A deep learning-based ground subsidence prediction and reinforcement effect dynamic evaluation method

Abstract

This invention discloses a deep learning-based method for predicting foundation settlement and dynamically evaluating reinforcement effects, belonging to the field of monitoring and intelligent evaluation technology. By deploying settlement monitoring sections, pore water pressure monitoring points, and a construction loading recording system, multi-source time-series data of the entire construction process are collected. The raw data undergoes anomaly identification, missing data compensation, and standardization. The standardized construction state time series is input into a dynamic prediction model, which outputs a settlement prediction curve under actual working conditions and a virtual baseline under no reinforcement intervention. By calculating the time-by-time difference between the two types of curves, the net contribution value of reinforcement measures to settlement control is obtained. The net contribution value is then dynamically compared with staged control thresholds to automatically generate control instructions and guide the adjustment of construction parameters, achieving real-time prediction of soft soil foundation settlement, quantitative evaluation of reinforcement effects, and dynamic optimization of construction parameters.

Description

Technical Field

[0001] This invention relates to the field of monitoring and intelligent assessment technology, specifically a method for predicting foundation settlement and dynamically assessing reinforcement effects based on deep learning. Background Technology

[0002] Foundation settlement control is a core aspect of civil engineering construction, directly impacting the safety and long-term service performance of the superstructure. In large-scale infrastructure and high-rise building construction, soft soil, characterized by low bearing capacity and high compressibility, often experiences continuous and uneven settlement. Therefore, engineering projects commonly employ reinforcement measures such as surcharge preloading, cement-soil mixing piles, and vacuum preloading to improve the overall performance of the foundation. Currently, mainstream settlement prediction methods are still based on pre-construction geological survey data, utilizing soil consolidation theory (such as Terzaghi's one-dimensional consolidation equation) or numerical simulation (such as finite element analysis) to perform a one-time static calculation, obtaining the settlement-time curve under given working conditions. This curve serves as the basis for design and a reference for construction control.

[0003] However, this static prediction model has revealed significant technical defects in practical engineering applications. On the one hand, the actual working conditions during construction, such as load grading, groundwater level fluctuations, precipitation effects, and construction disturbances, often differ significantly from the assumed conditions of the survey. This leads to a gradual deviation between the initial prediction curve and the on-site monitoring data. Furthermore, once the prediction model is determined, it can no longer absorb dynamic information such as settlement, pore pressure, and earth pressure collected in real time by subsequent sensors, and the model itself loses its self-correction ability. On the other hand, when the construction party takes reinforcement measures based on the monitoring signals exceeding the standard (such as adding mixing piles or increasing preload), due to the lack of a parallel prediction benchmark in the unreinforced background, it is difficult for engineering technicians to quantitatively evaluate the actual settlement reduction contribution of the measures taken. It is also impossible to distinguish whether the slowdown in settlement rate is a direct mechanical effect of the reinforcement measures or a delayed manifestation of the natural consolidation process of the soil under the original load. For example, in a coastal soft soil subgrade project, after monitoring that the settlement rate exceeded the limit, the construction team urgently added two rows of cement mixing piles. In the following three months, the settlement rate decreased significantly. However, the design team could not determine whether the decrease was due to the load sharing effect of the newly added piles or the delayed increase in the degree of consolidation of the previous surcharge preloading. This ambiguity of attribution directly led to the inability to accurately calculate the economic benefits of reinforcement investment and also made optimization decisions (such as whether to add more piles and when to unload) lack scientific quantitative basis.

[0004] In summary, existing technologies completely decouple settlement prediction from the construction process in the time dimension. The prediction results cannot be dynamically updated as the working conditions evolve, and the reinforcement effect cannot be objectively attributed based on reliable comparison benchmarks. As a result, on-site decision-making has long relied on engineering experience and lacks data-driven scientific support. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based method for predicting foundation settlement and dynamically evaluating reinforcement effects, enabling real-time prediction of soft soil foundation settlement, quantitative evaluation of reinforcement effects, and dynamic optimization of construction parameters.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] This application provides a deep learning-based method for predicting foundation settlement and dynamically evaluating reinforcement effects, comprising the following steps:

[0008] S1. Construction Status Perception and Data Standardization: Collect raw time-series data of foundation settlement, pore water pressure and graded loading information within the construction impact area, and perform anomaly identification, missing data compensation and trend smoothing on the raw time-series data to form a standardized construction status time series.

[0009] S2. Rolling prediction of settlement trend under actual working conditions: Input the standardized construction state time series into the dynamic prediction model and output the settlement development prediction curve under actual working conditions.

[0010] S3. Deduction of settlement trend under the scenario of no reinforcement intervention: Based on the dynamic prediction model, the characteristic quantities of reinforcement measures are removed from the input parameters, and only the natural foundation consolidation parameters and the current actual load conditions are retained to deduce the virtual baseline under the scenario of no reinforcement intervention.

[0011] S4. Quantitative separation of reinforcement contribution: Calculate the time-by-time difference between the settlement development prediction curve and the virtual baseline to obtain the net contribution value caused by the reinforcement measures;

[0012] S5. Phased achievement determination and control instruction generation: The net contribution value is dynamically compared with the control threshold of the corresponding construction stage, and a control instruction is generated when the threshold requirements are not met.

[0013] S6. On-site execution and iterative control of control commands: The control commands are transmitted to the construction control system, and the adjusted working condition data is re-integrated into step S1. Steps S1 to S5 are executed iteratively until each construction stage meets the control requirements.

[0014] The beneficial effects of this invention are as follows:

[0015] To address the shortcomings of traditional methods that rely on pre-construction survey data for one-time static calculations, where the predicted curve gradually deviates from the measured value as construction progresses and the model cannot absorb subsequent dynamic monitoring information, this method collects and standardizes multi-source monitoring data in real time through step S1, and then inputs the standardized time sequence into the dynamic prediction model in step S2, automatically correcting the prediction deviation time by time. This ensures that the settlement prediction curve remains synchronized with the actual loading rate and consolidation process on site, solving the problem of complete decoupling between prediction and construction process in the time dimension, and providing a real-time and reliable prediction benchmark for construction control.

[0016] To address the issue of attribution ambiguity—where engineers cannot distinguish whether the slowdown in settlement rate is a direct result of reinforcement measures or a delayed manifestation of natural soil consolidation, and the economic viability of reinforcement investment cannot be accurately calculated—step S3 sets the characteristic quantities of reinforcement measures to zero and derives a virtual baseline without reinforcement intervention. Step S4 then calculates the time-by-time difference between the actual working condition prediction curve and the virtual baseline to obtain the net contribution value of reinforcement measures at each stage. This enables the design team to accurately determine the actual settlement reduction contribution of reinforcement investments such as adding mixing piles and increasing preload, providing a quantitative basis for optimizing reinforcement schemes.

