Physical mechanism-driven building settlement intelligent prediction method, product and equipment

Abstract

The invention discloses a physical mechanism-driven building settlement intelligent prediction method, a product and equipment, and belongs to the technical field of civil engineering. The method comprises the following steps: S1, collecting building settlement data; s2, determining an empirical formula, and performing data enhancement on the building settlement data by empirical formula fitting to obtain settlement data constrained by a physical mechanism; s3, training an LSTM neural network by using the obtained settlement data constrained by the physical mechanism to obtain an LSTM model constrained by the physical mechanism; s4, building settlement data to be predicted are input into the LSTM model constrained by the physical mechanism, and future settlement of the building is predicted. The method can be effectively applied to a data set with characteristics of scarce data, discontinuity and the like, a data change rule is reasonably captured, and the purpose of high-precision prediction is achieved.

Description

Technical Field

[0001] This invention belongs to the field of civil engineering technology, specifically relating to a physical mechanism-driven intelligent prediction method, product, and equipment for building settlement. Background Technology

[0002] Building deformation is a core criterion and early warning signal for its structural safety and long-term stability. Uneven or excessive settlement may cause building damage or even structural instability. Settlement monitoring is a key component of deformation monitoring as clearly required by relevant standards and is a crucial indicator for construction, service life, and post-disaster assessment.

[0003] Existing settlement prediction methods include: Traditional methods, including empirical formulas and grey models, are simple to operate, but they generally suffer from significant fitting and prediction errors, and their prediction accuracy is difficult to meet practical needs. Intelligent methods: Represented by neural network methods such as BP and LSTM, when there is sufficient data, they can use their powerful autonomous learning and analysis capabilities to uncover the inherent patterns in the data. They have excellent performance and wide application in the fields of time series prediction and geotechnical engineering settlement prediction.

[0004] Existing intelligent prediction methods have rigid requirements for datasets, necessitating continuous time series data and sufficient sample size. For example, BP and LSTM rely on continuous time series data for training. However, monitoring data in engineering fields often exhibit discontinuous characteristics, large time intervals, and scarce samples due to non-continuous monitoring methods and high monitoring costs (e.g., subsidence data in urban geological disasters, where short-interval monitoring occurs during the emergency phase but becomes longer intervals later, requiring timely prediction based on limited data to support repairs). This leads to unreliable predictions by the models. Although T-LSTM attempts to process discontinuous data through time labels, its prediction accuracy is low for data with large time intervals (e.g., ...). Figure 1 As shown in the figure, it is difficult to meet the feasibility and reliability requirements of building settlement prediction in the field of urban geological disasters. Summary of the Invention

[0005] The present invention aims to at least partially solve one of the technical problems in the aforementioned related technologies.

[0006] Therefore, the purpose of this invention is to provide a physical mechanism-driven intelligent prediction method, product, and equipment for building settlement. Taking a ground settlement disaster in northern China as the research object, based on actual monitoring data, this invention integrates a physical mechanism model of building settlement and proposes a physical mechanism-constrained intelligent prediction method. By constructing a data augmentation mechanism for generating physical-driven virtual scene data, discontinuous measured data is coupled with a virtual sequence conforming to physical laws to generate a continuous and physically consistent dataset. This dataset is then combined with a physical mechanism-constrained neural network for model training and prediction, enabling prediction of scarce, large-span, and discontinuous time series data. Performance evaluation of the model shows that the mean absolute error (MAE) is 2.1325 mm, the root mean square error (RMSE) is 2.5445 mm, and the coefficient of determination (R²) is [not specified in the original text]. 2 The accuracy reached 0.9951, with an excellent prediction rate of 98.24%, demonstrating higher prediction accuracy than other methods. Subsequent experimental data further showed that the physical mechanism-driven intelligent prediction method can effectively predict the development trend of building settlement. The physical mechanism-driven intelligent prediction method proposed in this invention can be effectively applied to datasets with characteristics such as scarce and discontinuous data, reasonably capturing the data change patterns to achieve the goal of high-precision prediction.

