Intelligent prevention and control method for concrete construction rain damage

Abstract

The application discloses a kind of concrete construction rain damage intelligent prevention and control method, belong to concrete construction quality monitoring and maintenance control field, to solve the problem that rain intensity, raindrop kinetic energy, pouring surface damage and maintenance temperature and humidity cannot form continuous quantitative decision chain under the condition of sudden rainfall, the application is carried out attenuation correction to X band dual polarization radar data, the corrected polarization parameter is interpolated to Cartesian rule grid, then raindrop spectrum distribution classification is carried out, combined with pouring zoning map, double time scale memory and neural order preserving controller generates construction risk interval, risk interval, radar reliability and construction action are coded as the uncertainty boundary of physical information neural network, deduce intensity, temperature gradient and crack risk, and generate construction and maintenance instructions through interval robust gate, realize the continuous quantification and collaborative control of rain damage risk.

Description

Technical Field

[0001] This invention relates to the field of concrete construction quality monitoring and curing control, and in particular to an intelligent method for preventing and controlling rain damage in concrete construction. Background Technology

[0002] The construction of large-scale concrete structures (such as the Pinglu Canal Hub, long-span cross-sea bridges, shipping locks, and large aqueducts) and new energy infrastructure such as wind power generation foundations, photovoltaic support foundations, and energy storage power stations is highly sensitive to sudden rainfall events, often occurring within minutes. During the pouring stage, the impact of short-duration heavy rainfall and excessive rainwater mixing disrupt the surface water-cement ratio balance, easily leading to surface laitance, exposed aggregates, and even cold joints. During the curing stage, the cooling effect of rainfall couples with the internal temperature and humidity field of the concrete, causing nonlinear changes in hydration heat release and moisture migration, accumulating and increasing the risk of cracking. These rain-related hazards persist throughout the entire pouring and curing process and have become significant factors affecting project duration and structural durability.

[0003] Current rain damage prevention and control decisions are usually based on ground-based single-point rainfall observations and manual judgment of weather forecasts. The information sources are extensive and lack quantitative criteria with spatiotemporal precision at the hundred-meter or minute level. The response to the suddenness and localization of short-term heavy rainfall is lagging. Protection relies on static protective devices such as rain shelters and plastic film coverings, which cannot dynamically adjust the covering status and coordinate maintenance strategies with real-time rainfall. In terms of assessing surface damage to the poured surface after rain, existing methods mainly rely on on-site inspections and acceptance, judging the degree of damage based on visual characteristics such as floating slurry, exposed aggregate, and water accumulation. There is a lack of quantitative correlation with real-time rainfall and continuous quantification of damage status.

[0004] X-band dual-polarization radar has a spatial resolution of ≤75m and a temporal resolution of better than 5 minutes. It can invert micro-parameters such as raindrop spectrum based on polarization parameters such as reflectivity factor (ZH), differential reflectivity (ZDR), and differential propagation phase shift rate (KDP), and realize the perception and precise control of precipitation at tens of meters grid level for concrete pouring targets. The spatial granularity is significantly better than traditional kilometer-level precipitation monitoring and forecasting products. It is the preferred data source to support the prevention and control of rain damage in concrete construction. However, the existing applications have the following problems: (1) The application is extensive and the accuracy is insufficient. Existing QPE products generally rely on fixed empirical ZR relationships and have not been fully attenuated and corrected. The inversion of rainfall intensity in the core area of ​​the rainstorm is inaccurate. The difference is as high as 40% to 60%, which is enough to cause misjudgment of "stop watering / cover / continue"; (2) The precipitation dimension is singular; the existing precipitation products only output scalar indicators such as rainfall intensity and cumulative rainfall, which merge the two completely different physical mechanisms of instantaneous raindrop impact and continuous wetting into a single cumulative rainfall, ignoring the dynamic damage of impact kinetic energy to the surface water-cement ratio, and cannot quantify the damage process of "raindrop impact - surface water-cement ratio imbalance - slurry / exposed stone"; (3) Data and decision are separated; there is a lack of a complete mapping chain from "polarization parameter → raindrop spectrum → impact kinetic energy → surface damage → construction action", and high-precision precipitation information cannot be automatically converted into construction control instructions.

[0005] On the other hand, in terms of construction and maintenance control, although existing intelligent maintenance has the ability to collect real-time temperature and humidity inside concrete and manage construction data in an information-based manner, it has not yet been linked with precipitation sensing: the maturity method (ASTM C1074) only uses a single parameter of temperature, without coupling the humidity field and meteorological boundary, nor does it introduce the rainfall events experienced by the component during the pouring period; the temperature and humidity sensing system relies on manually fixed threshold alarms, lacks the ability to predict cracking risks in advance, and the fixed threshold cannot express the nonlinear relationship between rainfall input and maintenance actions; the BIM construction management platform does not embed meteorological remote sensing data as a dynamic boundary into the hydration heat-humidity coupling model.

[0006] Therefore, there is a need for an intelligent method for preventing and controlling rain damage in concrete construction that can address the shortcomings of existing technologies. Summary of the Invention

[0007] One objective of this invention is to propose an intelligent method for preventing and controlling rain damage in concrete construction. Addressing the problems of existing technologies where rain damage prevention decisions rely on manual judgment, maintenance criteria lack forward-looking projections of cracking risks during rainfall, and rainfall perception, pouring surface damage assessment, and construction and maintenance control are independent and difficult to coordinate, this invention proposes a method that attenuates and corrects X-band dual-polarization radar data, interpolates the corrected polarization parameters to a Cartesian regular grid, and then classifies raindrop spectrum distribution to form rainfall event characteristics and mass vectors. It utilizes pouring zoning maps, graph spatiotemporal models, and dual-timescale memory to generate damage states and risk trends. A neural order-preserving controller with monotonic constraints outputs construction risk intervals, and encodes the risk intervals, radar reliability, and initial construction instructions into the initial and boundary condition uncertainties of a physical information neural network, assimilating the intensity, temperature gradient, and crack risk projections from temperature, humidity, and meteorological data. Finally, it uses interval robust gating to generate construction and maintenance control instructions. This invention achieves the technical effect of continuously quantifying rain damage and collaboratively controlling the construction and maintenance processes.

[0008] This invention provides an intelligent method for preventing and controlling rain damage during concrete construction, comprising:

[0009] S1. Collect X-band dual-polarization radar, pouring surface condition, internal temperature and humidity of concrete, ground meteorological data, construction stage and sensor observation data, perform attenuation correction on radar data, interpolate the corrected polarization parameters to a Cartesian regular grid, and then perform raindrop spectrum distribution classification to generate rainfall event characteristics and rainfall event quality vector.

[0010] S2. Establish a pouring zone map, input rainfall event characteristics and pouring surface status data into the spatiotemporal model and perform dual time scale memory processing to generate the damage status and risk trend of each pouring zone.

[0011] S3. Determine the construction stage based on the construction stage data, and calculate the sensor residual based on the sensor observation data. Input the damage state, risk trend, construction stage and sensor residual into a neural sequence-preserving controller with monotonicity constraints to generate a construction risk interval consisting of the lower bound of construction risk, the center value of construction risk and the upper bound of construction risk.

[0012] S4. Generate initial construction instructions based on the construction risk interval, encode the construction risk interval, radar credibility represented by the quality vector of rainfall events, and initial construction instructions into the initial conditions and boundary condition uncertainties of the physical information neural network, assimilate the internal temperature and humidity of concrete and ground meteorological data, and generate the lower limit of strength, the upper limit of temperature gradient, and the crack risk interval.

[0013] S5. Input the construction risk range, lower limit of strength, upper limit of temperature gradient, and crack risk range into the range robust gating controller, and generate construction control instructions and maintenance control instructions according to the risk upper limit triggering and risk lower limit release rules.

[0014] Optionally, S1 includes:

[0015] The path integral attenuation correction amount is determined based on the spatiotemporal registration results of ground rainfall and radar reflectivity, and the differential reflectivity and differential propagation phase shift are corrected using the path integral attenuation correction amount.

[0016] The raindrop spectrum distribution type is identified and the rain intensity and raindrop kinetic energy are calculated based on the interpolated corrected polarization parameters.

[0017] The rainfall event quality vector is generated based on the attenuation correction residual, raindrop spectrum distribution classification confidence, data spatiotemporal integrity, and ground rainfall consistency. The rainfall event quality vector is then fed back to update the attenuation correction weight and raindrop spectrum distribution classification weight for subsequent radar data.

[0018] Optionally, S2 includes:

[0019] The pouring surface is divided into nodes according to the pouring sequence, pouring time, elevation and adjacent boundaries. Edges are constructed based on the material continuity relationship and time interval between adjacent nodes to form the pouring zoning diagram.

[0020] Each node is provided with a first timescale memory unit and a second timescale memory unit, wherein the decay period of the first timescale is less than the decay period of the second timescale.

[0021] The impact state caused by raindrop kinetic energy extracted from the rainfall event features is accumulated using the first time-scale memory unit, and the rainfall duration extracted from the rainfall event features and the surface water content extracted from the pouring surface state data are accumulated using the second time-scale memory unit, and then fused to obtain the dual time-scale memory state.

