high With ψ low Taken from the inflection point of the hydration curve with the same mix ratio, the information flux can be automatically changed with the hydration stage without manual switching, making it sensitive during the acceleration period and stable during the deceleration period, keeping it in sync with the actual exothermic trend. S45, Based on historical measured compressive equivalent strength of the same period Supervised training was performed on the entire network. During training, only the parameters of the backbone and prediction head were updated, while the hydration rate exponent ψ was kept constant. i As a fixed gating coefficient, it does not participate in gradient backpropagation, ensuring that its one-way mapping with the physical exothermic slope is not altered by model optimization; S46. During the online phase, the real-time input tensor is processed by the backbone and gating layer to obtain the latest multi-layer feature z. i (t k Then, the data is fed into a fully connected prediction head to generate a segmented compressive strength prediction sequence with several steps in the time domain for the next 6 hours. Immediately The data is written into the tunnel monitoring master station to provide real-time decision-making basis for subsequent screening of weak sections, linkage of sprinkler and ventilation, and phased opening of traffic, thus completing the closed loop of spatiotemporal neural intensity prediction driven by hydration rate.

6. The method for predicting the early-age strength of tunnel pavement concrete based on multi-network fusion modeling according to claim 5, characterized in that, S43 includes the following sub-steps: 1) Read the hydration rate exponent ψ at both ends of any edge (i, j) between segments. i , ψ j The absolute value of the difference between the two is taken to obtain the stage difference Δ. ij =|ψ i -ψ j The larger the stage difference, the more inconsistent the stages of the two heat release processes. A subtraction operation is then performed on the stage difference to obtain the compatibility ρ. ij =1-Δ ij , ρ ij ∈[0,1] reflects the degree of matching in the exothermic phase; 2) Adopting a rate lower limit one-stage compatible linear fusion strategy, first select the smaller of the two rate exponents. This represents the commonly available exothermic activity, then weighted by λ∈(0, 1). With ρ ij We perform a weighted summation; to further highlight the coupling sensitivity during the exothermic acceleration period, we introduce a power exponent γ>1 for nonlinear emphasis, and the gating coefficient is calculated as follows: Where: ψ i ,ψ j Δ is the hydration rate exponent for segment i and j; ij Difference in hydration stages; p ij For stage compatibility; λ is the lower limit of the rate; λ is the linear fusion weight, taken as 0.6 based on field experience; γ is the nonlinear emphasis exponent, recommended to be 1.5 for construction purposes; g ij The gating coefficient is in the range [0, 1]. 3) The feature vector h from the convolutional layer in the previous image comes from the neighboring segment j. j (t) according to element-wise product g ij h j (t) Inject node i, if or ρ ij If it is lower, then g ij Compressed, cross-segment information injection is suppressed; when both ends enter the exothermic acceleration period and the stages are compatible, g ij The power-law amplification value is higher than the linear superposition value, thus enhancing the cross-segment information throughput. For node self-loops, directly set... To ensure that the self-generated heat contribution of this segment amplifies synchronously with the heat release rate, the gating mechanism avoids excessive assimilation of the slowed segment at the tunnel entrance by the accelerated segment in the deep part by real-time tuning of message weights, while ensuring sensitive coupling during the heat release peak period, and providing physically consistent information flow allocation for weak segment identification.

7. The method for predicting the early-age strength of tunnel pavement concrete based on multi-network fusion modeling according to claim 1, characterized in that, Step S5 includes the following sub-steps: S51. Develop the intensity sequence of the next six hours output by the spatiotemporal graph neural prediction head. Perform threshold comparison for each segment number σ i Read the corresponding threshold τ set by the construction technology department. i The two are encapsulated in key-value pair format as follows: And write it to the weak segment detection cache; synchronously record mileage l i With three-dimensional physical coordinates (x) i y i , z i This forms a complete index row. S52. Perform a single-point check on each key-value pair in the cache: when Immediately change segment number α i Mark as lagging in intensity; otherwise mark as compliant. Write the judgment result and segment number into the risk bitmap. And add the current timestamp t for each lagging segment. flag ; S53, the inhibition factor g output by the hydration rate gating layer in this cycle. ii Align with the risk bitmap by segment number, and extract g for each one-dimensional element marked as hysteretic. ii As a credibility benchmark, a 1-minus mapping strategy is used to generate the risk confidence level κ. i =1-g ii Stronger suppression means more difficulty in transmitting information across segments, resulting in a more conservative prediction of the segment's intensity and higher confidence level. (The last part, "<α"), appears to be a separate, unrelated statement and is left untranslated. i , κ i >Write to the weak segment temporary table, and press κ i Set priority weights in descending order; S54. Sort the temporary table first by priority weight, then by mileage. i Arrange the segments in ascending order for stability and output the final list of weak segments. The four columns of information in the list—mileage, predicted intensity, threshold, and confidence level—are then packaged and pushed to the tunnel monitoring master station, triggering a control logic traversal. For each α i Adjust the spray nozzles. i Insulation curtain i Ventilation fan φ i The predefined parameter templates are automatically written into the corresponding PLC channels.

8. The method for predicting the early-age strength of tunnel pavement concrete based on multi-network fusion modeling according to claim 1, characterized in that, Step S6 includes the following sub-steps: S61. List of weak segments generated in S5 and its confidence level κ i By mileage i Import the tunnel monitoring master station in ascending order. Call a fixed risk classification rule for each segment: if κ i ≤θ1 is denoted as lower level, θ1<κ i ≤θ2 is denoted as intermediate, κ i >θ2 is designated as high-level, where θ1 and θ2 are the uniformly set grading thresholds during the construction period. Sub-queues are then generated within each of the three levels while maintaining mileage order. The data is then concatenated into a global priority queue Q based on high, medium, and low priority levels. The status of each segment of Q is checked against the PLC channel status, and segment number α is determined. i With spray Π i Insulation curtain i Ventilation fan Φ i Control interface binding; S62. Traverse the priority queue Q using the parameter template library to call the function Lookup(α). i , level, d i Retrieval and segment priority as well as the distance d from the opening i The three target parameters are: spray duration, insulation curtain closing time, and fan speed. Compare the target parameters with real-time measurements Comparison: If Then the delay factor δ is automatically added. Πi ;like Then for α i Mark for manual verification. After comparison is completed, package the segment-level maintenance instruction package. And write it to the monitoring master station's execution queue to await system scheduling; S63, Read all instruction packets Expected completion time Align it with the maintenance threshold decay curve of the same section, and calculate the earliest time when it can be opened to traffic using the rule of unlocking the highest priority first. Subsequently, each adjacent segment α not included in the maintenance was... i±1 Check the humidity gradient Λ i±1 When |Λ i±1 |≤∈ H Then it is considered stable and can be inserted into α. i Within the same release batch, a timetable for phased reopening of transportation, arranged in mileage order, will be generated. The information is simultaneously pushed to on-site display screens and mobile terminals, guiding workers to perform only localized delays or enhanced maintenance on the sections listed, thus achieving precise coordination between early-age tunnel maintenance and phased opening to traffic.

Images