[0007] To solve the above-mentioned technical problems, the present invention is implemented as follows: This invention provides a physical mechanism-driven intelligent prediction method for building settlement, the method comprising: S1. Collect building settlement data; S2. Determine the empirical formula, and use the empirical formula to fit and augment the building settlement data to obtain settlement data constrained by physical mechanisms; S3. Use the obtained physical mechanism-constrained settlement data to train the LSTM neural network to obtain a physical mechanism-constrained LSTM model; S4. Input the settlement data of the building to be predicted into the LSTM model constrained by the physical mechanism to predict the future settlement of the building.

[0008] In addition, the intelligent prediction method for building settlement driven by the physical mechanism of the present invention may also have the following additional technical features: In some implementations, step S2 includes: S21. Based on empirical formulas and collected building settlement data, perform data fitting and determine the fitting coefficients; S22. Obtain several virtual scene settlement data based on the fitting coefficients; S23. The obtained virtual scene settlement data and the collected real scene building settlement data are fused together to obtain settlement data constrained by physical mechanisms.

[0009] In some of these implementations, the empirical formula is a formula based on the hyperbola method, the exponential curve method, or the Hoshino method.

[0010] In some of these implementations, the empirical formula is expressed as: ; in, for Settlement at any given time Given the set of parameters in the empirical formula, curve fitting techniques, including the least squares method, are used to solve the optimization problem: ; in, Observed settlement at any given time This represents the number of data points within a local time interval.

[0011] In some implementations, the rules for generating mixed data during fusion in step S23 are as follows: for Real-time measured settlement data, for The virtual scene settling data at any given time is then fused into the final data. satisfy: .

[0012] In some implementations, step S3 includes: The input to an LSTM neural network is an enhanced data sequence. Output for the future Settlement prediction values ​​at each time step ; By minimizing the loss function between the predicted and actual values, the weight parameters of the LSTM neural network are optimized using the backpropagation algorithm. This enables the model to accurately map the complex relationship between input data and future subsidence trends, resulting in a physical mechanism-constrained LSTM neural network.

[0013] In some implementations, the loss function between the predicted and actual values ​​is: ; in, For the future The actual settlement value at any given moment.

[0014] In some of these implementations, physical constraints, including the requirement that the settlement rate must not change abruptly and must eventually stabilize within a specified range, are explicitly embedded during the training of the LSTM neural network to avoid unreasonable predictions, such as reverse settlement growth, when data is insufficient.

[0015] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the physical mechanism-driven intelligent prediction method for building settlement as described in any of the preceding embodiments.

[0016] This invention also provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the physical mechanism-driven intelligent prediction method for building settlement as described in any of the preceding embodiments.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: In this embodiment of the invention, the physical mechanism-driven intelligent prediction method for building settlement intelligently couples discontinuous and sparse measured data with a physically consistent virtual continuous sequence to generate an augmented dataset that is temporally continuous, physically self-consistent, and has a significantly expanded sample size. This solves the data bottleneck of "scarce samples, discontinuity, and large time intervals" that is common in engineering, and provides a high-quality training foundation for the model. In this embodiment of the invention, the physical mechanism-driven intelligent prediction method for building settlement optimizes the data fitting error and the degree of conformity with physical laws during model training, forcing the network to strictly adhere to known physical principles when mining statistical laws in the data, effectively avoiding predictions that violate physical common sense when there is insufficient data. In this embodiment of the invention, the physical mechanism-driven intelligent prediction method for building settlement provides no need for complex preprocessing (such as interpolation and resampling) of discontinuous, large-span intervals of original monitoring data. The data can be directly input into the model for training and prediction, greatly improving the practicality of engineering.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] Figure 1 The image shows the effect of the T-LSTM disclosed in this invention; Figure 2 This is a schematic diagram of a geological disaster occurrence disclosed in one embodiment of the present invention; Figure 3 This is a typical non-continuous data type data diagram of geological disasters disclosed in one embodiment of the present invention; Figure 4This is a structural diagram of the empirical formula method disclosed in one embodiment of the present invention; Figure 5 This is a diagram of an LSTM structure disclosed in one embodiment of the present invention; Figure 6 This is a basic flowchart of LSTM model prediction based on deep fusion of physical mechanisms disclosed in one embodiment of the present invention; Figure 7 This is a flowchart of a prediction process for an LSTM model based on deep fusion of physical mechanisms, as disclosed in one embodiment of the present invention. Figure 8 This is a comparison diagram of the data augmentation effects of physical mechanism constraints and interpolation methods disclosed in an embodiment of the present invention; Figure 9 This is a diagram illustrating the training process disclosed in one embodiment of the present invention; Figure 10 This is a graph showing the predicted values ​​of different models disclosed in one embodiment of the present invention; Figure 11 This is a comparison chart of measured and predicted values ​​disclosed in one embodiment of the present invention; Figure 12 This is a graph showing the predicted values ​​of different models disclosed in one embodiment of the present invention; Figure 13 This is a comparison chart of measured and predicted values ​​disclosed in one embodiment of the present invention; Figure 14 This is a diagram showing the physical constraint LSTM prediction results under different building data in one embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and specific examples and application scenarios.

