Land subsidence prediction method based on feature selection and gwo optimization hybrid model

Abstract

This invention belongs to the field of surface settlement monitoring technology, and relates to a surface settlement prediction method based on a hybrid model of feature selection and GWO optimization. The method includes: acquiring monitoring data of a target foundation pit project; performing Pearson correlation analysis on the monitoring data, and selecting monitoring points of other monitoring items whose correlation with the target predicted surface settlement point is greater than a set threshold as key feature inputs; the monitoring points of the other monitoring items include at least: pile top vertical displacement monitoring points, settlement monitoring points of surrounding buildings, and settlement monitoring points of underground pipelines around the foundation pit; inputting the key features from the monitoring data into a CNN-GRU-Attention model based on an improved GWO optimization to predict the surface settlement value of the target surface settlement monitoring point. The method and system provided by this invention can accurately and efficiently predict future surface settlement of foundation pits, facilitating effective monitoring of surface settlement and early warning of engineering problems associated with surface settlement.

Description

Technical Field

[0001] This invention belongs to the field of surface subsidence monitoring technology, and relates to a method for predicting surface subsidence based on a hybrid model of feature selection and GWO optimization. Background Technology

[0002] Surface settlement of foundation pits is a common engineering geological hazard, with a wide impact range, high difficulty in control, and complex and diverse inducing factors. Surface settlement of foundation pits not only causes direct damage to the foundation pit support structure and surrounding buildings, but may also trigger secondary disasters such as pipeline rupture and road collapse, leading to serious engineering accidents and economic losses. Therefore, accurate prediction of surface settlement of foundation pits is of great significance.

[0003] Surface settlement of foundation pits is caused by the combined effects of multiple factors, the main reasons of which include: the drop in groundwater level caused by dewatering or excavation of the foundation pit, leading to an increase in effective stress in the soil and a decrease in pore water pressure, thereby intensifying soil compression and consolidation and triggering settlement; the unloading of the soil on the pit walls and bottom during excavation, and the stress release causing the surrounding soil to shift into the pit, resulting in surface subsidence; deformation or failure of the support structure can further lead to instability of the surrounding soil and increase the risk of settlement; external factors such as construction vibration, changes in surrounding loads, and rainfall infiltration can also affect the mechanical properties of the soil and exacerbate the development of settlement. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for predicting surface settlement based on a hybrid model of feature selection and GWO optimization, which can accurately and efficiently predict future surface settlement of foundation pits, facilitate effective monitoring of surface settlement and early warning of engineering problems associated with surface settlement.

[0005] To achieve the above objectives, the present invention provides the following solution: Land subsidence prediction methods based on a hybrid model of feature selection and GWO optimization include: Obtain monitoring data for the target foundation pit project; Pearson correlation analysis was performed on the monitoring data, and monitoring points of other monitoring items with a correlation greater than a set threshold with the target predicted surface settlement point were selected as key feature inputs; the monitoring points of the other monitoring items included at least: vertical displacement monitoring points at the pile top, settlement monitoring points of surrounding buildings, and settlement monitoring points of underground pipelines around the foundation pit. Key features from the monitoring data are input into a CNN-GRU-Attention model based on an improved GWO optimization to predict the surface subsidence value of the target surface subsidence monitoring point. The improved GWO employs a hierarchical concatenated coding scheme, representing the individual position of a "gray wolf" as a hybrid vector, which is composed of three sequentially concatenated parts: (1) CNN encoding segment: length is ;in This is the maximum allowed number of convolutional layers: First dimension: a dimension in Integers within the range represent the actual number of convolutional layers used. ; Subsequent dimension Each dimension is an integer representing the number of convolutional kernels at the corresponding level; Further Dimension: Each dimension is an integer mapped from a discrete set, representing the size of the convolutional kernel at the corresponding level; (2) GRU encoded segment: length is ;in This is the maximum allowed number of GRU layers; First dimension: Integer, representing the number of GRU layers. ; Subsequent Dimension: An integer representing the number of hidden units in each layer of the GRU; (3) Attention encoding segment: length 2; The first dimension: a discretized integer value used to index a predefined set of attention types; The second dimension: an integer representing the attention weight dimension. .

[0006] Furthermore, the improved GWO also employs a probability-based perturbation update algorithm for position updates, specifically including: (1) Leadership probability calculation: based on , , Wolves in Dimensions The value on the table is used to calculate the individual. The attractiveness probability and random probability of learning from each leader wolf: ; ; ; ; in, and It is a coefficient vector. It is a function that maps distance and coefficients to probabilities; (2) Probability sampling update: based on the calculated probability distribution Random sampling is performed. If you are selected , or ,but Take the value of the corresponding leader wolf in this dimension; If you are selected If the parameter is selected, a new value will be randomly generated within the valid domain of the parameter.