[0022] Furthermore, it also includes: extracting surface texture changes, slurry coverage ratio, aggregate exposure ratio and water depth from the pouring surface state data, and calculating the time interval between adjacent pouring zones based on the pouring time used to divide the nodes;

[0023] The surface texture changes, slurry coverage ratio, aggregate exposure ratio and water depth are aggregated with the dual time scale memory state to generate scour damage component, slurry damage component and exposed stone damage component.

[0024] The continuous pouring time window is determined by the difference between the initial setting time obtained from the concrete setting test of the same mix proportion and the transportation, paving and vibration time in the construction record. The time interval between adjacent pouring zones is compared with the continuous pouring time window to generate a cold joint risk component.

[0025] The damage state is composed of each damage component and the cold seam risk component, and the risk trend is generated based on the damage state at continuous sampling times.

[0026] Optionally, S3 includes:

[0027] The neural order-preserving controller is trained using damage-labeled samples and construction action result samples, and the lower bound of construction risk is constrained to be no greater than the construction risk center value and the construction risk center value is constrained to be no greater than the construction risk upper bound.

[0028] Non-negative monotonic weights are set for the dual-timescale memory state, the damage state, and the sensor residual, respectively, so that the lower bound of construction risk, the center value of construction risk, and the upper bound of construction risk do not decrease when the corresponding input increases;

[0029] The width of the construction risk interval is adjusted according to the interval uncovered error quantile of the calibration sample, so that the coverage rate corresponding to the interval uncovered error quantile reaches the preset coverage rate.

[0030] Optionally, S4 includes:

[0031] Set an emergency cover threshold and a work stoppage gap threshold, wherein the emergency cover threshold is less than the work stoppage gap threshold;

[0032] When the construction risk center value is less than the emergency cover threshold, an initial construction instruction to continue construction is generated; when the construction risk center value reaches the emergency cover threshold but is less than the work stoppage and gap-retention threshold, an initial construction instruction to cover the emergency is generated; when the construction risk center value reaches the work stoppage and gap-retention threshold, an initial construction instruction to stop work and leave gaps is generated. The two thresholds are determined by the allowable scour test results and allowable interval time of the corresponding concrete mix proportion.

[0033] Map the width of the construction risk range to an initial condition parameter range;

[0034] The attenuation correction residual, raindrop spectrum distribution classification confidence, data spatiotemporal integrity and ground rainfall consistency are read from the rainfall event quality vector. The attenuation correction residual is converted into a residual confidence component that does not increase with the increase of the residual. Each confidence component is normalized to a closed interval of zero to one. The normalized confidence components are arranged in ascending order of value and the confidence component value at the first position is used as the radar confidence.

[0035] The difference between one and the radar confidence level is used as the confidence level complement, and the product of the confidence level complement and the preset upper limit of the rainfall boundary disturbance is determined as the rainfall boundary disturbance range. When any confidence component used in step S4 is invalid or the transmission verification of the rainfall event quality vector fails, the confidence level complement is set to one.

[0036] The initial construction command is converted into a pouring flux boundary, a heat exchange coverage boundary, or a construction interval boundary, thus forming the initial condition uncertainty and the boundary condition uncertainty.

[0037] Furthermore, it also includes: setting hydration reaction, heat conduction and water migration constraints in the physical information neural network, and constructing parameter value ranges based on the initial condition uncertainty and the boundary condition uncertainty;

[0038] The parameter value range is corrected by the observation residuals of the internal temperature and humidity data of the concrete and the ground meteorological data, and the concrete strength, temperature field and humidity field are deduced within the parameter value range;

[0039] The lower limit of strength and the upper limit of temperature gradient are generated based on the quantile boundaries of each simulation result, and the crack risk range is generated based on the comparison between temperature stress and tensile strength that varies with strength.

[0040] Optionally, S5 includes:

[0041] Map the construction risk zone to construction control, and map the crack risk zone to maintenance control;

[0042] When the upper limit of the construction risk reaches the emergency cover threshold, an emergency cover construction control command is triggered. When the upper limit of the construction risk reaches the stop-work and crack-leaving threshold, a stop-work and crack-leaving construction control command is triggered. When the upper limit of the crack risk range reaches the curing trigger threshold determined by the allowable water loss and allowable temperature stress, a curing control command is triggered and a holding timer is started.

[0043] The minimum holding time is set separately for construction control actions and maintenance control actions, and is determined by the preset quantile of the holding time in the corresponding action history execution record;

[0044] After the minimum holding time is reached, the control command is released only when the lower bound of the corresponding risk interval is less than the corresponding action release threshold. The action release threshold is determined by subtracting the hysteresis width from the corresponding action trigger threshold. The hysteresis width is determined by the preset quantile of the difference between adjacent trigger values ​​and release values ​​in the historical action switching record.

[0045] When the recovery threshold is set to be less than the failure threshold, the sensor abnormality branch is entered when the sensor residual reaches the failure threshold or the number of effective sensors is less than the preset number of effective sensors. The sensor abnormality branch takes priority over normal triggering, normal cancellation and action conflict handling.

[0046] The sensor abnormal branch remains in an abnormal state and the protection action based on the lower boundary of the risk range is prohibited. The sensor recovery branch serves as the exit condition for the sensor abnormal branch and exits the abnormal state only when all conditions of the sensor recovery branch are met.

[0047] If the current action is one of the following: stop work and leave gap, emergency cover, maintain cover or postpone demolding, then keep the current action; otherwise, switch to the preset protection action.

[0048] When multiple construction actions conflict, the priority is determined in the order of stopping work to leave gaps, emergency covering, and continuing construction. When formwork removal actions conflict, delaying formwork removal takes priority.

[0049] The sensor recovery branch is entered only when the sensor residual is less than the recovery threshold and the number of effective sensors reaches the preset effective number, so that all failure conditions are not met and the recovery stabilization time continues. The construction risk range is recalculated using the recovered sensor data, and the physical information neural network is rerun. The updated lower limit of intensity and upper limit of temperature gradient are generated based on the quantile boundaries of each inference result after rerun, and the updated crack risk range is generated. The maintained protection action or the preset protection action is then processed according to each updated result and the corresponding action to remove the threshold. The recovery stabilization time is determined by the product of the data sampling interval and the preset number of consecutive effective samplings.

[0050] Furthermore, it also includes: delaying demolding when the lower bound of the intensity generated by the quantile boundary of each of the inference results is less than the demolding intensity threshold, and releasing the delayed demolding command when the lower bound of the intensity reaches the demolding intensity threshold and continues for the minimum holding time.

[0051] When the upper limit of the temperature gradient reaches the temperature control trigger threshold, the insulation coverage thickness or the cooling medium flow rate is adjusted. When the upper limit of the temperature gradient is less than the temperature control release threshold and continues for the minimum holding time, the temperature control state before adjustment is restored. The temperature control release threshold is determined by subtracting the temperature control hysteresis width from the temperature control trigger threshold.

[0052] When the upper boundary of the crack risk zone reaches the maintenance trigger threshold, the spraying interval and coverage status are adjusted, and the adjustment is lifted according to the lower boundary of the crack risk zone and the corresponding action release threshold.

[0053] The demolding strength threshold is determined by the strength test results of specimens with the same mix proportion; the temperature control trigger threshold is determined by the allowable temperature difference design value; and the temperature control hysteresis width is determined by the preset quantile of the difference between the temperature gradient trigger value and the release value in the historical temperature control action record.

[0054] Furthermore, it also includes: recording the data version, risk range, radar reliability, sensor residual, start and end times of the action, and action results for each risk upper boundary trigger, risk lower boundary release, protection action switching, and sensor recovery;

[0055] Independent verification sample sets for construction risk intervals and crack risk intervals are established separately. The independent verification sample sets for construction risk intervals do not overlap with the training and calibration samples of the neural sequence-preserving controller, and the independent verification sample sets for crack risk intervals do not overlap with the parameter fitting samples of the physical information neural network.

[0056] The calibration samples and parameters of the neural sequence-preserving controller are updated based on the deviation between the action result and the subsequent damage state of the poured surface, and the observation weights and parameters of the physical information neural network are updated based on the deviation between the action result and the subsequent internal temperature and humidity data of the concrete.

[0057] The updated neural sequence-preserving controller is applied to the independent verification sample set of the construction risk interval, and the updated construction risk interval coverage is recalculated. The updated physical information neural network is applied to the independent verification sample set of the crack risk interval, and the updated crack risk interval coverage is recalculated.

[0058] The preset coverage rate is used as the target coverage rate of the construction risk zone. The target coverage rate of the crack risk zone is determined by the coverage requirements of the crack marking zone. When any updated zone coverage rate is less than the corresponding target coverage rate, the calibration samples and parameters of the neural sequence-preserving controller and the observation weights and parameters of the physical information neural network are restored to the same data version before the update.