[0022] In some embodiments of this invention, a physical mechanism-driven intelligent prediction method is proposed to address the discontinuous, large-span time intervals, and sample scarcity characteristics of engineering data. The core idea of ​​this method is to deeply integrate prior physical knowledge with a data-driven model, breaking through the rigid dependence of traditional neural networks on data continuity and sample size. First, this method constructs a settlement curve based on the physical mechanism of building settlement. Then, it uses limited, discontinuous actual monitoring data to drive the generation of a continuous virtual settlement sequence that conforms to physical laws, thus constructing a physical-driven virtual data generation mechanism. Next, it intelligently couples the sparse, discontinuous original measured data with the generated, physically consistent virtual continuous sequence, avoiding the problems of discontinuous time series and providing good supplementation for data with large time intervals. Finally, based on the standard data-driven loss function, it explicitly embeds constraints from the physical mechanism model, forcing the neural network to strictly adhere to known physical principles while learning statistical patterns in the data, avoiding predictions that violate physical common sense due to insufficient or discontinuous data.

[0023] In some embodiments of this invention, taking the settlement of buildings due to geological disasters in a certain city as an example, a typical dataset is used as training samples for learning, realizing the training and prediction of small-sample, large-span, discontinuous time series data. To demonstrate the superiority of the physical mechanism constraint module, a horizontal comparison is made between linear interpolation and traditional empirical formulas and the physical mechanism constraint LSTM model, and a vertical comparison is made between the effects of combining LSTM, GRU, RNN with the physical mechanism module. Finally, an urban geological disaster building settlement prediction model is constructed, providing a predictive means for urban safety.

[0024] Example 1:

[0025] This case study illustrates a large-scale ground subsidence disaster that occurred in a deep drilling area in northern China. The Ordovician limestone in this region, due to long-term intermittent deposition, underwent exposure, leaching, and weathering erosion, resulting in intense karstification of the upper Ordovician carbonate rocks, making it highly susceptible to the formation of potential caves and fissures. During drilling, the well encountered a large, latent cave in the Ordovician limestone at a depth of 1300m. Drilling fluid leakage caused pressurized water carrying soil into the cave, leading to large-scale ground subsidence centered on the wellbore, with a maximum subsidence of 4.87m at the center. The mechanism of this geological disaster is as follows: Figure 2 As shown.

[0026] The study area comprises a high-rise residential complex consisting of buildings I-X and VI, located immediately west of the geothermal well. The two buildings are separated only by a road, with the closest distance between the buildings and the geothermal well being approximately 130 meters. Due to ground subsidence, the high-rise buildings in this area experienced significant uneven settlement, leading to building tilting, wall cracking, and other phenomena. Within days of the disaster, Building III experienced a maximum settlement of approximately 1 meter, with a cumulative eastward tilt of 15.8‰.

[0027] Because the buildings in the study area are close to the wells where the disaster occurred, the strata in this area are greatly affected by geological disaster disturbances. This places stricter safety requirements on the repair and correction of damaged buildings. The stability of building settlement and deformation is an important prerequisite for the safe control of correction construction. Therefore, it is necessary to establish a predictive model based on existing monitoring data to predict the average daily settlement of buildings, thereby assessing the stability of buildings and providing strong support for the rational and accurate planning of repair design and implementation.