[0017] To address the decision-making dilemma where engineers can only rely on experience to determine whether to adjust the reinforcement plan after monitoring data exceeds the standard, lacking scientific quantitative basis, step S5 pre-sets control thresholds for each construction stage based on the engineering design, dynamically compares the net contribution quantification results with the thresholds, and automatically outputs control instructions containing specific adjustment ranges when the data is continuously below the threshold. This enables decisions such as the number of additional piles, adjustment of loading rate, and extension of preloading time to be automatically matched and generated based on the degree of deviation, achieving objectivity, standardization, and traceability of reinforcement control.

[0018] To address the technical shortcomings of the complete decoupling of settlement prediction and construction process, and the separation of prediction and control, which leads to long-term reliance on engineering experience and lack of data-driven decision-making on site, step S6 transmits control instructions to the construction control system for execution. The adjusted working condition data is then reintegrated into step S1 for iterative cycles until all stages meet the control requirements. This ensures that the effect of each round of control can be verified and corrected by the next round of monitoring data, ensuring that the foundation settlement remains under control throughout the entire construction cycle and providing complete data support for project acceptance.

[0019] To address the misjudgment issues caused by using a unified control standard for different consolidation stages—such as false alarms triggered by instantaneous loading fluctuations in the primary consolidation stage and persistent failure to meet standards in the secondary consolidation stage due to slow settlement rate decay—an encoder module automatically identifies the three stages: instantaneous loading, primary consolidation, and secondary consolidation. Based on the identification results, the control threshold is dynamically adjusted, ensuring that the control strategy accurately matches the actual consolidation state of the foundation. This improves the scientific rigor and rationality of the compliance determination and avoids ineffective warnings or over-reinforcement caused by stage misjudgments in engineering practice. Attached Figure Description

[0020] To better understand and implement this application, the technical solution is described in detail below with reference to the accompanying drawings.

[0021] Figure 1 This is a flowchart illustrating a deep learning-based method for predicting foundation settlement and dynamically evaluating reinforcement effects, as provided in Embodiment 1 of this application. Detailed Implementation

[0022] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, exemplary embodiments will be described in detail below, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0023] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0024] The following detailed description of the specific implementation methods, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided in detail.

[0025] Example 1, please refer to Figure 1 This embodiment provides a method for predicting foundation settlement and dynamically evaluating reinforcement effects based on deep learning, including the following steps:

[0026] S1. Construction Status Perception and Data Standardization: Monitoring sections are set up within the construction impact area to continuously collect raw time-series data of foundation settlement, pore water pressure, and graded loading information at different construction stages. Anomalies are removed and trends are smoothed from the raw time-series data to generate a standardized construction status time series.

[0027] Further, step S1 specifically includes:

[0028] S11. Monitoring section layout and multi-source data acquisition: Settlement monitoring sections are set up around the surcharge preloading zone or within the influence range of the mixing pile construction section. Settlement plates or layered settlement meters are buried in layers along the foundation depth direction to obtain the settlement of soil layers at different depths. Pore water pressure gauges are set up within the key influence range of the vacuum preloading zone to monitor the dissipation process of excess pore water pressure in real time. The construction recorder is connected simultaneously to obtain the graded loading height, loading rate and preloading duration of the current stage.

[0029] Among them, the principles for unifying the spatial coordinate benchmarks and setting up cross-sections for monitored data include:

[0030] All monitoring data (settlement, pore water pressure, loading information) adopt the same construction coordinate system and elevation benchmark, and the data of each monitoring channel are comparable in time and space;

[0031] Settlement monitoring section layout principle: In the surcharge preloading zone, monitoring sections are set up at the surcharge centerline, slope toe, and edge of the affected area; in the mixing pile construction section, monitoring points are set up at the center of the pile group, the soil between piles, and the edge of the pile group; each monitoring section is buried with settlement plates or layered settlement meters at 1m to 2m intervals along the depth direction.

[0032] The principle for the placement of pore water pressure gauges is to place them at different depths (including shallow, medium and deep layers) at the most unfavorable stress section. The specific depth is determined according to the thickness of the soft soil layer, and should at least cover the main compressible layer range to ensure that the complete dissipation process of excess pore water pressure can be captured.

[0033] S12. Raw signal anomaly identification and removal: For the raw time series data of each monitoring channel collected, the outlier judgment criterion based on statistical distribution is adopted to identify jump points that deviate from the mean by more than three times the standard deviation as outliers and remove them. At the same time, missing time periods caused by poor sensor contact or signal transmission interruption are marked.

[0034] Specifically, when the number of missing periods exceeds a predetermined time window (the length of which is set according to the monitoring sampling frequency, generally three consecutive sampling cycles), the system automatically activates the data compensation mechanism.

[0035] By employing linear interpolation or physical trend compensation methods based on the degree of consolidation-time relationship, and utilizing effective monitoring data before and after the missing period, combined with the loading plan and consolidation theory of the current construction stage, the missing data can be reasonably estimated and filled.

[0036] All data generated after compensation is marked with an "estimated value" label during storage, and the compensation method and basis are recorded.

[0037] If the continuous missing time exceeds 10% of the total time of the construction phase, the system will generate a sensor maintenance warning, suspend the settlement prediction cycle, restart it after the data returns to normal, and mark the missing data period as "unavailable" and not participate in the contribution quantification calculation of that phase.

[0038] S13. Time series data smoothing and standardized output: The sliding window filtering method is used to smooth the trend of the time series data after removing anomalies, eliminating high-frequency random noise caused by diurnal temperature difference, mechanical vibration or instrument zero drift, retaining the true consolidation deformation trend of the soil under load, and forming a standardized construction state time series with a unified time base and sampling interval.

[0039] Specifically, S1 collects settlement, pore pressure and loading data by deploying a multi-source monitoring network, and performs anomaly identification, missing data compensation and trend smoothing on the original time series. This solves the problem of discontinuous and unreliable data caused by environmental interference, sensor failure and signal noise in traditional methods, and provides a continuous, reliable and standardized input data foundation for subsequent dynamic prediction models.

[0040] S2. Rolling prediction of settlement trend under actual working conditions: Using the standardized construction state time series as input, the pre-established dynamic prediction model is driven to output the settlement development prediction curve under the current construction state in time period by time. The dynamic prediction model automatically corrects the prediction deviation of the previous moment after the arrival of new data, so that the prediction results are synchronized with the actual loading rate and consolidation process on site.

[0041] Further, step S2 specifically includes:

[0042] S21. Settlement evolution time sequence feature encoding: The standardized construction state time sequence is input into the encoder module of the dynamic prediction model. The encoder extracts the time sequence dependency in the monitoring data along the time axis moment by moment, identifies the consolidation state of the current construction stage, including the evolution features of the instantaneous loading stage, the main consolidation stage and the secondary consolidation stage, and compresses the above features into a fixed-dimensional context feature vector.

[0043] The encoder calculates the settlement rate characteristic value, pore pressure dissipation rate characteristic value, and load increment characteristic value at the current moment based on the data change rate of each monitoring channel in the standardized construction state time series. It then matches the above three characteristic values ​​with a pre-stored consolidation stage discrimination rule library to output the consolidation stage label corresponding to the current moment. The discrimination rule library is pre-calibrated according to different soil types and loading conditions.