[0059] The beneficial effects of this invention are:

[0060] 1. By performing attenuation correction, raindrop spectrum distribution classification, and mass vector feedback on the X-band dual polarization radar, and combining it with the pouring zoning map and dual time scale memory, the raindrop impact state and the continuous wetting state can be distinguished in the same data link, so that the zoning damage and risk trend of the pouring surface have a continuous data source.

[0061] 2. By generating the lower bound, center value and upper bound of construction risk through a neural sequence-preserving controller with monotonicity constraints, and adopting interval robust gating with upper bound triggering, lower bound release, holding time and sensor anomaly priority, construction and maintenance control actions with hysteresis can be formed under conditions of risk uncertainty and sensor anomaly.

[0062] 3. By encoding the construction risk range, radar reliability, and initial construction instructions into the initial and boundary condition uncertainties of the physical information neural network, and assimilating the internal temperature and humidity of the concrete and ground meteorological data, the rain damage status during the construction period can be continuously transmitted to the strength, temperature, and crack risk projection during the curing period, providing a consistent status basis for spraying, covering, temperature control, and formwork removal control. Attached Figure Description

[0063] 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:

[0064] Figure 1 This is an overall flowchart of the intelligent prevention and control method for rain damage in concrete construction according to the present invention.

[0065] Figure 2 This is a detailed flowchart of the simulation of intensity, temperature gradient, and crack risk range based on physical information neural network in S4. Detailed Implementation

[0066] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0067] refer to Figure 1 A smart method for preventing and controlling rain damage during concrete construction, comprising:

[0068] S1. Collect X-band dual-polarization radar, pouring surface condition, internal temperature and humidity of concrete, ground meteorological data, construction stage and sensor observation data, perform attenuation correction on radar data, interpolate the corrected polarization parameters to a Cartesian regular grid, and then perform raindrop spectrum distribution classification to generate rainfall event characteristics and rainfall event quality vector.

[0069] S2. Establish a pouring zone map, input rainfall event characteristics and pouring surface status data into the spatiotemporal model and perform dual time scale memory processing to generate the damage status and risk trend of each pouring zone.

[0070] S3. Determine the construction stage based on the construction stage data, and calculate the sensor residual based on the sensor observation data. Input the damage state, risk trend, construction stage and sensor residual into a neural sequence-preserving controller with monotonicity constraints to generate a construction risk interval consisting of the lower bound of construction risk, the center value of construction risk and the upper bound of construction risk.

[0071] S4. Generate initial construction instructions based on the construction risk interval, encode the construction risk interval, radar credibility represented by the quality vector of rainfall events, and initial construction instructions into the initial conditions and boundary condition uncertainties of the physical information neural network, assimilate the internal temperature and humidity of concrete and ground meteorological data, and generate the lower limit of strength, the upper limit of temperature gradient, and the crack risk interval.

[0072] S5. Input the construction risk range, lower limit of strength, upper limit of temperature gradient, and crack risk range into the range robust gating controller, and generate construction control instructions and maintenance control instructions according to the risk upper limit triggering and risk lower limit release rules.

[0073] In this specific embodiment, S1 includes:

[0074] In this embodiment, the data processing server at the construction site uses the pouring area number, sampling time, and equipment clock as primary keys to synchronously receive the horizontal reflectivity, differential reflectivity, differential propagation phase shift, and echo quality fields of the dual-polarization phased array radar. It also receives the pouring progress, surface texture, slurry coverage ratio, aggregate exposure ratio, and water depth uploaded by the pouring surface status terminal. Furthermore, it receives the depth location, temperature, relative humidity, and sensor status codes of the temperature and humidity probes inside the concrete. The server also reads ground rainfall, wind speed, construction stage, and sensor observation data from the weather station and construction logs. The server unifies all records to a 5s sampling time base of UTC+8. For fields missing for no more than one sampling period, it uses adjacent valid values ​​to linearly fill in the missing data. For fields that are continuously missing or have a fault status code, it writes an invalid flag and retains the original record. The filled-in values ​​are not used as independent observation samples for radar attenuation correction.

[0075] For each radar propagation path, the server reads the path length and the spatiotemporal registration results of ground rainfall, and corrects the reflectivity according to the path integral attenuation model, whereby the model is... ,in This represents the reflectivity after attenuation correction, in radar calibration reflectivity units. This represents the original radar reflectivity, expressed in radar calibration reflectivity units. This represents the path attenuation coefficient obtained by fitting paired samples of rain gauge and radar echo. The unit is the reciprocal of the path length. It can be found by searching the equipment number, frequency band and rain type version in the radar equipment calibration table. The value range is 0 to 0.08 reciprocal of the path length. This represents the propagation path length from the radar to the sampling voxel, in meters (m). When the path length or attenuation coefficient is missing, the valid calibration record of the same device within the last 30 minutes is used. If no valid record is found, the voxel is marked as unusable for subsequent interpolation. Path correction is also performed on differential reflectivity and differential propagation phase shift using the same propagation path: the differential reflectivity attenuation coefficient and propagation phase correction coefficient from the device calibration table are used respectively, and paired queries are performed using the same path length, device number, frequency band, and rain pattern version. The correction value is output only when the path length, coefficient, and input fields are all within the valid calibration range. If missing, out of range, or pairing fails, the valid correction record of the same device within the last 30 minutes is used. If no record is found, the channel is marked as invalid and the original value is retained. The correction residuals of differential reflectivity and differential propagation phase shift are written into the residual field and aggregated with the reflectivity correction residual in the S1 quality vector according to the order specified in the quality table.

[0076] The server establishes a regular grid of raindrop spectra on three-dimensional coordinates of latitude, longitude, distance, and altitude. The grid cells store corrected reflectance, differential reflectance, differential propagation phase shift, ground rainfall, raindrop diameter quantiles, and data integrity. Trilinear interpolation is performed on available voxels, with the interpolation weights determined by the normalized distances from the eight neighboring voxels to the target voxel. After interpolation, reflectance, differential reflectance, and propagation phase shift are restricted to the valid ranges given in the equipment calibration table. Values ​​outside the range are only retained as outlier fields and do not participate in raindrop spectrum classification.

[0077] The raindrop spectral classification module uses corrected reflectance, differential reflectance, differential propagation phase shift, rainfall, and wind speed as feature vectors. It employs a Gaussian mixture classifier fitted to a historical artificial rain pattern annotation set to label each grid cell as either small-drop dominant, convective large-drop dominant, or a mixed rain pattern, and calculates rainfall intensity and raindrop kinetic energy. Rainfall intensity is converted from the cumulative amount per unit time at ground rain gauges to the hourly rainfall intensity unit specified by the equipment. Raindrop kinetic energy is calculated from particle size, number concentration, and terminal velocity in the spectral analysis box; velocity is calculated using the particle size-velocity parameter from the equipment parameter table. The classifier output fields include category number, category posterior probability, rainfall intensity, kinetic energy, and classification version number. When the category posterior probability is below 0.60, the mixed rainfall pattern is retained and enters the low-confidence branch. The radar-corrected reflectivity, differential reflectivity, differential propagation phase shift, and raindrop spectral binning results are preferentially used to calculate rainfall intensity. The calculation rules are determined by the polarization parameter table and particle size spectrum calibration table corresponding to the rainfall pattern version. Ground rain gauges are only used to calibrate radar rainfall intensity, calculate consistency, and for backoff when radar features are invalid; they cannot replace valid radar rainfall intensity. The classifier outputs the rainfall intensity source field, calibration source field, and version number to close the input link from radar features to rainfall intensity.

[0078] To generate rainfall event characteristics, the server normalized the attenuation correction residuals, raindrop spectrum classification confidence, data spatiotemporal integrity, and ground rainfall consistency to [0,1], where the residual quality was determined by... calculate, This represents the current attenuation correction residual, expressed in radar calibration reflectivity units. This represents the 95th percentile residual of the same radar equipment and rain pattern calibration samples, with the same unit as the current residual. This means the result is restricted to 0 to 1; the other three items are each divided by the upper limit of the same dimension or the sample size benchmark in their respective mass calibration tables and then clipped to [0,1]; and then processed according to the mass vector formula. Calculate the mass components of the event mass vector, where For dimensionless event mass components; Indicates the quality of the attenuation correction residual; Indicates the confidence level for raindrop spectrum classification; Indicates temporal and spatial completeness; Indicates the uniformity of rainfall across the ground; , , , The dimensionless weights stored in the quality weight table have a sum of 1, with initial values ​​of 0.30, 0.25, 0.20, and 0.25 respectively, and are updated by subsequent valid event feedback; when any required field is invalid, the corresponding quality component is set to zero and the weights are renormalized according to the valid components; the above This only considers radar propagation distance as the independent variable, without involving force, torque, or length-to-force conversion. The residual quality rules described above are only applied when the 95th percentile residual is positive and finite. When the 95th percentile residual is invalid or not greater than zero, a conservative quality lower limit from the most recent valid quality calibration of the same equipment is used. If no calibration is available, the attenuation correction residual quality component is marked as invalid. Quality aggregation first verifies the four components and their corresponding weights, then re-normalizes the valid components. When all four components are invalid or the sum of valid weights is zero, the event quality component is marked as invalid, a preset conservative quality lower limit is used, and protective actions are triggered, prohibiting the use of zero-weight sums as valid quality. Q, each component, rollback source, and quality table version are all written into the event table.