[0028] The following explains the discontinuous nature of the dataset in the example.

[0029] Settlement monitoring was conducted on 16 buildings within the study area for 22 months. The dataset exhibits significant characteristics of small sample size, large span, and discontinuous time series. Following the disaster, relevant departments immediately deployed monitoring points across the area for comprehensive monitoring. Due to localized ground subsidence exceeding 4 meters within a short period after the disaster, and significant settlement fluctuations in some buildings during the emergency response phase, continuous data updates were necessary. Severely affected buildings underwent daily monitoring for 3 months, while buildings in better condition underwent daily monitoring for 1 month. As the building settlement rate slowed, the measurement frequency gradually shifted to 3 days / time or 7 days / time, and later to 30 days / time, ensuring dynamic control of building safety while avoiding resource waste. Data analysis revealed that a small number of buildings had over 124 data points, while most buildings had fewer than 80 monitoring data points. Figure 3 As shown in type a. At the same time, by collecting various types of data in the project, the characteristics of the project data were sorted out and summarized: type b is characterized by a gradual change in the monitoring time interval from large to small; type c is characterized by non-continuous large-span time intervals; and type d is characterized by sudden emergency monitoring situations in the large-span time interval data.

[0030] This invention investigates three prediction methods: empirical formula method, standard LSTM neural network, and physical mechanism-constrained LSTM neural network. The three methods are described below.

[0031] Empirical formula methods are prediction methods based on engineering measured data to summarize the quantitative relationship between settlement and influencing factors. They achieve rapid settlement estimation by constructing mathematical models. The most common empirical formula methods are the hyperbolic method, the exponential curve method, and the Hoshino method. The hyperbolic method assumes that settlement develops in a hyperbolic trend over time, fitting parameters based on measured settlement-time data to estimate the final settlement. The exponential curve method assumes that the settlement process exhibits an exponential convergence characteristic with a rapid initial increase followed by gradual stabilization, and is often used to describe settlement scenarios with significant time-dependent effects, such as soft soil foundations. The Hoshino method combines the advantages of the hyperbolic and exponential curve methods, allowing for a more flexible fitting of the change patterns from the initial to the later transition stages of settlement, and is suitable for settlement prediction under complex foundation conditions. Empirical formula methods have significant limitations. They require high-quality data; quality defects in the data can significantly amplify prediction errors. They also require a certain amount of data; scarce data will affect prediction accuracy. Figure 4 As shown.

[0032] Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network (RNN) designed to address the long-term dependency problem encountered by standard RNNs when processing long sequence data. RNNs are advantageous in modeling sequential data (such as time series and natural language), but their gradients are prone to vanishing or exploding during backpropagation at time steps, making it difficult to learn the correlations between information at distant time steps.

[0033] like Figure 5 As shown. The core innovation of LSTM lies in its cell states and gating mechanism. Cell states act like a "conveyor belt" in a network, transmitting critical information over a relatively long period. The gating mechanism precisely controls the information flow, mainly including the forget gate: determining which historical information in the cell should be discarded; the input gate: controlling which new information should be updated into the cell; and the output gate: determining the output at the current moment based on the updated cell state.

[0034] Through this sophisticated gating design, LSTM can selectively retain long-term information, update new information, and output relevant results, greatly enhancing the model's ability to capture long-term dependencies in sequences. This has led to its widespread application and significant effectiveness in machine translation, speech recognition, and time series prediction. However, LSTM still has limitations in time series prediction. When time series have irregular intervals, missing data, or abrupt changes, the original LSTM struggles to directly capture temporal relationships between discontinuous points. Furthermore, it is prone to overfitting in small sample scenarios and requires a large amount of labeled data to learn robust temporal patterns.

[0035] This invention innovatively proposes an LSTM model framework based on deep integration of physical mechanisms. Through the organic combination of physical mechanism constraint modules and LSTM neural network modules, it delves into the intrinsic laws of physical processes to address the complex characteristics of building settlement caused by geological disasters, thereby achieving more accurate settlement analysis and prediction. Figure 6 As shown.