[0007] Furthermore, when updating the attention type parameter, if learning from the leader wolf is selected, its category is directly copied; if random exploration is selected, it is uniformly and randomly selected from the entire category set.

[0008] In a second aspect, the present invention also provides a computer system, comprising: 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 land subsidence prediction method based on feature selection and GWO optimized CNN-GRU-Attention model.

[0009] Compared with existing models, this invention comprehensively considers the collaborative working relationship between the foundation pit support structure, the surrounding soil, and adjacent buildings during the foundation pit excavation process. In order to further deepen the close relationship between each monitoring item and surface settlement, the Pearson correlation analysis method is used to calculate the correlation degree between each monitoring item and surface settlement, and the most relevant monitoring point among the monitoring items with high correlation degree is selected as the key feature input. Furthermore, this invention improves the general GWO (Gross Flow of Words) by including a hierarchical concatenated coding scheme and a probability-based perturbation update algorithm. Six prediction models were established: unimproved GWO with no key features in the input, unimproved GWO with key features in the input, improved GWO with no key features in the input but with improved model hyperparameters and network structure, improved GWO with key features in the input and with improved model hyperparameters and network structure, improved GWO with no key features in the input but with standard GWO optimization, and improved GWO with key features in the input and with standard GWO optimization. The impact of whether the input contains key features and whether the improved GWO is used to optimize the model hyperparameters and network structure was compared and analyzed. Finally, the optimal prediction model was selected to achieve accurate prediction of surface subsidence, which is more conducive to the monitoring and early warning of surface subsidence in foundation pit engineering. Attached Figure Description

[0010] Figure 1 A flowchart of a land subsidence prediction method based on feature selection and GWO optimization of the CNN-GRU-Attention model provided in an embodiment of the present invention; Figure 2 Comparison chart of the predicted settlement values ​​of the first model, third model, and fifth model provided in the embodiments of the present invention with the actual settlement values; Figure 3A comparison chart of the predicted settlement values ​​of the second, fourth, and sixth models provided in the embodiments of the present invention with the actual settlement values; Figure 4 A comparison chart of the surface subsidence prediction results of six different models provided in the embodiments of the present invention; Figure 5 A comparison chart of surface subsidence prediction results for the test set provided in this embodiment of the invention; Figure 6 An internal structure diagram of a computer system provided for an embodiment of the present invention. Detailed Implementation

[0011] 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 embodiments of the present invention, and not all embodiments. 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.

[0012] The purpose of this invention is to provide a method and system for predicting surface settlement based on feature selection and GWO-optimized CNN-GRU-Attention model, which can accurately predict the future surface settlement of foundation pits and facilitate effective monitoring and early warning of surface settlement generated during foundation pit excavation.

[0013] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0014] Example 1: This example provides a method for predicting land subsidence based on a hybrid model of feature selection and GWO optimization, such as... Figure 1 As shown, the specific steps of this method are as follows: Step S1: Obtain historical monitoring data for the target foundation pit project. This historical monitoring data mainly includes: surface settlement, vertical displacement of retaining piles, horizontal displacement of retaining piles, and settlement of surrounding underground pipelines. Additionally, [the data is then used to] monitor the previous [projects / data]. The cumulative displacement can be obtained by summing the displacements of each measurement.

[0015] Step S2: Perform Pearson correlation analysis on historical monitoring data to determine key feature inputs, specifically including: The first step is to categorize the historical monitoring data and determine the various monitoring items. The second step is to use Pearson correlation analysis to calculate the correlation between the target prediction monitoring point and each of its nearby monitoring points, and to obtain the Pearson correlation coefficient for different monitoring points. The third step is to identify monitoring points with a Pearson correlation coefficient greater than or equal to a set threshold as key feature inputs. In this embodiment, the set threshold is 0.5, meaning that monitoring points with a Pearson correlation coefficient greater than or equal to 0.5 are identified as key feature inputs. The key feature inputs finally identified in this embodiment include: ZQC27, ZQS28, GXC6-7, and DBC28-1.

[0016] In this embodiment, the Pearson correlation coefficient threshold is set to 0.5. The threshold is determined based on the general understanding of correlation strength in statistics: generally, a Pearson correlation coefficient absolute value above 0.5 indicates a moderate to strong correlation, reflecting a meaningful association between variables. In the field of geotechnical engineering monitoring, 0.5 is a common empirical threshold used to screen monitoring indicators that have a significant impact on target variables (such as surface subsidence). In practical applications, the threshold can be adjusted within the range of 0.4-0.6 according to the characteristics of the project.