[0079] The server writes rainfall event characteristics into an event table, with fields including event number, start and end time, affected grid set, rainfall type, peak rainfall intensity, raindrop kinetic energy, mass vector, set of anomalous fields, and version number. The mass vector is then fed back into the mass weight table: when the deviation between the radar data verified after construction and the ground rainfall is less than the preset error, the corresponding consistency weight is increased; when the deviation exceeds the error, the corresponding weight is decreased and the correction residual weight is increased. The event characteristics and mass vector in the event table serve as inputs to the S2 graph spatiotemporal model, and the corrected radar data continues to be used for attenuation correction and raindrop spectrum classification in the next sampling period.

[0080] In this specific embodiment, S2 includes:

[0081] The spatiotemporal processing module reads the rainfall event characteristics and pouring surface status records from the S1 event table. It divides the pouring surface into nodes with partition numbers according to the pouring sequence, planned pouring time, elevation, and adjacent boundaries. Node fields include geometric boundary, polygon area, planned start and end times, current pouring status, material batch, surface texture, slurry coverage ratio, aggregate exposure ratio, and water depth. Edges are established between adjacent nodes only when the material connection is continuous and the time interval does not exceed the maximum connection interval set in the construction log. The edge field stores the adjacent node number, boundary length, time interval, elevation difference, and material continuity flag. Duplicate edges are deduplicated using the joint key of boundary length and time interval.

[0082] For each node, a first-timescale memory unit and a second-timescale memory unit are set up. The first timescale is used to retain the rapid state caused by raindrop impact, and the second timescale is used to retain the slow state caused by continuous rainfall and surface water content. The two memory units are respectively configured according to... and Update, in which , Representing time respectively The fast and slow memory states; This represents the node input vector, which is spliced ​​and normalized from the characteristics of the S1 rainfall event, surface condition, and construction condition. , The attenuation factor is a dimensionless factor, retrieved from the memory parameter table by rain type and nodal material batch, and satisfies... In this embodiment, 0.65 and 0.92 are used; when a node has no effective rainfall observation within a sampling period, the previous memory state is maintained and the state staleness count is increased;

[0083] The first timescale unit accumulates the impact state from the S1 raindrop kinetic energy sequence, and the second timescale unit accumulates the wetting state from the rainfall duration and the water content of the pouring surface. The graph neighborhood aggregation module reads the two types of memories and edge fields of the first-order adjacent nodes for each node, uses the product of the reciprocal of the boundary length and the time interval decay as the aggregation weight, and writes the neighborhood state into the node neighborhood cache. When the material batches of adjacent nodes are not continuous or the time interval exceeds the connection interval, no state transfer is performed, but the isolation reason is recorded in the edge field.

[0084] The damage calculation module extracts the surface texture change rate, slurry coverage ratio, aggregate exposure ratio, and water depth from the neighborhood cache, and merges them with the dual time-scale memory into three damage components: the scour damage component uses the normalized product of raindrop kinetic energy and surface texture change rate, the slurry damage component uses the normalized product of slurry coverage ratio change and water depth, and the exposed stone damage component uses the normalized product of aggregate exposure ratio change and rapid memory state; each component is first divided by the reference upper limit of the same material batch in the damage calibration table, clipped to [0,1], and then written into the node damage field.

[0085] The continuous pouring time window consists of the concrete setting test records and construction logs of the same mix proportion. The module reads the initial setting time, transportation time, paving time and vibration time, calculates the actual time interval between adjacent zones and compares it with the continuous pouring time window; when the interval between adjacent zones is greater than the continuous pouring time window, a cold joint risk component is established. The risk component is calculated as the ratio of the excess amount to the allowable interval and is limited to [0,1]. When any time field is missing in the construction log, the most recent valid test record of the same mix proportion is used and the estimation source is marked.

[0086] The node damage state vector consists of four components: scour damage, slurry damage, exposed aggregate damage, and cold joint risk. The risk trend module performs robust linear regression on the state vector at continuous sampling times. A positive slope indicates an increase in risk, while a slope not greater than zero indicates that the risk is stable or decreasing. When the regression residual exceeds three times the median of the historical residuals, the trend update is frozen and the module waits for the next valid sampling point.

[0087] The spatiotemporal module writes the dual-timescale memory, neighborhood aggregation state, damage state vector, risk trend, node boundary, and version number of each node into the node status table of the casting zoning diagram. The node status table establishes a joint index according to the node number and sampling time. S3 reads the damage state, trend, and node construction stage fields in this table to determine the controller input. The node status table also stores the state staleness count and the pending review flag; the pending review flag is read by S3 and passed as an input quality field, and must not be lost during cross-module transmission.

[0088] During node state updates, the module first normalizes the raindrop kinetic energy according to the impact reference value of the same material batch, and then adjusts the impact state according to... Calculation, where It is a dimensionless impact state. For the kinetic energy of raindrops in the current time window, The impact reference value for the material batch is used, and both use the same kinetic energy unit. When the kinetic energy record is missing, the previous valid state is retained and the state staleness count is incremented by one. When the staleness count reaches three, the node is marked as pending verification. The impact reference value for the material batch must be a finite positive number. When the impact reference value for the material batch is missing, invalid, or not greater than zero, the normalization calculation is not performed. The most recent valid impact state of the same material batch is used first and marked as rollback. If there is no valid state, the impact state is marked as invalid and conservative protection is triggered.

[0089] The neighborhood aggregation module assigns lookup table weights to boundary length, time interval, and material continuity, with the edge weights following the rules. Calculation, where Let be the dimensionless edge weight from node i to node j. As a marker of material continuity, For the boundary length, For the same structural type, the boundary reference length For time intervals, To allow for connection time reference values; when the edge weight is below 0.05, the edge is deleted and the deletion reason is saved in the graph version record; the key fields of the memory parameter table are rain type, material batch, and node elevation, and the value fields are two attenuation factors, update time, and valid range. The table version is updated with the material test results. When no match is found in the table, the latest version of the same strength grade is used and the rollback source is marked; the material continuity flag can only be 0 or 1, indicating that the material is discontinuous or continuous, respectively; the boundary reference length and connection time reference value must be finite positive numbers, the boundary length must be positive, and the time interval cannot be negative. When any parameter is invalid, the edge is deleted and the reason is recorded. Invalid parameters cannot be used to generate edge weights; when rolling back from the table, the edge weight version and rollback flag are written simultaneously.

[0090] Material batch calibration was performed before writing the scour, laitance, and exposed aggregate components into the damage record. The damage components were calculated according to... Calculation, where Represents the dimensionless component of the k-th type of damage. For the current observation features, and These are the baseline value and the 95th percentile value of the calibration samples in the same batch, respectively; when the denominator is less than the minimum effective difference in the calibration table, the minimum effective difference is used and written into the calibration anomaly flag;

[0091] The risk of cold joints is assessed using a rule that addresses exceeding the limits of the continuous pouring time window and the actual interval. ,in Risk of dimensionless cold seams The time interval between adjacent partitions. The continuous pouring time window is determined by the initial setting test, transportation, paving, and vibration records; when The cold joint component is recorded as zero. When the time window is exceeded, the excess quantity is written into the node damage status and used for S3 risk interval calculation. The continuous pouring time window must be a finite positive number and the source record must be valid. When the continuous pouring time window is missing, zero, or cannot form a valid time window, the cold joint risk component is marked as invalid and it is prohibited to send it into risk calculation. The node retains the previous protection action and records the time window abnormality until a valid test or log record is obtained.

[0092] In this specific embodiment, S3 includes:

[0093] The risk interval control module reads the current pouring, curing, formwork removal, or work stoppage and joint leaving stage from the construction log, and converts the construction stage into a stage code using the partition number and sampling time in the node status table as the key. Simultaneously, it reads the damage state vector, risk trend, and sensor observation data from S2, and calculates the residual for each sensor based on the median of the nearest reference sensors at the same sampling time. The sensor residual is defined as... ,in Indicates the first Each sensor at time The residuals are in the same units as the sensor measurements; This represents the sensor's raw observations; S1 represents the reference observation set of the same node, the same measurement type, and in normal condition; when the reference set is empty, the effective median of the sensor in the last five minutes is used; if it is still empty, the residual is marked as invalid and the sensor abnormality flag is triggered; S3 reads the state stale count and the pending verification flag of S2 at the same time; when the node is pending verification, the input quality is reduced to the conservative level, the generation of release type output is prohibited, the protection type risk interval is output first or wait for at least one effective sampling window, and the flag is written into the risk interval index and passed to S5.