[0036] The core of the physical mechanism constraint module lies in fully utilizing empirical formulas universally applicable in the field of surface subsidence to construct a physical-driven computational engine suitable for geological disaster building subsidence analysis. First, let the complete time series be... To more precisely uncover the physical patterns in settlement data, it is necessary to divide it according to specific rules. Specifically, the complete time series should be rationally divided into m local time steps. ,in The partitioning needs to satisfy two conditions: first, the union of all local time steps equals the complete time series, i.e. First, to ensure that no data information is omitted; second, there is no overlap between any two different local time steps, that is... Ensure the independence and accuracy of data partitioning.

[0037] For each local time interval Based on the empirical formula for surface subsidence followed by the research system: ; in, for Settlement at any given time Given the set of parameters in the empirical formula, curve fitting techniques, including the least squares method, are used to solve the optimization problem: ; in, Observed settlement at any given time This represents the number of data points within a local time interval. A data curve conforming to the constraints of the physical mechanism is fitted. This curve not only accurately reflects the inherent physical laws of the settlement process, but also calculates the settlement displacement of buildings at each time step within a local time step, thereby constructing a continuous settlement change trajectory constrained by physical laws and generating a virtual scene data sequence. These virtual scene data fill the information gaps between discrete measured data based on physical mechanisms, effectively alleviating problems such as data sparsity, uneven distribution, and missing samples, and significantly improving the integrity and consistency of the data.

[0038] In the data fusion phase, rules for generating hybrid data are constructed: [The following is a list of rules and their implications.] for Real-time measured settlement data, for The physical mechanism that generates data at any given time will result in fused data. satisfy: ; This rule couples the original discrete real-world data with the virtual-world data generated by the physical mechanism, forming an augmented dataset that is continuous in the time dimension and self-consistent in physical logic. This dataset retains the authenticity of actual monitoring data while incorporating intermediate change processes that conform to physical laws through a physical mechanism constraint module.

[0039] In the LSTM neural network module, the physically augmented dataset is utilized. The LSTM neural network is trained using the data as input to the model. The LSTM neural network can deeply learn the dynamic patterns of building settlement caused by geological disasters, capture long-term dependencies in the data, and, combined with the inherent laws of settlement reflected by physical constraints, achieve more accurate predictions of future settlement trends. Let the enhanced data sequence be the input to the LSTM neural network. Output for the future Settlement prediction values ​​at each time step By minimizing the loss function between the predicted and actual values. ( For the future (Real-time settlement value), and optimize the weight parameters of the LSTM neural network using the backpropagation algorithm. This enables the model to accurately map the complex relationship between input data and future settlement trends, resulting in a physically constrained LSTM neural network applicable to actual building settlement prediction. This provides robust technical support for geological disaster early warning and building safety maintenance. Figure 7 As shown.

[0040] Figure 8 The figure compares the data augmentation effects of the physical mechanism constraint module and traditional linear interpolation. As can be seen from the figure, linear interpolation appears to better approximate the data distribution; however, due to the prevalent noise and potential uncertainties in the observed data, the augmented dataset produced by this method often introduces additional interference information, weakening its gain effect on model training. In contrast, the data augmentation of the physical mechanism constraint module better conforms to the variation patterns of settlement, introduces less noise interference, and achieves better augmentation results.

[0041] The results of the first three methods will be analyzed and explained below.

[0042] Model training and evaluation metrics: The mean absolute error (MAE), root mean square error (RMSE), and sample regression value (R²) were used. 2As performance evaluation metrics for models, MAE, RMSE, and R0 are used. 2 The calculation formulas are as follows: ; ; ; In the formula: N The number of data samples. Y i The predicted settlement value for existing buildings. Y i 'This represents the measured settlement value of the existing building.' Y This represents the average measured settlement value of existing buildings.