[0044] In the consolidation stage discrimination rule base: the discrimination condition for the instantaneous loading stage is that the settlement rate exceeds the first threshold for three consecutive sampling points and the pore pressure increment exceeds the second threshold; the discrimination condition for the primary consolidation stage is that the settlement rate shows a monotonically decreasing trend and the pore pressure dissipation rate is greater than the third threshold; the discrimination condition for the secondary consolidation stage is that the settlement rate is lower than the fourth threshold and the pore pressure change rate is less than the fifth threshold for more than a preset duration; the first to fifth thresholds are preset according to the site soil survey parameters and loading scheme.

[0045] The first to fifth thresholds are set as follows: the first threshold is 1.2 to 1.5 times the theoretical maximum settlement rate of the normally consolidated soil at the site under the design loading rate; the second threshold is 70% to 80% of the theoretical incremental value of the design loading; the third, fourth, and fifth thresholds are jointly calibrated by historical engineering data, indoor consolidation test results, and field test section monitoring data, and stored in the threshold parameter library.

[0046] The threshold parameter library includes at least: soft soil layer thickness parameters; natural moisture content parameters; compression modulus parameters; consolidation coefficient parameters; construction loading rate parameters; and stage identification rule parameters corresponding to the construction stage.

[0047] When the conditions at the engineering site change significantly, the threshold parameter library supports recalibration and updates.

[0048] S22, Future Settlement Sequence Rolling Decoding Output: The decoder module takes the context feature vector as the initial state and generates the settlement prediction value within the set time window in the future through autoregression. The output at each time moment depends on the prediction result of the previous time moment and the load input conditions of the current time moment, forming a settlement development prediction curve that progresses with time.

[0049] In step S22, when the decoder generates the future settlement prediction sequence through autoregression, the output value at each moment depends on both the prediction result of the previous moment and the load input conditions of the current moment. Specifically, the predicted settlement value of the previous moment is used as a component of the decoder input for the next moment, and together with the graded loading height and loading rate information of the current moment, they constitute the decoder's input vector for the current moment. This ensures that the load change information is transmitted moment by moment in the time-series prediction, and that the prediction curve remains sensitive to changes in the construction loading rate.

[0050] S23. Online fine-tuning of model parameters and self-correction of deviation: After each newly collected monitoring data arrives, the deviation between the current model output settlement prediction value and the actual monitoring value is used as a feedback signal. The model parameters are slightly adjusted in the direction of reducing prediction error, so that the model's prediction benchmark gradually approaches the actual response on site as the construction progresses, ensuring that the prediction curve always remains synchronized with the measured data.

[0051] Furthermore, in step S23, the parameter fine-tuning of the dynamic prediction model is performed according to the following recursive formula: ;

[0052] in, This is the model parameter vector at the k-th update round. The learning rate is updated in the kth round. For predicting the loss function, This represents the predicted settlement value output by the model in the k-th round. The actual settlement value monitored in the kth round is... Let λ be the gradient of the loss function with respect to the model parameters, and λ be the elasticity weight solidification coefficient. This refers to the parameter vector when the model's initial training is complete; the last term in the recursive formula... This is an elastic constraint term that allows the model to fine-tune its parameters while absorbing new monitoring data, while retaining the knowledge learned in the initial training phase, thus avoiding catastrophic forgetting during the parameter update process.

[0053] Specifically, S2 extracts temporal dependencies and identifies consolidation stages through the encoder, and generates prediction sequences by autoregression based on load conditions through the decoder. At the same time, it fine-tunes model parameters online after new data arrives to automatically correct prediction deviations. This solves the problems of traditional static prediction models being unable to absorb subsequent monitoring data, prediction results gradually deviating from measured values ​​as construction progresses, and being unable to respond to changes in loading rate. It achieves real-time dynamic synchronization between the settlement prediction curve and the actual working conditions on site.

[0054] S3. Settlement trend deduction under the scenario of no reinforcement intervention: Based on the dynamic prediction model, the characteristic quantities of the reinforcement measures implemented have been set to zero in the input parameters, and only the natural foundation consolidation parameters and the current actual load conditions are retained. The settlement time series evolution curve under the assumption that no reinforcement intervention has been taken is simultaneously deduced, and this curve is used as a virtual baseline.

[0055] Furthermore, step S3 specifically includes:

[0056] S31. Identification and zeroing of reinforcement measure characteristic quantities: Identify the characteristic quantities corresponding to the implemented reinforcement measures from the current input parameters, including the area replacement ratio, pile length, pile spacing of cement mixing piles, the additional load increment and preloading duration of surcharge preloading, or the vacuum degree and drainage board spacing of vacuum preloading. Set all the values ​​of these characteristic quantities to zero so that the model can remove all reinforcement intervention information at the input level.

[0057] S32. Independent input of natural foundation consolidation parameters: Input the natural foundation consolidation parameters obtained during the site investigation stage into the model, including consolidation coefficient, compression modulus and permeability coefficient. Combine the current graded loading height and load application time to construct consolidation calculation conditions that depend only on the performance of the natural foundation.

[0058] S33. Virtual Unreinforced Settlement Curve Deduction: Drive the dynamic prediction model to perform forward calculations under the above input conditions, and simultaneously deduce the temporal evolution of foundation settlement from the current moment to a future set time window, assuming no reinforcement measures are taken. Use the deduction results as a virtual baseline for evaluating the reinforcement contribution. This baseline represents the theoretical settlement background determined only by the natural foundation consolidation and the current load conditions.

[0059] In step S33, the derivation result of the virtual baseline is jointly constrained by the natural foundation consolidation parameters and the current actual load conditions. Furthermore, the current consolidation stage information identified by the encoder is simultaneously injected into the model derivation process, so that the shape of the virtual baseline matches the settlement law that the natural foundation should have in the current consolidation stage.

[0060] The specific implementation method of simultaneously injecting the consolidation stage information into the model derivation process is as follows: the consolidation stage label output by the encoder is converted into a learnable embedding vector, and features are concatenated with the load input of the decoder at each time step. This allows the decoder to clearly know the type of consolidation stage it is currently in when generating the virtual baseline, thereby controlling the output curve to focus on peak capture in the instantaneous loading stage, on decay rate in the main consolidation stage, and on linear creep trend in the secondary consolidation stage, ensuring that the shape of the virtual baseline is consistent with the theoretical settlement law of the natural foundation in each consolidation stage.

[0061] S3 sets the characteristic quantities of the implemented reinforcement measures to zero, retains only the natural foundation consolidation parameters and the current load conditions, and deduces a virtual unreinforced settlement curve as a virtual baseline. It also sets consistency constraints and physical rationality verification mechanisms, which solves the problem that traditional methods cannot distinguish between reinforcement effectiveness and natural consolidation effects and the attribution of reinforcement contribution is ambiguous due to the lack of parallel prediction benchmarks for unreinforced scenarios. It provides a common, physically reliable and directly comparable reference benchmark for the subsequent quantitative separation of reinforcement contribution.