[0094] The controller takes damage state, risk trend, construction stage, and sensor residuals as input vectors. It uses a monotonic constrained neural differential equation controller trained with damage-labeled samples and construction action result samples. The controller outputs a lower bound, a risk center value, and an upper bound for risk. An application is applied to the output of the risk interval. Constraints, among which Indicates the lower limit of construction risk. Indicates the construction risk center value. The upper bound of construction risk is represented by three variables, all of which are dimensionless risk values ​​and normalized to [0,1] according to the risk calibration set. During training, non-negative monotonic weights are set for damage memory, trend, and sensor residuals so that the corresponding risk output does not decrease when any input increases. The weights are stored in the controller version file and updated with labeled samples. Monotonic constraints cover all risk-related input channels: non-negative monotonic weights or corresponding penalties are applied to the components representing increased impact, water content, or other risks in fast and slow memory, each component of damage state, positive risk trend, and absolute value of sensor residuals. If a memory channel uses reverse protection coding, it is first converted to a risk vector according to the coding definition before constraints are applied. Stage embedding, missing mask, and version identifier do not participate in monotonicity comparison, but their invalid states are only allowed to tighten the output and must not reduce the protection level.

[0095] The physical state encoding of the controller includes fast memory, slow memory, damage state, trend, and residual. The encoder is first standardized according to the mean and standard deviation of the material batch at the node, and then the risk state trajectory is obtained by the hidden layer integration with monotonic constraints. The construction stage encoding adopts independent embedding vectors and sets stage switching gates. The risk output of the pouring stage cannot directly reuse the threshold of the curing stage. When the controller input contains invalid residuals, the mask averaging of the valid sensors is used and the invalid number is written into the input quality field. If the number of valid sensors is lower than the preset number, the output interval is marked as pending review.

[0096] To ensure the interval width covers the error quantiles of the calibration samples, the calibration module retains an interval calibration set that is not used in training for each material batch, and calculates the coverage of the risk prediction interval to the actual action results. When the coverage is lower than the target coverage, the interval width is increased according to the calibration error quantiles. The increase only affects the upper and lower bounds and does not change the center value. The adjusted calibration version number is written to the controller configuration. When the coverage reaches the target value, the interval parameters for this batch are frozen until the coverage decreases for two consecutive sampling windows.

[0097] The risk interval record output by the controller includes node number, stage code, lower bound, center value, upper bound, input quality, sensor residual set, check flag, calibration flag, model version, and calibration version. The previous cycle interval is retained in the state cache for use by hysteresis control. This record is passed to S4 as the initial construction instruction code, boundary condition uncertainty, and interval constraint input for the physical information neural network. When sensor residuals are invalid, the module prioritizes finding alternative sensors for the same node and measurement type, and then uses the most recent five-minute valid reference set to re-obtain the median residual. While alternative references remain unavailable, the residuals remain invalid, and the input quality is reduced to a conservative level. S5 prohibits deprotection and records the manual check task.

[0098] The center value of the risk interval is obtained by integrating the controller state, and the interval width is determined according to... Calculation, where Indicates the width of the dimensionless risk interval. and These are the upper and lower bounds of the risk, respectively; when the width exceeds the 95th percentile of the calibration set width, the controller will mark the input quality as low and prioritize reading the new sensor residual in the next cycle;

[0099] The controller performs inference and interval calibration every 5 seconds. Training samples are extracted stratified by node number, rain type, and construction stage. Action result samples are obtained from the S5 action log. The training loss consists of interval coverage error, center value error, and monotonicity violation penalty. The monotonicity penalty is uniformly applied to the risk-in-the-direction input channels corresponding to fast memory, slow memory, damage state, trend, and sensor residual, ensuring that the risk lower bound, center value, and upper bound do not decrease when any channel increases. Stage embedding, missing mask, and version identifier do not participate in monotonicity comparison, and their invalid states are only allowed to tighten the output.

[0100] When the interval calibration sample is less than the minimum number of a material batch, the controller does not update the batch parameters, uses the previous version, and outputs a calibration flag; when the calibration sample is sufficient but the coverage is still lower than the target, the boundary quantile is increased first and the interval is recalculated. If the target is not met after two consecutive attempts, the center value and controller version are frozen to prevent unverified intervals from entering S4.

[0101] The controller associates the lower bound, center value, and upper bound of risk with the set of actions allowed in the construction phase, respectively. The action set is recorded in the phase action table and includes the action code, allowed throughput range, allowed interval range, and release condition. S4 reads the action set and interval width and writes the initial construction instructions and uncertainties of each node into the physical simulation input cache.

[0102] To construct training and calibration samples, the module establishes a hierarchical index based on node number, rain type, construction stage, and effective number of sensors. Inputs are extracted from the damage annotation sample set, and corresponding results are extracted from the action log. Observations after the action occurs are removed according to their chronological order to prevent control results from leaking to the input. The sample record saves the input vector, interval label, action code, sampling time, node number, and data version. If the index is missing, it is not included in the training.

[0103] The output constraints of the interval controller are executed in each inference cycle. If the lower bound obtained by the solution is greater than the center value or the center value is greater than the upper bound, the interval is corrected according to the boundary and written into the constraint correction flag. If the interval still violates the allowed range of the stage action table after correction, the interval of the previous cycle is frozen and the node is transferred to the queue for review. The review result is fed back by the physical deduction of S4.

[0104] The controller writes the interval, stage code, and input quality of each node into the risk interval index. The index key is the node number and the sampling time, and the value fields are the three risk boundaries, trend slope, input quality, pending review flag, pending calibration flag, calibration version, and abnormal flag. S4 reads the complete record of the same time according to the index. If the record is missing, the boundaries of different sampling times must not be spliced.

[0105] In this specific embodiment, S4 includes:

[0106] The physical information inference module reads the S3 risk interval and construction stage. First, it maps the risk interval to the initial construction instructions of pouring flux, covering heat exchange boundary, or construction interval according to the construction action table. Then, it extracts the radar confidence, attenuation correction residual, raindrop spectrum classification confidence, data spatiotemporal integrity, and ground rainfall consistency from the S1 event quality vector as boundary condition uncertainties. The two thresholds are determined by the allowable scour test results and allowable interval time of concrete with the same mix proportion. The emergency cover threshold is less than the work stoppage and gap leaving threshold. The threshold record includes material batch, test number, unit, expiration date, and version number. The radar confidence is uniformly taken as the aggregated result of the S1 event quality Q: first, the validity is verified in the order of attenuation correction residual quality, raindrop spectrum classification confidence, spatiotemporal integrity, and ground rainfall consistency. Then, the valid components are normalized and weighted according to the quality weight table. The output fields include radar confidence, four quality components, weight, anomaly field, rollback source, and quality version. If any component is invalid, set it to zero and renormalize it; if all components are invalid, use the most recent valid Q or adopt a conservative lower limit and mark the confidence level as a fallback. Invalid values ​​should not be treated as high confidence levels.

[0107] The module encodes the initial construction instructions into the initial conditions of the physical information neural network, and maps the risk interval width, radar confidence and sensor residuals into the initial conditions and boundary condition uncertainties. The interval width mapping uses a piecewise linear rule: when the interval width is less than the first quantile of the calibration table, a normal boundary is used; when the width is between the first and second quantiles, the boundary perturbation is linearly increased; when the width is greater than the second quantile, the boundary update of the affected node is frozen and the most recent valid period is used.

[0108] The physical information neural network simultaneously extrapolates the strength field, temperature field, and humidity field on the node grid of the casting zoning map. The network input includes node coordinates, material batch, casting age, internal temperature and humidity, ground meteorological data, initial construction instructions, and uncertainty coding. The network output is the strength prediction, temperature, moisture content, and quantile boundary of each node. The training objective consists of strength test samples, temperature probe samples, humidity probe samples, and constraints on hydration reaction, heat conduction, and moisture migration. The constraint residuals are weighted according to the material batch weight table within each time window.

[0109] During the simulation, the temperature and humidity fields are updated step by step according to the node boundary conditions. The rainfall input at the boundary uses S1 rainfall intensity and raindrop kinetic energy, while the internal nodes use the heat conduction and moisture migration flux of the adjacent nodes. When the temperature or humidity observation exceeds the sensor range, the physical constraint prediction value is used and the range anomaly is recorded. The anomaly value is not used to update the probe bias. When the simulation residual exceeds the 95th percentile of the calibration residual of the same batch, the output confidence of the node is reduced and it is included in the review queue.

[0110] The lower bound of strength is taken as the preset quantile boundary of the strength prediction distribution, the upper bound of temperature gradient is taken as the temperature difference between adjacent nodes divided by the center distance of the nodes, and the crack risk range is obtained by comparing temperature stress with tensile strength. The comparison results are used to calculate the dimensionless crack risk upper and lower bounds according to the temperature gradient, humidity gradient, lower bound of strength and allowable tensile stress table of materials. The larger the upper bound of risk, the higher the crack risk. S4 only outputs the crack risk range with version number and valid flag, and does not directly execute maintenance triggering or cancellation. Maintenance triggering is done by S5 comparing the crack risk upper bound with the trigger threshold, and maintenance cancellation is done by S5 comparing the crack risk lower bound with the cancellation threshold.