[0043] The physical mechanisms constraining the LSTM training process involve adjusting neuron weights and hyperparameters. Based on previous research, the range of hyperparameter selection has been determined. To determine the optimal hyperparameter values, this invention divides the dataset into 80% for model training and 20% for model testing, optimizing the parameters using a grid search method. The experiments will be conducted using the TensorFlow-GPU 2.3.0 deep learning framework in a Python environment, with the following hardware: i5 CPU, 32GB RAM, and an NVIDIA GeForce RTX 4060 graphics card.

[0044] The training process is demonstrated using the settlement data of Building III as an example. Figure 9 The figure shows the loss convergence process of the LSTM model on the training and validation sets due to the physical mechanism. In the figure, MSE is the mean squared error, and ep is the number of iterations.

[0045] To compare the reliability of physically constrained LSTM in solving discontinuous time series problems in the field of geological disasters, four models—LSTM (linear interpolation), T-LSTM, exponential curve method, and physically constrained LSTM—were selected for training and prediction on a typical discontinuous time series dataset with limited data and large spans in this geological disaster case, thereby comparing the prediction performance. The dataset contains 124 data points with varying time intervals that gradually increase in span, from 1 day / time to 3 days / time and then to 30 days / time.

[0046] After determining the structural parameters of LSTM (linear interpolation), T-LSTM, and physically constrained LSTM, each model was iteratively trained using the training set with its weights and thresholds to obtain the final model. The prediction results were then compared and analyzed using the test set. Figure 10 , 11As shown in Table 1, among the four methods, LSTM (linear interpolation) and physical mechanism-constrained LSTM outperform T-LSTM and the exponential curve method. Furthermore, compared to LSTM (linear interpolation), the physical mechanism-constrained LSTM model has lower MAE and RMSE, at 2.1325 mm and 2.5445 mm respectively, and its R... 2 The physical mechanism-constrained LSTM model achieved the highest score of 0.9951 among the four methods, indicating that it can capture long-term settlement patterns and achieve high-accuracy predictions. LSTM (linear interpolation) performed slightly worse than the physical mechanism-constrained LSTM, showing a significant deviation trend in later data sets. This is because linear interpolation suffers from larger errors and noise when dealing with data over longer time intervals, lacking a built-in physical mechanism to constrain data accuracy, thus failing to effectively simulate dynamic changes in discontinuous time series data. The T-LSTM model and the exponential curve method performed poorly, with the exponential curve method achieving an RMSE of 20.9411 mm, a MAE of 30.565 mm, and an R... 2 The value is 0.6255. The RMSE of the T-LSTM model is 69.1153 mm, the MAE is 336.9276 mm, and the R... 2 The result was -2.601. The poor performance of both methods indicates that traditional empirical formulas and T-LSTM have significant shortcomings and large errors when dealing with settlement datasets with unique characteristics in the field of geological disasters. Based on the predicted data, a 0.05% error was used as the standard for classifying the prediction excellence rate; data with a relative error of less than 0.5% between the predicted and measured values ​​were considered excellent. Compared to the exponential curve method, T-LSTM, and LSTM (linear interpolation), the physically constrained LSTM showed a significant lead in the prediction excellence rate, reaching 98.24%, and its maximum absolute error was also much smaller than the other three methods, at only 6.305 mm. This demonstrates that through physical constraints, neural networks can achieve better fitting accuracy and prediction performance when dealing with small datasets and large-span discontinuous time series data, enabling more accurate prediction of geological disaster settlement.

[0047] Table 1 Prediction Error Table

[0048] To demonstrate the superiority of the physical mechanism constraint combined with the LSTM model, two common methods were selected for comparison: RNN and GRU combined with physical mechanism constraints. The number of input parameter units, the number of output parameter units, and the number of training iterations were kept consistent with the physical mechanism constraint LSTM model in both comparison models. The remaining hyperparameters were selected through optimization algorithms, and the final hyperparameter values ​​for each comparison model are shown in Table 2.