[0062] S4. Quantitative separation of reinforcement contribution: The settlement development prediction curve and the virtual baseline are subjected to point-by-point difference calculation on the same time axis to obtain the time series value of settlement reduction directly caused by reinforcement measures. This time series value is used as the net contribution quantification result of reinforcement measures to settlement control at each construction stage.

[0063] Furthermore, step S4 specifically includes:

[0064] S41. Calculation of settlement difference at each time step: Subtract the settlement development prediction curve under actual working conditions from the virtual baseline at each monitoring sampling time, calculate the difference according to the settlement difference at each time step, and obtain the settlement reduction amount generated by the reinforcement measures. If the settlement reduction amount is greater than zero, it means that the reinforcement measures have made a positive contribution to settlement reduction. If the settlement reduction amount is less than or equal to zero, it means that the reinforcement measures have not yet played an effective role.

[0065] The formula for calculating the settlement difference at each time step is as follows: C(t) is the net contribution value at time t; This represents the settlement value of the virtual baseline without reinforcement intervention. This represents the predicted settlement value under actual working conditions.

[0066] S42. Statistics of cumulative contribution within a stage: The net contribution values ​​at all sampling times within the same construction stage are summed according to the time interval to obtain the total cumulative net contribution of the reinforcement measures in that stage. The total cumulative net contribution reflects the overall settlement reduction effect of the reinforcement measures in that stage.

[0067] The total cumulative net contribution is expressed as follows: The Let be the total cumulative net contribution of the j-th construction stage, m be the number of sampling points in this stage, and Δt be the sampling time interval. When the total cumulative net contribution is greater than the threshold of the corresponding stage, the reinforcement measures in this stage are deemed to meet the control requirements. When the total cumulative net contribution is less than the threshold of the corresponding stage, the reinforcement effect in this stage is deemed insufficient.

[0068] In step S42, the value of the monitoring sampling time interval Δt is automatically adjusted according to the construction stage: during the active loading stage where the loading rate changes frequently, Δt is less than or equal to 1 hour to capture the rapid changing trend of the reinforcement contribution; during the consolidation stabilization period in the surcharge maintenance or preloading stage, Δt is extended to 6 to 24 hours to reduce data redundancy and computational burden; the time interval for each construction stage... Based on the dynamic division of the loading plan in the engineering design and the actual construction progress, the start and end times of each stage are automatically determined according to the execution time of the actual loading operation on site.

[0069] S43. Dynamic output of contribution sequence for each stage: The net contribution value calculated at each time step and the total cumulative net contribution of each stage are organized into a contribution quantification result sequence according to time sequence, which is used as the input data for the compliance judgment in step S5.

[0070] S44. Segmented setting of control thresholds for each construction stage: Based on the allowable total final settlement and allowable differential settlement value given in the engineering design, and combined with the site consolidation degree-time relationship curve, the total settlement control target is allocated according to the construction period of each construction stage.

[0071] The site consolidation degree-time relationship curve is calculated based on Terzaghi consolidation theory or inverted and fitted based on field measured consolidation degree data.

[0072] The settlement is allocated according to the proportion of the construction period for each stage, including: if the design documents have clearly defined the maximum settlement rate and final settlement control value for each construction stage, then the value specified in the design documents is directly used as the threshold for that stage; if the design documents only give the final total settlement control target without detailing it by stage, then the percentage of consolidation degree that the foundation should reach at the end of each stage is determined according to the consolidation degree-time relationship curve, and this percentage is multiplied by the allowable final total settlement to calculate the cumulative settlement control value for each construction stage. Then, combined with the planned construction period of that stage, the cumulative settlement control value is converted into the maximum allowable settlement rate threshold for that stage; once the threshold for each stage is set, it is stored in the control parameter library and remains unchanged in subsequent iterations unless there are design changes or the supervising engineer confirms in writing that an adjustment is required.

[0073] Specifically, S4 calculates the time-by-time difference between the actual settlement prediction curve and the virtual baseline on the same time axis, and sums the net contribution values ​​of each stage according to time intervals. At the same time, it sets control thresholds for each construction stage according to engineering design requirements. This solves the technical problem that traditional methods cannot quantify the actual contribution of reinforcement measures and cannot distinguish between reinforcement effectiveness and natural consolidation effect. It realizes real-time quantitative output and stage cumulative statistics of the net contribution value of reinforcement measures, providing a comparable quantitative basis for subsequent stage-by-stage compliance judgment.

[0074] S5. Phased achievement judgment and control instruction generation: Based on the pre-set settlement control thresholds for each construction stage in the engineering design, the net contribution quantification results are dynamically compared with the corresponding stage thresholds. When the net contribution quantification results for a certain stage are continuously lower than the stage threshold, control instructions containing suggestions for adjusting reinforcement parameters are automatically output.

[0075] Furthermore, step S5 specifically includes:

[0076] S51. Presetting and storing control thresholds for each stage: Before construction begins, the maximum allowable settlement rate and final settlement control values ​​for each construction stage are extracted and set according to the engineering design documents. The thresholds for different stages are stored in the control parameter library in segments according to the time axis, serving as the benchmark for subsequent dynamic comparison.

[0077] The settlement control thresholds for each construction stage are automatically adjusted based on the consolidation stage type identified by the encoder: when it is determined that the current stage is instantaneous loading, only the settlement rate threshold is monitored and the reinforcement contribution compliance determination is not initiated; when it is determined that the current stage is primary consolidation, the reinforcement contribution compliance determination is activated and the control threshold corresponding to the primary consolidation stage is adopted; when it is determined that the current stage is secondary consolidation, the control target is switched to the final settlement threshold, while the settlement rate compliance requirement is relaxed.

[0078] When the consolidation stage type automatically adjusts the control threshold, the following transition period confirmation mechanism is executed:

[0079] When the consolidation stage label output by the encoder changes from the current stage to an adjacent stage, the control threshold is not switched immediately. Instead, a confirmation window is set. That is, the label remains unchanged for three consecutive sampling cycles before the control threshold of the corresponding stage is officially switched. If the label jumps back to the original stage during the confirmation window, the switch is not performed and the original control threshold remains unchanged. The net contribution quantification result during this transition period is still judged according to the original stage control threshold to avoid the control target from changing constantly due to frequent switching of the consolidation stage discrimination label, which would affect the continuity and stability of on-site construction.

[0080] S52. Real-time dynamic comparison of contribution value and threshold: At each monitoring sampling time, the net contribution quantification result output in step S4 is compared with the control threshold corresponding to the current construction stage in real time. The safety margin or defect range of the net contribution value relative to the threshold is calculated, and the current reinforcement measures are evaluated to determine whether they meet the expected settlement reduction requirements.