[0111] The physical information simulation results are written into the node physical state table, with fields including lower bound of intensity, upper bound of temperature gradient, crack risk interval, temperature field version, humidity field version, boundary uncertainty, simulation residual, and anomaly flag. The module also generates initial condition interval records, passing the pouring flux boundary, covering heat transfer boundary, construction interval boundary, and their uncertainties to the S5 interval robust gating device. S4 simultaneously writes the maintenance gating status field, including interval validity flag, pending verification flag, pending calibration flag, version number, and anomaly reason. The maintenance gating status field containing the pending calibration flag is then passed to S5 as is according to the node number and sampling time. S5 executes the construction and maintenance state machine uniformly according to this field and risk interval, with the status field having higher priority than ordinary triggering or deactivation judgment.

[0112] Strength derivation uses the strength lower bound formula ,in This represents the lower bound of the node strength, expressed in strength calibration units. This is the set of intensity samples output by the network. Indicates taking the 5th percentile; upper bound of the temperature gradient is based on... Calculation, where This is the upper bound of the temperature gradient. , Temperature of adjacent nodes The node center distance, along with all three parameters, are provided by the node physical state table. The strength prediction sample set must be non-empty and meet the minimum sample size specified for the material batch. The samples must also pass validity and version consistency checks. If the sample set is insufficient or empty, the strength quantile boundary of the most recent valid batch will be used and a fallback will be marked. If no usable boundary is available, the lower bound of the strength will be marked invalid and a protection-type risk interval will be output. The upper bound of the temperature gradient will only take the maximum value on all valid and adjacent node pairs with positive node center distances. If there are no valid node pairs, the upper bound of the temperature gradient will be marked invalid and a conservative temperature boundary will be used.

[0113] The crack risk zone is determined by temperature stress, humidity gradient, and tensile strength, with the risk center based on... Calculation, where The dimensionless crack risk center, For temperature stress, Tensile strength at the current age; when When the risk center is zero, physical constraint predictions are used and the upper and lower bounds are expanded when temperature or humidity observations are abnormal; the tensile strength at the current age must be a finite positive value; when the prediction is invalid, missing, or not greater than zero, the lower bound of physical constraints is preferred and must still be positive; if a positive lower bound cannot be obtained, the crack risk is marked as invalid and the removal of maintenance protection is prohibited.

[0114] The boundary parameters for interval extrapolation are established into boundary tables according to three categories: construction flux, overlay heat exchange, and intermittent. The tables store node type, material batch, boundary field, unit, effective range, calibration sample number, and version number. When no matching record is found in the table, a conservative boundary of the same strength level is used, and the boundary source is marked as a backtracking record. Backtracking records are not used for subsequent parameter fitting.

[0115] The network inference cycle is consistent with the S1 sampling cycle of 5 seconds. At the end of each time window, the inference residual is compared with the 95th percentile of the material batch calibration residual. Nodes that exceed the threshold are entered into the verification queue and temporarily maintain the boundary of the previous cycle. When a node in the verification queue obtains two consecutive valid temperature and humidity observations, the inference is rerun and the node's physical state is replaced.

[0116] When selecting the lower bound of intensity, the upper bound of temperature gradient, and the crack risk range, the module sorts the multi-group dimensional boundaries from largest to smallest according to the upper bound of risk. If the upper bounds of risk are the same, the processing order is determined by the node number in ascending order. If any boundary is missing, no release control quantity is generated, and only the protection initial instruction is output to avoid the selection result lacking executable boundaries.

[0117] Each output of the inference network simultaneously saves the initial condition version, boundary condition version, material batch, and inference time window. During training, a validation time window that does not participate in fitting is used to evaluate temperature, humidity, and intensity errors. When the validation error exceeds the allowable error of the material batch, the previous model version is retained and only the observation weights are updated. The updated weights need to be reconfirmed in the next validation window.

[0118] The physical constraint solution is performed in the order of residuals of hydration reaction, heat conduction and water migration. If any residual exceeds the corresponding constraint threshold, the inference time step is shortened first. If convergence is still not achieved, the strength and temperature boundaries of the node are frozen and the protection risk range is output. After convergence, the quantile boundary and residual are written into the node physical state table and passed to S5.

[0119] The attenuation correction residual of S1 is converted into a residual confidence component that does not increase with the increase of the residual, and the original residual and normalized value of this component are retained in the physical deduction input. The confidence component uses the 95th percentile residual of S1 of the same device as the denominator. If it exceeds the denominator, it is directly truncated to zero. If it does not exceed the denominator, it is calculated according to the reverse ratio to ensure that the increase of the residual does not increase the radar confidence. When the radar confidence is invalid, the difference between it and the radar confidence is used as the confidence complement, and the source of the complement is written into the boundary condition uncertainty record. The aforementioned confidence complement is only applicable to numerical transformations of Q that have passed the validity check. If Q itself is invalid, the transformation result of subtracting the confidence is not used. Instead, the most recent valid Q or the conservative lower limit is used, and the fallback source and version are written into the boundary condition uncertainty record.

[0120] In this specific embodiment, S5 includes:

[0121] The interval robust gating controller reads the S3 risk interval and the S4 node physical status table, and establishes construction control, maintenance control and current action status registers according to the node number. The gating controller maps the upper limit of construction risk to the construction action trigger quantity, the upper limit of crack risk to the maintenance action trigger quantity, and the lower limit of crack risk to the maintenance action release quantity. It also saves the action code, action node, start and end time, minimum holding time, trigger threshold, release threshold and data version in the action table. When the upper limit of risk reaches the emergency cover threshold and is lower than the stop-work gap threshold, an emergency cover command is output. When the stop-work gap threshold is reached, a stop-work gap command is output. Otherwise, the operation is maintained or a continue construction command is generated. Both thresholds are retrieved from the S4 mix proportion test record and the allowable interval time table.

[0122] The maintenance gating system maps crack risk zones to adjustments in spray intervals, coverage conditions, and insulation thickness. When the upper limit of crack risk reaches the maintenance trigger threshold determined by allowable water loss and allowable temperature stress, the spray frequency or insulation thickness is increased. When the lower limit of crack risk is not higher than the unified maintenance release threshold and continues to reach the minimum holding time, the system reverts to the pre-adjustment state. The trigger threshold and hysteresis width are stored in the material batch maintenance parameter table, which includes the unit, test source, version, and effective time. For batches not matching, a conservative record of the same strength grade is used. Write the calibration flag; maintenance release is uniformly determined by the lower limit of crack risk: the state before adjustment is allowed only when the lower limit of crack risk is not higher than the unified maintenance release threshold (trigger threshold minus hysteresis width), the minimum holding time is reached, and the check flag is false; the upper limit of crack risk is only used to trigger maintenance protection; the above maintenance release determination must also simultaneously satisfy the interval valid flag being true, the input quality reaching the valid threshold, the check flag being false, the calibration flag being false, and the sensor unlock being false. If any condition is not met, the current protection action is maintained and release is prohibited.

[0123] The gating controller converts the risk interval width into initial condition interval parameters and establishes upper and lower limits for the construction flux boundary, the heat exchange coverage boundary, and the construction intermittent boundary, respectively. The boundary uncertainty is recorded using S4 intervals. If any input confidence is invalid or the event quality vector transmission verification fails, the corresponding confidence value is set to 1 and the release action is stopped. Only the current protection action is allowed to continue.

[0124] The gating controller uses the current action, lower limit of the risk range, upper limit of the risk range, effective number of sensors, and action timer to form a state vector. It executes action switching according to a state machine with hysteresis: the trigger condition is determined by the upper limit reaching the trigger threshold; the construction action release condition is determined by the lower limit of construction risk not exceeding the construction release threshold; the maintenance action release condition is determined by the lower limit of crack risk not exceeding the unified maintenance release threshold. The release threshold equals the trigger threshold minus the preset hysteresis width in the historical action switching record. Even after the action reaches the minimum hold time, it is only released when the release condition is met, avoiding frequent switching caused by short-term rainfall drops. The release condition of the maintenance state machine is consistent with the above rules, uniformly comparing the lower limit of crack risk with the release threshold, and simultaneously checking the hysteresis width, minimum hold time, input quality, and pending verification flag. The above maintenance release determination must also simultaneously satisfy the following conditions: the interval valid flag is true, the input quality reaches the valid threshold, the pending verification flag is false, the pending calibration flag is false, and the sensor unlock is false. If any condition is not met, the current protection action is maintained and release is prohibited.

[0125] When the sensor residual reaches the failure threshold or the number of valid sensors falls below the preset number, the sensor anomaly branch is entered. This branch takes precedence over normal triggering, normal release, and action conflict handling, and maintains the abnormal state. Releasing protection actions based on the lower limit of the risk interval is prohibited. Only when the residuals of all abnormal sensors are below the recovery threshold, the number of valid sensors reaches the preset number, and continuous valid sampling reaches the recovery stabilization time, will the sensor recovery branch be entered and the risk interval be recalculated. When the reference set is empty or the residual is invalid, the system will first switch to the alternative sensor or the valid reference set of the most recent five minutes according to the same node and the same measurement type. Only after obtaining a valid residual again will the recovery judgment be allowed. If it remains invalid, the current protection action will be maintained, the release command will be prohibited, the boundary will be limited to a conservative range, and an abnormal task requiring manual review will be generated.