[0017] Step S3: Filter historical data for key feature input based on historical monitoring data.

[0018] In this embodiment, since the historical monitoring data also includes monitoring items such as foundation pit pipeline settlement, it is necessary to select the historical data corresponding to the key feature input from the historical monitoring data based on the key feature input, and divide the selected historical data into training set, test set and validation set in a ratio of 8:1:1, that is, 80% for training, 10% for testing and 10% for validation, so as to facilitate the later model training and optimization.

[0019] Step S4: Create the first, second, third, fourth, fifth, and sixth models based on the CNN-GRU-Attention model. The first model's input does not contain key features and is not optimized using GWO. The second model's input contains key features but is not optimized using GWO. The third model's input does not contain key features but its hyperparameters and network structure are optimized using GWO. The fourth model's input contains key features and its hyperparameters and network structure are optimized using GWO. The fifth model's input does not contain key features but the model is optimized using standard GWO. The sixth model's input contains key features and the model is optimized using standard GWO.

[0020] In this embodiment, step S4 is as follows: The first step is to build six CNN-GRU-Attention base models; The second step is to use one of the aforementioned basic models as the first model; The third step is to change the input of one of the base models to the key feature inputs to obtain the second model; The fourth step is to optimize the hyperparameters and network structure of the aforementioned basic model using the improved GWO to obtain the third model; The fifth step involves changing the input of one of the base models to key feature inputs and optimizing the model hyperparameters and network structure using the improved GWO to obtain the fourth model. The sixth step is to optimize the aforementioned basic model using the standard GWO to obtain the fifth model; The seventh step involves changing the input of one of the base models to key feature inputs and optimizing the model using standard GWO to obtain the sixth model.

[0021] The input method, hyperparameters, and network structure of the CNN-GRU-Attention model, and whether they are optimized using the improved GWO, significantly impact the accuracy and efficiency of the model's predictions. When key features are used as input, more complete contextual information can be provided for predicting future land subsidence. The GRU neural network captures the long-term dependencies between land subsidence and various key features through memory units and gating mechanisms, enabling the model to better understand long-term dependencies in time series. Furthermore, compared to the standard GWO, the improved GWO can better optimize the model's hyperparameters and network structure, allowing the model to find optimal parameters more quickly, better fit the training data, and better extract local spatial features, thereby improving the model's prediction accuracy and efficiency.

[0022] The standard GWO individual position is a continuous vector of fixed length, while the architecture of the CNN-GRU-Attention model consists of parameters of different properties and variable lengths. To enable GWO to optimize the number of CNN convolutional layers and the size and number of filters, the number of GRU layers and the number of hidden units per layer, and the type and weights of the attention mechanism in the CNN-GRU-Attention model, this invention makes the following improvements to the existing general standard GWO: A hierarchical concatenated encoding scheme is designed to represent the individual position of a "gray wolf" as a hybrid vector, which is composed of three sequentially concatenated parts: (1) CNN encoding segment: length is ;in This is the maximum allowed number of convolutional layers: First dimension: a dimension in Integers within the range represent the actual number of convolutional layers used. ; Subsequent Dimension: Each dimension is an integer representing the number of convolutional kernels at the corresponding layer; in practice, only the first... One dimension is valid, and the other dimensions are ignored during decoding.

[0023] Further Dimension: Each dimension is an integer mapped from a discrete set (e.g., 1→3, 2→5, 3→7), representing the convolutional kernel size of the corresponding layer; similarly, only the first... It is effective.

[0024] (2) GRU encoded segment: length is ;in This is the maximum allowed number of GRU layers; First dimension: Integer, representing the number of GRU layers. ; Subsequent Dimension: An integer representing the number of hidden units in each layer of the GRU; (3) Attention encoding segment: length 2; The first dimension: a discretized integer value used to index a predefined set of attention types (e.g., 0 → "dot-product", 1 → "additive", 2 → "self-attention"...). The second dimension: an integer representing the attention weight dimension. .

[0025] For example: Suppose that the decoded part of an individual's position vector is: CNN segment = [3, 64, 128, 64, 5, 3, 5], GRU segment = [2, 128, 64], and Attention segment = [2, 32]. This means: The CNN has 3 layers with 64, 128, and 64 channels respectively, and kernel sizes of 5×5, 3×3, and 5×5 respectively. The GRU has 2 layers with 128 and 64 hidden units respectively. The attention mechanism type is "self-attention" corresponding to index 2, with a weight dimension of 32.