[0081] S53. Continuous non-compliance judgment and early warning trigger: When the net contribution quantification result is lower than the corresponding stage threshold for N consecutive sampling periods, the system automatically judges that the reinforcement effect of the stage is insufficient and triggers an early warning signal. The value of N is preset by engineering technicians according to the monitoring sampling frequency and the duration of the construction stage.

[0082] S54. Automatic matching and output of control commands: Based on the degree of deviation between the net contribution value and the threshold, the corresponding reinforcement parameter adjustment scheme is automatically matched. The adjustment scheme includes increasing the number of mixing piles or adjusting the pile position arrangement, increasing the surcharge preload value or extending the preload time, adjusting the vacuum degree of vacuum preload or the spacing of drainage boards, reducing the loading rate or adding temporary unloading measures, and outputting the matched control commands to the construction control system.

[0083] In step S54, the control command automatically matches the corresponding adjustment scheme according to the degree of deviation between the net contribution value and the threshold: when the deviation value is less than the first predetermined range, a light adjustment suggestion of reducing the loading rate or extending the preloading time is output; when the deviation value is greater than or equal to the first predetermined range and less than the second predetermined range, a moderate adjustment suggestion of increasing the preloading load or adding temporary unloading is output; when the deviation value is greater than or equal to the second predetermined range, a strength adjustment suggestion of increasing the number of mixing piles or adding piles is output.

[0084] The first predetermined range of the deviation is 20% of the control threshold for this stage, and the second predetermined range is 50% of the control threshold for this stage. This baseline value allows engineers to manually adjust it within ±10% based on the importance level of the project; wherein:

[0085] Suggested minor adjustments: Output "Reduce the current loading rate to 80%~90% of the original rate" or "Extend the pre-compression time of this stage by 3~7 days";

[0086] Moderate adjustment suggestion: Output "Increase the preload to 110%~120% of the design value" or "Add temporary unloading, with the unloading height being 20%~30% of the current surcharge height";

[0087] Strength adjustment suggestion: When outputting "Increase the number of mixing piles" or "Add piles", follow the formula.

[0088]

[0089] Calculate the number of additional piles, where To achieve the target area replacement rate required to reach the control threshold, Given the current area replacement rate, This represents the total area of ​​the reinforcement zone; it also outputs the recommended locations for additional piles (which should be placed at the section with the largest current settlement).

[0090] Specifically, S5 dynamically compares the control thresholds of each construction stage with the net contribution quantification results in real time, and automatically adjusts the control strategy and sets a transition period confirmation mechanism according to the consolidation stage type identified by the encoder to avoid frequent label switching. At the same time, when the standard is not met continuously, it matches the corresponding reinforcement parameter adjustment scheme according to the degree of deviation and outputs specific executable control instructions. This solves the decision-making dilemma of engineers relying on experience and lacking scientific quantitative basis after monitoring exceeds the standard in traditional methods. It realizes the objectivity, standardization and traceability of reinforcement control, and provides clear and quantitative execution basis for construction parameter adjustment.

[0091] S6. On-site execution and iterative control of control instructions: The control instructions are transmitted to the construction control system to guide the on-site adjustment of reinforcement parameters or construction rate. The adjusted new working condition data is then re-integrated into step S1. Steps S1 to S5 are executed iteratively until the net contribution quantification results of all construction stages meet the corresponding control thresholds.

[0092] Furthermore, step S6 specifically includes:

[0093] S61. On-site transmission and reception of control commands: The control commands generated in step S5 are transmitted to the automatic control platform or handheld terminal device set up at the construction site through a wired or wireless communication network. The automatic control platform is connected to the control system of the construction machinery, and the handheld terminal device is for on-site engineers to view and confirm.

[0094] The automatic control platform is a central control unit set up at the construction site. It is connected to the control systems of construction machinery such as mixing pile construction machinery, surcharge loading equipment and vacuum preloading pump group through industrial Ethernet or wireless communication network. It has the functions of receiving control commands, parsing command parameters and automatically adjusting the operating status of construction machinery. It also supports two operation modes: manual confirmation and execution and fully automatic execution.

[0095] S62. Selection and implementation of command execution method: The control command is displayed in a visual form on the interactive interface, including the suggested parameter type, adjustment range and execution time window. The on-site engineer confirms the command according to the actual situation and then executes it manually, or the automatic control platform directly adjusts the construction parameters according to the command content.

[0096] S63. Re-integration of adjusted working condition data: The adjusted reinforcement parameters and construction rate are used as new working condition data and re-integrated into the standardized construction status time series of step S1.

[0097] S64. Full-stage compliance inspection and iteration termination: Repeat steps S1 to S5. After each cycle, check the net contribution quantification results of all construction stages. When the net contribution value of all stages at each sampling time is not lower than the corresponding control threshold, and the net contribution value of three consecutive sampling cycles shows a stable or increasing trend, it is determined that the foundation settlement has been effectively controlled. The cycle iteration is terminated and the reinforcement contribution compliance report of each stage is output.

[0098] The criteria for determining whether the net contribution value of the three consecutive sampling periods shows a stable or increasing trend are as follows: when the absolute value of the rate of change of the net contribution value of each two adjacent sampling periods is less than 5% of the stage control threshold, it is determined to be a stable trend; when the net contribution value of each sampling period is not less than 95% of the net contribution value of the previous period, it is determined to be an increasing trend; satisfying either a stable trend or an increasing trend is considered to meet the termination condition.

[0099] Furthermore, following step S64, the system also includes: unloading and completion condition determination and construction completion confirmation. After step S64 confirms that the net contribution quantification results of all construction stages meet the corresponding control thresholds, the system further performs the following completion determination:

[0100] Unloading condition verification: Based on the current measured settling rate and excess pore water pressure dissipation, determine whether the unloading conditions specified in the design documents are met (including settling rate less than the design allowable unloading rate and consolidation degree greater than the design requirement for unloading consolidation degree). If met, output a graded unloading suggestion instruction, including unloading height, unloading rate and unloading interval time for each stage.

[0101] If the current settlement rate or degree of consolidation does not meet the unloading conditions, a suggestion to extend the preloading maintenance time is output, and the loop is restarted to execute steps S1 to S5 until the unloading conditions are met.

[0102] Final compliance report output: The report will summarize the cumulative net contribution of each construction stage and the corresponding control threshold, the settlement rate time history curve, the excess pore water pressure dissipation curve, and the confirmation of the reinforcement contribution of each stage, and generate a full-process compliance report for foundation settlement control, which will serve as the technical basis for project acceptance.

[0103] Furthermore, following step S6, the process also includes: data storage and digital traceability archive construction for the entire construction process, storing data from each iteration in a hierarchical manner to the construction database according to the following four dimensions:

[0104] This includes: Input data layer: storing standardized construction status time series, raw data from each monitoring channel, and preprocessed logs, including outlier removal records and missing time period markers;

[0105] Predictive data layer: Stores the actual settlement development prediction curve, virtual baseline, and prediction model parameter update records for each round (model parameter vector and update time after each round).