[0126] When there is a conflict between actions such as stopping work to leave gaps, emergency covering, maintaining coverage, and delaying formwork removal, the gate controller will execute according to the priority of stopping work to leave gaps, emergency covering, and continuing construction. When the formwork removal action conflicts with other actions, the delayed formwork removal will be placed with the highest priority. Each action switch records the action code, trigger and release time, risk range, radar confidence level, sensor residual, action result, and data version. The action result consists of the texture of the subsequent poured surface, water depth, internal temperature and humidity, and strength test results.

[0127] After the sensor recovery conditions are met, the gate controller uses the recovered sensor data to recalculate the S3 risk range and rerun the S4 physical information neural network. It updates the protection action based on the new lower limit of intensity, upper limit of temperature gradient, and crack risk range. The recovery stabilization time is determined by the product of the sampling interval and the preset number of consecutive effective samplings. The updated action must still meet the minimum holding time and release threshold. The holding action will not be skipped due to a single recalculation.

[0128] The system establishes independent validation sample sets for construction risk zones and crack risk zones respectively. The construction risk validation sample set does not overlap with the training and calibration samples of the neural sequence-preserving controller, and the crack risk validation sample set does not overlap with the parameter fitting samples of the physical information neural network. Each trigger, release, protection action switch, and sensor recovery is written to a versioned action log. The system calculates the zone coverage rate based on the validation sample set. If any updated zone coverage rate is less than the corresponding target coverage rate, the controller calibration sample, physical information network observation weights, and parameters are rolled back to the same version before the update to avoid further relaxing the control boundary when the target coverage rate has not been reached. The action log and the updated model version are available for the next construction cycle and serve as the feedback basis for S1 quality weights and S2 damage trends, forming a closed loop of rain damage observation, risk zone, physical deduction, and construction and maintenance control.

[0129] Construction control quantity according to Mapping, where Construction control instructions. This is the risk range. The set of actions for each stage is mapped. Actions that do not meet the throughput and intermittent constraints are first filtered, and then actions are selected in descending order of risk upper bound. If the risk upper bounds are the same, the current action is retained to avoid meaningless switching. If the set of actions for the construction stage is empty after filtering, the current effective protection action is retained. If the current action is invalid or does not exist, the preset stop-work gap, emergency cover or other conservative protection instructions are output, and the reason for no action is recorded. The output of the continue construction or release instruction is prohibited.

[0130] Maintenance control amount according to Mapping, where For maintenance control instructions, and These are the lower and upper bounds for crack risk. This is the upper bound of the temperature gradient. The set of maintenance actions is defined as follows: when the upper bound reaches the trigger threshold, the covering and spraying actions are selected first; the maintenance is only allowed to be released when the lower bound of crack risk is not higher than the unified maintenance release threshold and the minimum holding time is met; when the set of maintenance actions is empty, the input is invalid, or the upper and lower bounds of maintenance are unavailable, the current maintenance protection action is maintained; if the current action is invalid, a preset covering, spraying, or insulation protection command is output, the output of maintenance release command is prohibited, and the abnormality and rollback source are recorded; the above maintenance release judgment must also meet the following conditions simultaneously: the interval valid flag is true, the input quality reaches the valid threshold, the pending verification flag is false, the pending calibration flag is false, and the sensor unlock is false. If any condition is not met, the current protection action is maintained and release is prohibited.

[0131] Action conflict handling adopts fixed priority and state lock. The state lock fields include the current action, lock start time, minimum hold time, release threshold and data version. During the lock period, even if the trigger condition of another action is met, only the candidate action field is updated and the currently executed action is not rewritten. After the lock ends, the action is reselected according to the priority of stop work and leave gap, emergency cover, and continue construction.

[0132] Trigger quantity of construction instructions calculate, For construction trigger margin, This is the upper limit of construction risk. The trigger threshold for the corresponding action; maintenance release margin according to calculate, To maintain the trigger threshold, The lower bound of crack risk is set; the trigger branch is entered when the trigger margin is not less than zero, and the release branch is entered only when the release margin is not less than the hysteresis width and the minimum holding time is reached; the formulaic release branch is only allowed to be executed when the interval valid flag is true, the input quality reaches the valid threshold, the check flag is false, the calibration flag is false and the sensor unlock is false. If any condition is not met, the current protection action is maintained.

[0133] When the abnormal branch of the sensor persists, the system sets the unlock in the status register to true and limits the construction flux, overlay heat exchange and intermittent boundary to a conservative range of protection actions; after the branch is unlocked, the risk interval is recalculated using the recovery data, and then the switching is performed according to the minimum hold time, action priority and executable constraints of the candidate actions.

[0134] Action logs are indexed by version number, node number, and start and end times of actions. The log fields also store trigger margin, release margin, risk upper and lower bounds, radar quality vector, sensor residuals, and execution results. The S1 quality weight update and S2 trend regression for the next construction cycle only read completed and verified log records. Incomplete or rolled-back versions are not included in the feedback samples.

[0135] The interval width and hysteresis parameter undergo a consistency check before each action switch, with the width calculated according to... , Residual release amount according to Minimum hold time according to Calculation, where The width of the risk range. The difference between the trigger and release thresholds. To maintain the duration, For the number of consecutive valid samples, The sampling interval is specified. The three results are compared with the allowable range in the action table. If the allowable range is not met, the current protection action is retained and an abnormal parameter flag is written.

[0136] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0137] This invention uses the mass vector of rainfall events to control the reliability of radar data, uses dual time-scale memory to characterize raindrop impact and continuous wetting, and converts pouring zone damage, construction stages and sensor residuals into construction risk intervals, so that rainfall observation, surface damage and construction actions can enter subsequent physical deduction and interval gating along a unified data link.

[0138] This invention changes the previously separate processing method of construction rain damage control and maintenance simulation by using monotonicity constraints, risk interval uncertainty coding, and gating structures with upper bound triggering and lower bound release. It also maintains the consistency of control state and model version by prioritizing sensor anomalies, recalculating after recovery, and independent verification rollback, thereby achieving collaborative prevention and control of risks during the construction and maintenance periods.

Claims

1. A method for intelligent prevention and control of rain damage in concrete construction, characterized in that, include: S1. Collect X-band dual-polarization radar, pouring surface condition, internal temperature and humidity of concrete, ground meteorological data, construction stage and sensor observation data, perform attenuation correction on radar data, interpolate the corrected polarization parameters to a Cartesian regular grid, and then perform raindrop spectrum distribution classification to generate rainfall event characteristics and rainfall event quality vector. S2. Establish a pouring zone map, input rainfall event characteristics and pouring surface status data into the spatiotemporal model and perform dual time scale memory processing to generate the damage status and risk trend of each pouring zone. S3. Determine the construction stage based on the construction stage data, and calculate the sensor residual based on the sensor observation data. Input the damage state, risk trend, construction stage and sensor residual into a neural sequence-preserving controller with monotonicity constraints to generate a construction risk interval consisting of the lower bound of construction risk, the center value of construction risk and the upper bound of construction risk. S4. Generate initial construction instructions based on the construction risk interval, encode the construction risk interval, radar credibility represented by the quality vector of rainfall events, and initial construction instructions into the initial conditions and boundary condition uncertainties of the physical information neural network, assimilate the internal temperature and humidity of concrete and ground meteorological data, and generate the lower limit of strength, the upper limit of temperature gradient, and the crack risk interval. S5. Input the construction risk range, lower limit of strength, upper limit of temperature gradient, and crack risk range into the range robust gating controller, and generate construction control instructions and maintenance control instructions according to the risk upper limit triggering and risk lower limit release rules.

2. The intelligent prevention and control method for rain damage in concrete construction according to claim 1, characterized in that, S1 includes: The path integral attenuation correction amount is determined based on the spatiotemporal registration results of ground rainfall and radar reflectivity, and the differential reflectivity and differential propagation phase shift are corrected using the path integral attenuation correction amount. The raindrop spectrum distribution type is identified and the rain intensity and raindrop kinetic energy are calculated based on the interpolated corrected polarization parameters. The rainfall event quality vector is generated based on the attenuation correction residual, raindrop spectrum distribution classification confidence, data spatiotemporal integrity, and ground rainfall consistency. The rainfall event quality vector is then fed back to update the attenuation correction weight and raindrop spectrum distribution classification weight for subsequent radar data.

3. The intelligent prevention and control method for rain damage in concrete construction according to claim 1, characterized in that, S2 includes: dividing the pouring surface into nodes according to the pouring sequence, pouring time, elevation, and adjacent boundaries; constructing edges based on the material continuity relationship and time interval between adjacent nodes to form the pouring zoning map; setting a first time-scale memory unit and a second time-scale memory unit for each node, wherein the decay period of the first time-scale is less than the decay period of the second time-scale; using the first time-scale memory unit to accumulate the impact state caused by the raindrop kinetic energy extracted from the rainfall event features, and using the second time-scale memory unit to accumulate the rainfall duration extracted from the rainfall event features and the surface water content extracted from the pouring surface state data, and fusing them to obtain the dual time-scale memory state.