[0049] Table 2 Model Hyperparameter Selection Table

[0050] The prediction results of each model are as follows Figure 12 , 13 As shown in the figure, LSTM performs well on the entire test set, while the predicted values ​​of GRU and RNN gradually deviate from the measured values ​​after day 215, with the error continuously increasing. This indicates that LSTM can more accurately capture the changing trend of landslide subsidence, ensuring more accurate predictions in later stages. Table 3 shows that the R-values ​​of the three methods... 2 All three methods achieved an accuracy above 0.85, with LSTM performing best at 0.9951. This indicates that all three methods effectively reflect the development trend and patterns of building settlement, with LSTM showing the best fit. Comparing the RMSE indices of LSTM, GRU, and RNN, which were 2.5445 mm, 4.076 mm, and 7.9368 mm respectively, it can be seen that the LSTM model exhibits smaller deviations from the measured values, demonstrating superior accuracy and robustness. Based on the predicted data, a 0.05% error was used as the standard for classifying the prediction excellence rate; data with a relative error of less than 0.5% between the predicted and measured values ​​were considered excellent, as shown in Table X. The prediction excellence rates of the LSTM, GRU, and RNN models were all above 85%, but only LSTM exceeded 90%, reaching 98.24%, meaning that almost all predicted and measured errors were controlled within 0.5%. Comparing the mean absolute error and maximum absolute error, the LSTM model also shows significant advantages, with a mean absolute error of 2.1325 mm and a maximum absolute error of 6.305 mm. This indicates that the LSTM model has the smallest prediction error and the highest accuracy.

[0051] Table 3 Prediction Error Table

[0052] The following section explains the prediction and determination of building settlement.

[0053] According to the specifications, a building is considered to have reached a stable settlement state when the maximum settlement rate in the last 100 days is less than 0.04 mm / d. Based on the building's tilt and other factors, a grading standard for settlement stabilization rates was established: Level 1 stability index is a settlement rate less than 0.01 mm / d, Level 2 stability index is a settlement rate less than 0.02 mm / d, and Level 3 stability index is a settlement rate less than 0.04 mm / d. This standard was used to predict the stabilization time. A physical constraint mechanism LSTM model was applied to the settlement monitoring of three buildings (Buildings VII, VIII, and IX), and the settlement development of the three buildings over a period of 180 days, from 540 days to 720 days after the accident, was predicted and compared with the latest measured data for verification. Figure 14 As shown, the settlement of the buildings all exhibits a slowing trend. The predicted MAE and RMSE for the three buildings are 0.0882-0.0328 mm and 0.1061-0.0527 mm, respectively, indicating that the model's predictions have small errors and high accuracy, making them valuable for reference. Both the predicted and measured settlement values ​​for the three buildings show that buildings VIII and IX have reached the level three stability standard, while building VII has not yet reached the level three stability standard, with a settlement rate of 0.05 mm / d, which is greater than 0.04 mm / d.

[0054] Furthermore, using the physical mechanism constrained by the LSTM model proposed in this invention, a 306-day prediction was made for the settlement development of buildings VII, VIII, and IX. Figure 14 As shown in the figure, the predicted data indicates that Building VII will reach Level III stability standards 958 days after the accident, Building VIII will meet Level II stability standards 886 days after the accident, and Level I stability standards 997 days after the accident, while Building IX will meet Level I stability standards 829 days after the accident. This invention suggests that the design and construction of the building repairs can be carried out according to the above timelines, ensuring that the building settlement development fully meets the requirements while guaranteeing timely commencement of repair work.

[0055] This invention employs a physical mechanism to constrain and optimize LSTM, proposing a predictive model to address the challenge of handling small-sample, large-span, discontinuous time series data in engineering fields. The model is validated through a typical case study of urban geological disasters. A comparative analysis is conducted on the model's performance in processing discontinuous time series data, comparing it horizontally with traditional interpolation methods, T-LSTM, and empirical formula methods. Furthermore, a longitudinal comparative analysis is performed using RNN and GRU models constrained by the physical mechanism and this model. The main research results and discussion are summarized below: (1) Monitoring data in engineering fields such as urban geological disasters, tunnels, and foundation pits are characterized by "discontinuity", "large monitoring intervals", and "scarce data". The physical mechanism constraint proposed in this invention, as virtual scene data, combined with measured data, can effectively capture settlement patterns and improve the reliability of neural network prediction.