[0106] Evaluation data layer: Stores the quantitative value of net contribution at each time step, the total cumulative net contribution at each stage, the achievement judgment results at each stage, and the corresponding control thresholds;

[0107] Decision data layer: Stores the time of each warning trigger, the content of the control instruction (including the type of parameter to be adjusted, the adjustment range and the execution time window), the instruction execution method (manual / automatic) and the execution result feedback;

[0108] The database adopts a structured time-series database, which organizes data according to a three-level index of project number, construction stage number, and cycle number, to ensure that all decisions and predictions throughout the construction process are verifiable, traceable, and reproducible.

[0109] Specifically, S6 transmits control commands to the site for execution and re-integrates the adjusted working condition data into step S1 to form a loop iteration. At the same time, after all stages meet the standards, it performs unloading condition verification and outputs graded unloading suggestions and a full-process compliance report. This solves the problems of settlement prediction and construction control being disconnected in traditional methods, the inability to continuously track the effectiveness of reinforcement measures after adjustment, and the lack of quantitative unloading basis in the construction closing stage. It realizes dynamic closed-loop control throughout the entire process and quantitative judgment of construction completion, ensuring that foundation settlement remains under control throughout the entire construction cycle.

[0110] A deep learning-based method for predicting foundation settlement and dynamically evaluating reinforcement effects is applicable to settlement control in any of the following construction scenarios under soft soil foundation conditions: surcharge preloading for deep soft soil roadbeds, cement-soil mixing pile composite foundation reinforcement for high-rise building foundations, vacuum preloading for coastal soft soil sites, and dynamic settlement control during graded loading construction. The characteristic quantities of the reinforcement measures under different construction scenarios correspond to different combinations of input parameters. The dynamic prediction model automatically identifies the current construction scenario type and adapts to the corresponding settlement evolution law based on the differences in the input parameter combinations.

[0111] The applicable prerequisites and boundary conditions for this method include: it is applicable to homogeneous or layered soft soil foundations, and the soil consolidation coefficient is within [a certain range]. Within the scope of conventional engineering experience, soft soil foundation treatment projects with construction loading rates within the range of 0.5~2.0m / month.

[0112] For complex sites such as karst development areas, highly sensitive structural soft soils (sensitivity > 4), or sites with significant groundwater seepage field influence (seepage velocity greater than the preset threshold Vg), the model can only be applied after the corresponding working condition data is included in the model training phase. Among them, Vg is determined by the field hydrogeological conditions and in-situ test results, and the model can only be applied after the corresponding working condition data is included in the model training phase.

[0113] When the actual engineering geological conditions exceed the above applicable range, the system will automatically output the prompt message "The current site conditions exceed the applicable range of the model. It is recommended to verify the model in conjunction with the on-site in-situ test data before using it", and will not directly output the prediction and evaluation results.

[0114] Before the model is put into use, the initial prediction results of the model must be verified using the measured data of at least the first three monitoring cycles of the site. Only after the verification is passed can it be officially used for subsequent dynamic prediction and reinforcement effect evaluation.

[0115] Example 2: This example provides a deep learning-based method for predicting foundation settlement and dynamically evaluating reinforcement effects. The difference between this example and Example 1 is that this example is applied to a soft soil foundation project in a port area's stockyard, using vacuum preloading combined with plastic drainage boards for reinforcement. In addition to settlement, the monitoring objects also include parameters such as vacuum degree, pore water pressure dissipation rate, and drainage flux. The model input features, threshold settings, and control instructions are all different from those in Example 1.

[0116] The reinforced area of ​​this project is approximately 86,000 square meters. The foundation mainly consists of dredged silty clay and silty soil, with high natural moisture content and high compressibility. The design requires control of the final post-construction settlement and differential settlement. To achieve dynamic prediction of settlement evolution and evaluation of reinforcement effect throughout the construction process, this embodiment is implemented according to the following steps.

[0117] Construction status perception and data standardization: Six monitoring sections are set up along the longitudinal and transverse directions within the construction area, including three main control sections and three auxiliary sections; five settlement monitoring points are set up in each section, located in the center of the vacuum zone, the edge zone, the area with dense drainage boards, the area with sparse drainage boards, and the influence transition zone.

[0118] Furthermore, pore water pressure gauges were installed at different depths of 2m, 6m, 10m and 14m to cover the main soft soil compression layer; at the same time, vacuum gauges and flow meters were installed on the main vacuum pipeline to collect construction parameters such as vacuum pressure, pumping flow rate and maintenance time in real time.

[0119] The sampling period for raw monitoring data is set as follows:

[0120] During the vacuum pressurization start-up phase, the sampling period is 15 minutes.

[0121] During the main body vacuum stabilization phase, the sampling cycle is 30 minutes.

[0122] During the later consolidation and stabilization stage, the sampling period was 2 hours.

[0123] The collected raw data undergoes anomaly detection, missing data compensation, and standardization, including:

[0124] The three-standard-deviation criterion was used to remove abrupt outliers;

[0125] For missing measurements caused by short-term air leakage in the vacuum line or loose sensor, piecewise linear interpolation combined with consolidation trend compensation method is used to fill the gaps.

[0126] High-frequency noise is smoothed using a sliding window mid-range filter, with a window length of 5 sampling points;

[0127] Settlement, pore water pressure, vacuum degree, air extraction flow rate, and drainage volume are uniformly mapped to the 0-1 interval to form a standardized construction state time series.

[0128] Rolling prediction of settlement trend under actual working conditions: Input the standardized construction state time series obtained in step S1 into the dynamic prediction model, and output the settlement development curve under the current construction state.

[0129] In this embodiment, the dynamic prediction model employs a time-series prediction network with a bidirectional gated recurrent unit (BiGRU) and an attention mechanism, which differs from the prediction structure used in Embodiment 1. This model includes an encoding layer, an attention weight layer, and a decoding layer, and is capable of simultaneously learning the coupling relationship between vacuum degree changes, pore pressure dissipation, and sedimentation response.

[0130] Specifically, the encoder input features include: current settlement value, pore water pressure, vacuum level, pumping flow rate, drainage flux, and cumulative vacuuming time;

[0131] Temporal dependencies are extracted using a bidirectional GRU to form a contextual feature vector;

[0132] The attention layer assigns higher weights to near-moment vacuum abrupt changes and the rapid dissipation range of pore pressure;

[0133] The decoder outputs the predicted settlement values ​​every 6 hours for the next 7 days.

[0134] Meanwhile, whenever new monitoring data is input, the model performs an online fine-tuning, so that subsequent prediction results can reflect in a timely manner the impact of vacuum pump start-up and shutdown, pipeline maintenance, and changes in local sealing status on foundation settlement.

[0135] Subsidence trend projection under no reinforcement intervention scenario: Based on the dynamic prediction model in step S2, a virtual baseline is constructed under the no reinforcement intervention scenario.