4. The intelligent prevention and control method for rain damage in concrete construction according to claim 3, characterized in that, Step S2 further includes: extracting surface texture changes, slurry coverage ratio, aggregate exposure ratio, and water depth from the pouring surface state data; calculating the time interval between adjacent pouring zones based on the pouring time used to divide the nodes; performing graph neighborhood aggregation on the surface texture changes, slurry coverage ratio, aggregate exposure ratio, and water depth with the dual-timescale memory state to generate scour damage components, laitance damage components, and exposed aggregate damage components; determining a continuous pouring time window based on the difference between the initial setting time obtained from the concrete setting test with the same mix proportion and the transportation, paving, and vibration time in the construction record; comparing the time interval between adjacent pouring zones with the continuous pouring time window to generate a cold joint risk component; and constructing the damage state from each damage component and the cold joint risk component, and generating the risk trend based on the damage state at continuous sampling times.

5. The intelligent prevention and control method for rain damage in concrete construction according to claim 4, characterized in that, S3 includes: training the neural order-preserving controller with damage annotation samples and construction action result samples, and constraining the lower bound of construction risk to be no greater than the center value of construction risk and the center value of construction risk to be no greater than the upper bound of construction risk; setting non-negative monotonic weights for the dual-timescale memory state, the damage state and the sensor residual, so that the lower bound of construction risk, the center value of construction risk and the upper bound of construction risk do not decrease when the corresponding input increases; and correcting the width of the construction risk interval according to the interval uncovered error quantile of the calibration sample, so that the coverage rate corresponding to the interval uncovered error quantile reaches the preset coverage rate.

6. The intelligent prevention and control method for rain damage in concrete construction according to claim 5, characterized in that, S4 includes: setting an emergency cover threshold and a work stoppage / gap threshold, wherein the emergency cover threshold is less than the work stoppage / gap threshold; generating an initial construction instruction to continue construction when the construction risk center value is less than the emergency cover threshold, generating an initial construction instruction to cover the emergency when the construction risk center value reaches the emergency cover threshold but is less than the work stoppage / gap threshold, and generating an initial construction instruction to stop construction and leave a gap when the construction risk center value reaches the work stoppage / gap threshold, wherein the two thresholds are determined by the allowable scour test results and allowable interval time of the corresponding concrete mix proportion; mapping the width of the construction risk interval to an initial condition parameter interval; and reading the attenuation correction residual, raindrop spectrum distribution classification confidence, data spatiotemporal integrity, and other parameters from the rainfall event quality vector. The ground rainfall consistency is determined by converting the attenuation correction residual into a residual confidence component that does not increase with the increase of the residual, and normalizing each confidence component to a closed interval of zero to one. The normalized confidence components are arranged in ascending order of value, and the confidence component value at the top of the sort is used as the radar confidence. The difference between one and the radar confidence is used as the confidence complement value, and the product of the confidence complement value and the preset upper limit of rainfall boundary disturbance is determined as the rainfall boundary disturbance range. When any confidence component used in step S4 is invalid or the transmission verification of the rainfall event quality vector fails, the confidence complement value is set to one. The initial construction command is converted into a pouring flux boundary, a covering heat exchange boundary, or a construction interval boundary, forming the initial condition uncertainty and the boundary condition uncertainty.

7. The intelligent prevention and control method for rain damage in concrete construction according to claim 6, characterized in that, Step S4 further includes: setting hydration reaction, heat conduction, and moisture migration constraints in the physical information neural network; constructing parameter value intervals based on the initial condition uncertainty and the boundary condition uncertainty; correcting the parameter value intervals using the observation residuals of the internal temperature and humidity data of the concrete and the ground meteorological data; and extrapolating the concrete strength, temperature field, and humidity field within the parameter value intervals; generating the lower bound of strength and the upper bound of temperature gradient based on the quantile boundaries of each extrapolation result; and generating the crack risk interval based on the comparison between temperature stress and tensile strength that varies with strength.

8. The intelligent prevention and control method for rain damage in concrete construction according to claim 7, characterized in that, S5 includes: mapping the construction risk zone to construction control, and mapping the crack risk zone to maintenance control; triggering an emergency covering construction control command when the upper boundary of the construction risk reaches the emergency covering threshold, triggering a work stoppage and crack retention construction control command when the upper boundary of the construction risk reaches the work stoppage and crack retention threshold, and triggering a maintenance control command and starting a holding timer when the upper boundary of the crack risk zone reaches a maintenance trigger threshold jointly determined by allowable water loss and allowable temperature stress; the minimum holding time is set according to the construction control action and the maintenance control action respectively, and is recorded in the corresponding action history execution record. The duration of hold is determined by a preset quantile; after the minimum hold duration is reached, the control command is released only when the lower bound of the corresponding risk interval is less than the corresponding action release threshold. The action release threshold is determined by subtracting the hysteresis width from the corresponding action trigger threshold. The hysteresis width is determined by a preset quantile of the difference between adjacent trigger values ​​and release values ​​in the historical action switching record; a recovery threshold is set to be less than the failure threshold. When the sensor residual reaches the failure threshold or the number of effective sensors is less than a preset number of effective sensors, a sensor anomaly branch is entered. The sensor anomaly branch takes precedence over normal triggering, normal release, and action conflict handling; in the sensor The abnormal branch remains in an abnormal state and is prohibited from releasing the protection action based on the lower bound of the risk interval. The sensor recovery branch serves as the exit condition for the abnormal sensor branch, and exits the abnormal state only when all conditions of the sensor recovery branch are met. If the current action belongs to one of the following: work stoppage with gap, emergency covering, maintaining coverage, or delayed formwork removal, the current action is maintained; otherwise, the preset protection action is switched to. When multiple construction actions conflict, the priority is determined according to the order of work stoppage with gap, emergency covering, and continuing construction. When formwork removal actions conflict, delayed formwork removal takes priority. The protection action is only effective when the sensor residual is less than the recovery threshold and effective sensing is achieved. When the number of sensors reaches the preset effective number, all failure conditions are no longer met, and the system continues to recover and stabilize for a certain period of time, the system enters the sensor recovery branch. The construction risk range is recalculated using the recovered sensor data, and the physical information neural network is rerun. Based on the quantile boundaries of each rerunned simulation result, updated lower bounds for intensity and upper bounds for temperature gradient are generated, and updated crack risk ranges are generated. Then, the maintained protection action or the preset protection action is processed according to each updated result and the corresponding action release threshold. The recovery and stabilization time is determined by the product of the data sampling interval and the preset number of consecutive effective samplings.

9. A method for intelligent prevention and control of rain damage in concrete construction according to claim 8, characterized in that, Step S5 further includes: delaying demolding when the lower limit of the strength generated by the quantile boundaries of each of the inference results is less than the demolding strength threshold; releasing the delayed demolding command when the lower limit of the strength reaches the demolding strength threshold and continues for the minimum holding time; adjusting the insulation coverage thickness or cooling medium flow rate when the upper limit of the temperature gradient reaches the temperature control trigger threshold; restoring the temperature control state before adjustment when the upper limit of the temperature gradient is less than the temperature control release threshold and continues for the minimum holding time; the temperature control release threshold is determined by subtracting the temperature control hysteresis width from the temperature control trigger threshold; adjusting the spraying interval and coverage state when the upper limit of the crack risk zone reaches the maintenance trigger threshold, and releasing the adjustment according to the lower limit of the crack risk zone and the corresponding action release threshold; the demolding strength threshold is determined by the strength test results of the specimen with the same mix proportion; the temperature control trigger threshold is determined by the allowable temperature difference design value; and the temperature control hysteresis width is determined by the preset quantile of the difference between the temperature gradient trigger value and the release value in the historical temperature control action record.

10. The intelligent prevention and control method for rain damage in concrete construction according to claim 8, characterized in that, Step S5 further includes: recording the data version, risk range, radar reliability, sensor residual, start and end times of the action, and action result corresponding to each risk upper bound triggering, risk lower bound release, protection action switching, and sensor recovery; establishing independent verification sample sets for construction risk ranges and crack risk ranges respectively, wherein the independent verification sample set for construction risk ranges does not overlap with the training and calibration samples of the neural sequence-preserving controller, and the independent verification sample set for crack risk ranges does not overlap with the parameter fitting samples of the physical information neural network; updating the calibration samples and parameters of the neural sequence-preserving controller based on the deviation between the action result and the subsequent damage state of the poured surface, and updating the calibration samples and parameters of the neural sequence-preserving controller based on the deviation between the action result and the subsequent internal temperature and humidity data of the concrete. The observation weights and parameters of the physical information neural network are updated accordingly. The updated neural order-preserving controller is applied to the independent verification sample set of the construction risk interval, and the updated construction risk interval coverage is recalculated. The updated physical information neural network is applied to the independent verification sample set of the crack risk interval, and the updated crack risk interval coverage is recalculated. The preset coverage is used as the target coverage of the construction risk interval, and the target coverage of the crack risk interval is determined according to the coverage requirements of the crack marking interval. When any updated interval coverage is less than the corresponding target coverage, the calibration samples and parameters of the neural order-preserving controller and the observation weights and parameters of the physical information neural network are restored to the same data version before the update.

Images