[0056] (2) With the combination of physical mechanism constraint modules, neural network models (LSTM / GRU / RNN) significantly improved the subsidence prediction capability for data with limited data, large spans, and discontinuous time series. Among them, the physical mechanism constraint LSTM performed best (MAE=2.1325, RMSE=2.5445, R 2 =0.9951), its accuracy and stability surpass traditional methods such as LSTM (linear interpolation), T-LSTM, and exponential curve methods. Further experiments show that LSTM has the highest overall performance within this framework and is more suitable for fields requiring high prediction accuracy.

[0057] (3) The physical mechanism-constrained LST method is applicable to civil engineering safety monitoring scenarios such as bridge settlement and foundation pit deformation. Its innovation lies in: intelligently coupling actual measurements and virtual sequences to generate high-quality datasets, solving problems such as sample scarcity; taking into account data fitting and physical laws during training to avoid unreasonable predictions; and eliminating the need for complex preprocessing of raw data, thus improving practicality. This method provides a solution for predicting discontinuous, large-span interval data, and strongly supports engineering repair and stability assessment work.

[0058] Any part of this invention not described in detail can be referred to in the prior art or in the art known to those skilled in the art. This embodiment does not limit such part and will not describe it in detail here.

[0059] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A physical mechanism-driven intelligent prediction method for building settlement, characterized in that, The method includes: S1. Collect building settlement data; S2. Determine the empirical formula, and use the empirical formula to fit and augment the building settlement data to obtain settlement data constrained by physical mechanisms; S3. Use the obtained physical mechanism-constrained settlement data to train the LSTM neural network to obtain a physical mechanism-constrained LSTM model; S4. Input the settlement data of the building to be predicted into the LSTM model constrained by the physical mechanism to predict the future settlement of the building.

2. The physical mechanism-driven intelligent prediction method for building settlement according to claim 1, characterized in that, Step S2 includes: S21. Based on empirical formulas and collected building settlement data, perform data fitting and determine the fitting coefficients; S22. Obtain several virtual scene settlement data based on the fitting coefficients; S23. The obtained virtual scene settlement data and the collected real scene building settlement data are fused together to obtain settlement data constrained by physical mechanisms.

3. The physical mechanism-driven intelligent prediction method for building settlement according to claim 2, characterized in that, The empirical formula is a formula based on the hyperbola method, the exponential curve method, or the Hoshino method.

4. The intelligent prediction method for building settlement driven by physical mechanisms according to claim 1, characterized in that, The empirical formula is expressed as follows: ； in, for Settlement at any given time Given the set of parameters in the empirical formula, curve fitting techniques, including the least squares method, are used to solve the optimization problem: ; in, Observed settlement at any given time This represents the number of data points within a local time interval.

5. The physical mechanism-driven intelligent prediction method for building settlement according to claim 2, characterized in that, The rules for generating mixed data during fusion in step S23 are as follows: for Real-time measured settlement data, for The virtual scene settling data at any given time is then fused into the final data. satisfy: 。 6. The physical mechanism-driven intelligent prediction method for building settlement according to claim 1, characterized in that, Step S3 includes: The input to an LSTM neural network is an enhanced data sequence. Output for the future Settlement prediction values ​​at each time step ; By minimizing the loss function between the predicted and actual values, the weight parameters of the LSTM neural network are optimized using the backpropagation algorithm. This enables the model to accurately map the complex relationship between input data and future subsidence trends, resulting in a physical mechanism-constrained LSTM neural network.

7. The physical mechanism-driven intelligent prediction method for building settlement according to claim 6, characterized in that, The loss function between the predicted and actual values ​​is: ； in, For the future The actual settlement value at any given moment.

8. The physical mechanism-driven intelligent prediction method for building settlement according to claim 6, characterized in that, During the training of LSTM neural networks, physical constraints, including the requirement that the settlement rate must not change abruptly and must eventually stabilize within the specified range, are explicitly embedded to avoid unreasonable predictions, such as reverse growth of settlement, when there is insufficient data.

9. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the intelligent prediction method for building settlement driven by the physical mechanism described in any one of claims 1-8.

10. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the physical mechanism-driven intelligent prediction method for building settlement according to any one of claims 1-8.

Images