[0136] Unlike Example 1, in this example, the following reinforcement characteristic quantities are set to zero when simulating the unreinforced scenario:

[0137] Vacuum negative pressure value; drainage board spacing optimization effect; pumping duration; enhancement of drainage effect of surface sand pad layer.

[0138] Simultaneously, the natural foundation consolidation parameters, overlying fill load, and natural drainage conditions during construction are preserved to form a settlement evolution curve controlled solely by natural consolidation. This curve serves as a virtual baseline for subsequent calculations of the net settlement reduction contribution of vacuum preloading.

[0139] Quantitative separation of reinforcement contribution: The actual settlement prediction curve obtained in step S2 and the virtual baseline obtained in step S3 are calculated by time-by-time difference to obtain the net contribution value of vacuum preloading measures.

[0140] Specifically, it is expressed as follows: ; The net contribution value at time t; This represents the settlement value under the scenario of no reinforcement intervention; ; represents the predicted settlement value under actual working conditions.

[0141] When C(t)>0, it indicates that the vacuum preloading and drainage board have an effective effect on reducing settlement; when C(t)≤0, it indicates that the reinforcement effect is insufficient or has not yet manifested at the current stage.

[0142] Furthermore, the net contribution values ​​within the same construction stage are summed to obtain the total cumulative net contribution for that stage, which is used to measure the overall effect of vacuum preloading in that stage.

[0143] Phased achievement determination and control instruction generation: Based on the construction characteristics of this embodiment, the vacuum preloading process is divided into three phases: vacuum establishment phase; main consolidation phase; and stability maintenance phase.

[0144] Different control thresholds are set for different stages, specifically:

[0145] Vacuum establishment phase: The key is to monitor whether the vacuum level reaches more than 90% of the design value within 72 hours;

[0146] Primary consolidation stage: Focus on monitoring whether the settlement rate is consistently below 1.2 mm / d and whether the pore pressure dissipation rate reaches the predetermined threshold.

[0147] Stable maintenance phase: The focus is on monitoring whether the cumulative settlement increment tends to stabilize, requiring the settlement increment to be less than 3mm for 5 consecutive days.

[0148] When the net contribution value is lower than the corresponding threshold for N consecutive sampling periods, the system automatically generates a control command.

[0149] Unlike Example 1, the control commands output in this example mainly include:

[0150] Increase the vacuum level to 95%–100% of the design limit; check and repair any leaks in the sealing membrane; adjust the start-stop frequency of the vacuum pump; locally reinforce the drainage plates or supplement drainage channels; extend the vacuum maintenance time by 3–7 days.

[0151] Specifically, when the net contribution is insufficient but the vacuum level fluctuates significantly, a light adjustment command such as "seal check + increase vacuum level" is output first; when the pore pressure dissipation is significantly slower, a moderate adjustment command such as "extend maintenance time + localized drainage" is output; when the main consolidation stage continues to fail to meet the standard, a strength adjustment command such as "add drainage channels or reconstruct the pumping system" is output.

[0152] Control command execution and iterative cycle: The control commands generated in step S5 are transmitted to the field control system to guide construction personnel or automatic control platform to perform corresponding adjustments.

[0153] In this embodiment, the construction control system is linked with the vacuum pump unit, the air extraction valve and the sealing monitoring terminal. After the control command is executed, the construction parameters are automatically updated and returned to step S1, forming a closed-loop iterative process of monitoring, prediction, evaluation, control and re-monitoring.

[0154] The loop terminates when the following conditions are met:

[0155] The net contribution value of each stage is greater than the corresponding threshold for multiple consecutive sampling periods;

[0156] The settlement rate is consistently lower than the design allowable value;

[0157] The pore water pressure has basically dissipated, and the target degree of consolidation has been achieved.

[0158] The vacuum state is maintained continuously and stably.

[0159] The final output includes a settlement control assessment report, a reinforcement contribution analysis table, and a phased achievement confirmation result for the project.

[0160] Compared to Example 1, this example primarily targets the vacuum preloading engineering scenario of soft soil foundation reclaimed in port areas. The monitoring parameters have been expanded from settlement and pore pressure to more suitable features for vacuum preloading conditions, such as vacuum degree, air extraction flow rate, and drainage volume. The model structure employs a BiGRU combined with an attention mechanism, which is better suited to handling rapid fluctuations in vacuum degree and multi-temporal coupling relationships. The control strategy has also been adjusted accordingly to focus on increasing vacuum degree, leak detection, extending air extraction time, and densifying drainage channels, rather than adding piles or increasing surcharge. Although the core logic of reinforcement assessment still quantifies the net contribution by the difference between the actual working condition prediction curve and the unreinforced virtual baseline, its application targets, input characteristics, model adaptability, and dynamic control methods are more aligned with the actual needs of vacuum preloading engineering.

[0161] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any brief modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A deep learning-based ground settlement prediction and reinforcement effect dynamic evaluation method, characterized in that: Includes the following steps: S1. Construction Status Perception and Data Standardization: Collect raw time-series data of foundation settlement, pore water pressure and graded loading information within the construction impact area, and perform anomaly identification, missing data compensation and trend smoothing on the raw time-series data to form a standardized construction status time series. S2. Rolling prediction of settlement trend under actual working conditions: Input the standardized construction state time series into the dynamic prediction model and output the settlement development prediction curve under actual working conditions. Step S2 specifically includes: S21. Settlement evolution time series feature encoding: The encoder of the dynamic prediction model compresses the standardized construction state time series into a context feature vector and identifies the consolidation state of the current construction stage. The consolidation state includes the instantaneous loading stage, the primary consolidation stage, and the secondary consolidation stage. S22, Future Settlement Sequence Rolling Decoding Output: The decoder of the dynamic prediction model uses the context feature vector as the initial state and combines it with the load input conditions at each time step to generate a future settlement prediction sequence within a set time window. S23. Online fine-tuning of model parameters and self-correction of deviation: After each newly acquired monitoring data arrives, the deviation between the current model output settlement prediction value and the actual monitoring value is used as a feedback signal to fine-tune the model parameters in the direction of reducing prediction error, so that the prediction results are kept in sync with the field measured data. In step S21, the encoder calculates the current settlement rate characteristic value, pore pressure dissipation rate characteristic value, and load increment characteristic value based on the data change rate of each monitoring channel, and matches them with the pre-stored consolidation stage discrimination rule library to output the consolidation stage label corresponding to the current moment. In the discrimination rule base, the discrimination conditions for the instantaneous loading stage are that the settling rate exceeds the first threshold and the pore pressure increment exceeds the second threshold; the discrimination conditions for the primary consolidation stage are that the settling rate shows a monotonically decreasing trend and the pore pressure dissipation rate is greater than the third threshold; and the discrimination conditions for the secondary consolidation stage are that the settling rate is lower than the fourth threshold and the pore pressure change rate is less than the fifth threshold for a continuous period of time. S3. Deduction of settlement trend under the scenario of no reinforcement intervention: Based on the dynamic prediction model, the characteristic quantities of reinforcement measures are removed from the input parameters, and only the natural foundation consolidation parameters and the current actual load conditions are retained to deduce the virtual baseline under the scenario of no reinforcement intervention. S4. Quantitative separation of reinforcement contribution: Calculate the time-by-time difference between the settlement development prediction curve and the virtual baseline to obtain the net contribution value caused by the reinforcement measures; S5. Phased achievement determination and control instruction generation: The net contribution value is dynamically compared with the control threshold of the corresponding construction stage, and a control instruction is generated when the threshold requirements are not met. S6. On-site execution and iterative control of control commands: The control commands are transmitted to the construction control system, and the adjusted working condition data is re-integrated into step S1. Steps S1 to S5 are executed iteratively until each construction stage meets the control requirements.

2. The deep learning-based ground settlement prediction and reinforcement effect dynamic evaluation method according to claim 1, characterized in that: Step S1 specifically includes: S11. Monitoring deployment and multi-source data acquisition: Deploy settlement monitoring sections, pore water pressure monitoring points and construction loading recording devices to collect foundation settlement data, pore water pressure data and graded loading information respectively, and unify all monitoring data into the same construction coordinate system and elevation datum. S12. Anomaly identification and missing data compensation in raw data: Outlier detection methods based on statistical distribution are used to remove outliers, and linear interpolation or trend compensation based on the relationship between consolidation degree and time is performed on missing time periods to obtain complete time series data. S13. Time series data smoothing and standardization output: The time series data after anomaly removal and missing data compensation are smoothed and standardized by sliding window to generate a construction status time series with a unified time base and sampling interval.

3. The deep learning-based ground settlement prediction and reinforcement effect dynamic evaluation method according to claim 1, characterized in that: Step S3 specifically includes: S31. Identification and zeroing of reinforcement measure characteristic quantities: Identify the characteristic quantities corresponding to the implemented reinforcement measures from the current input parameters, including the area replacement ratio, pile length, and pile spacing of cement mixing piles, the additional load increment and preloading duration of surcharge preloading, or the vacuum degree and drainage board spacing of vacuum preloading, and set all the values ​​of these characteristic quantities to zero. S32. Independent input of natural foundation consolidation parameters: Input the natural foundation consolidation parameters obtained during the site investigation stage into the model, including consolidation coefficient, compression modulus and permeability coefficient. Combine the current graded loading height and load application time to construct consolidation calculation conditions that depend only on the performance of the natural foundation. S33. Virtual Unreinforced Settlement Curve Deduction: Drive the dynamic prediction model to perform forward calculations under the above input conditions, and deduce the temporal evolution of foundation settlement when no reinforcement measures are taken. The deduction results are used as a virtual baseline.

4. The deep learning-based ground settlement prediction and reinforcement effect dynamic evaluation method according to claim 1, characterized in that: Step S4 specifically includes: S41. Calculation of settlement difference at each time step: Subtract the settlement development prediction curve under actual working conditions from the virtual baseline at each monitoring sampling time to obtain the settlement reduction amount generated by the reinforcement measures at that time as the net contribution value. S42. Statistics of cumulative contribution within a stage: The net contribution values ​​at all sampling times within the same construction stage are summed according to the time interval to obtain the total cumulative net contribution of the reinforcement measures in that stage. S43. Dynamic output of contribution sequence for each stage: The net contribution value calculated for each construction stage and the total cumulative net contribution for each stage are organized into a contribution quantification result sequence according to time sequence.

5. The method for predicting foundation settlement and dynamically evaluating reinforcement effects based on deep learning according to claim 4, characterized in that: The total cumulative net contribution is represented as: ; the total cumulative net contribution of the jth construction phase is represented as: C(t) is the net contribution value at time t, m is the number of sampling points in the phase, and Δt is the sampling time interval.

6. The deep learning-based ground settlement prediction and reinforcement effect dynamic evaluation method according to claim 1, characterized in that: Step S5 specifically includes: S51. Presetting and storing control thresholds for each stage: Before construction begins, the maximum allowable settlement rate and final settlement control values ​​for each construction stage are extracted and set according to the engineering design documents. The thresholds for different stages are stored in the control parameter library in segments according to the time axis. S52. Real-time dynamic comparison of contribution value and threshold: At each monitoring sampling time, the net contribution quantification result output in step S4 is compared with the control threshold corresponding to the current construction stage in real time, and the safety margin or defect range of the net contribution value relative to the threshold is calculated. S53. Continuous non-compliance judgment and early warning trigger: When the net contribution quantification result is lower than the corresponding stage threshold for N consecutive sampling periods, the system automatically judges that the reinforcement effect of the stage is insufficient and triggers an early warning signal. S54. Automatic matching and output of control commands: Based on the degree of deviation between the net contribution value and the threshold, the corresponding reinforcement parameter adjustment scheme is automatically matched, and the matched control commands are output to the construction control system.

7. The method for predicting foundation settlement and dynamically evaluating reinforcement effects based on deep learning according to claim 1, characterized in that: Step S6 specifically includes: S61. On-site transmission and reception of control commands: Transmit the control commands generated in step S5 to the automatic control platform or handheld terminal device set up at the construction site through a wired or wireless communication network. S62. Selection and implementation of command execution method: Control commands are displayed in a visual form on the interactive interface, and are manually executed after confirmation by the on-site engineer, or the automatic control platform directly and automatically adjusts the construction parameters according to the command content. S63. Re-integration of adjusted working condition data: The adjusted reinforcement parameters and construction rate are used as new working condition data and re-integrated into the standardized construction status time series of step S1. S64. Full-stage compliance inspection and iteration termination: Repeat steps S1 to S5. After each cycle, check the net contribution quantification results of all construction stages until the net contribution value of all stages is not lower than the corresponding control threshold.

8. The method for predicting foundation settlement and dynamically evaluating reinforcement effects based on deep learning according to claim 7, characterized in that: Following step S64, the process further includes: unloading completion condition determination and construction completion confirmation. After step S64 confirms that the net contribution quantification results of all construction stages meet the corresponding control thresholds, a completion determination is performed, including: Unloading condition verification: Based on the current measured settling rate and excess pore water pressure dissipation, determine whether the unloading conditions specified in the design documents are met, including the settling rate being less than the design allowable unloading rate and the degree of consolidation being greater than the design requirement for unloading consolidation. If met, output a graded unloading suggestion instruction, including the unloading height, unloading rate and unloading interval time for each stage. If the current settlement rate or degree of consolidation does not meet the unloading conditions, a suggestion to extend the preloading maintenance time is output, and the loop is restarted to execute steps S1 to S5 until the unloading conditions are met. Final compliance report output: The report summarizes the cumulative net contribution of each construction stage and the corresponding control threshold, the settlement rate time history curve, the excess pore water pressure dissipation curve, and the confirmation of compliance of reinforcement contribution at each stage, and generates a full-process compliance report for foundation settlement control.

Images