[0026] Standard GWO position update formula This can produce continuous values, which, when directly applied to integer or categorical parameters in this invention, result in a large number of invalid intermediate values. Therefore, this invention abandons direct arithmetic averaging and designs a probability-based perturbation update operator. For individual... Each dimension Its update no longer relies on continuous vector operations, but generates new discrete values ​​through the following steps: (1) Leadership probability calculation: based on , , Wolves in Dimensions The value on the table is used to calculate the individual. The attractiveness probability and random probability of learning from each leader wolf: ; ; ; ; in, and It is a coefficient vector. It is a function that maps distance and coefficients to probabilities; (2) Probability sampling update: based on the calculated probability distribution Perform random sampling; If you are selected , or ,but Take the value of the corresponding leader wolf in this dimension; If you are selected If the parameter is selected, a new value will be randomly generated within the valid domain of the parameter.

[0027] (3) For purely discrete category parameters such as “attention type”, its domain is an unordered set. During the update, if learning from the leader wolf is selected, its category is directly copied; if random exploration is selected, it is uniformly and randomly selected from the entire category set.

[0028] Step S5: Using historical monitoring data from the previous period The time-based surface settlement value is used as the input, based on the historical monitoring data of the [number]th [time]. The surface subsidence value at time 1 is output, and the first, third, and fifth models are trained and optimized respectively to determine the optimal model among the first, third, and fifth models; among them, In this embodiment, the model training in step S5 uses the Adam optimizer with an initial learning rate of 0.001 and 200 training rounds. The early stopping mechanism is set to stop training if the loss on the validation set does not decrease for 10 consecutive rounds.

[0029] In this embodiment, step S5 is as follows: The first step is to analyze the historical monitoring data from the previous period. Time and the The surface subsidence value at time t is used to train and optimize the first and third models respectively, and the mean absolute error, root mean square error and coefficient of determination of the first, third and fifth models are calculated respectively. The second step is to select the model with the smallest mean absolute error, root mean square error, and largest coefficient of determination as the better model among the first, third, and fifth models.

[0030] Step S6: Prioritize key features Previous in historical data and historical monitoring values The time-based surface settlement value is used as the input, based on the historical monitoring data of the [number]th [time]. The surface subsidence value at time 1 is output, and the second, fourth, and sixth models are trained and optimized respectively to determine the optimal model among the second, fourth, and sixth models. .

[0031] In this embodiment, step S6 is as follows: The first step is to base the analysis on key features. The previous i-th time and the historical data and historical monitoring values The surface subsidence value at time t is used to train and optimize the second, fourth, and sixth models respectively, and the mean absolute error, root mean square error, and coefficient of determination of the second, fourth, and sixth models are calculated respectively. The second step is to select the model with the smallest mean absolute error, root mean square error, and largest coefficient of determination as the better model among the second, fourth, and sixth models.

[0032] Step S7: Compare the evaluation index values ​​of the better models in the first, third, and fifth models and the better models in the second, fourth, and sixth models, and determine the final land subsidence prediction model based on the evaluation index values; wherein, the evaluation index values ​​include: mean absolute error value, root mean square error value, and coefficient of determination value.

[0033] In this embodiment, step S7 is as follows: The first step is to calculate the mean absolute error, root mean square error, and coefficient of determination for the first, second, third, fourth, fifth, and sixth models after training and optimization, respectively. The second step involves selecting the prediction model with the smallest mean absolute error, root mean square error, and largest coefficient of determination as the final land subsidence prediction model. The input terms of the final determined land subsidence prediction model contain key features, and the model undergoes improved GWO optimization.

[0034] Step S8: Collect the current key feature data of the target foundation pit project and input them into the surface settlement prediction model to determine the future surface settlement value of the target foundation pit project.

[0035] Based on the above analysis, the surface settlement prediction method provided in this embodiment has high accuracy in predicting surface settlement caused by foundation pit excavation. It can predict future changes in surface settlement in real time and provides efficient and accurate technical support for foundation pit support and engineering accident prevention.

[0036] Example 2: Taking the foundation pit project of Huicheng Road Station on Qingdao Metro Line 7 as an example, the specific implementation is as follows:

[0037] The foundation pit for Huicheng Road Station on Qingdao Metro Line 7 is 223m long, 20.1-24m wide, and 18.5-23m deep. The soil structure, from top to bottom, consists of plain fill, silty clay, strongly weathered coarse andesite, moderately weathered coarse andesite, and slightly weathered coarse andesite.

[0038] A total of 110 surface settlement monitoring points, 225 pipeline settlement monitoring points, and 28 pile top horizontal and vertical displacement monitoring points were set up around the foundation pit, with the codes DBC (surface settlement), GXC (pipeline settlement), ZQS (pile top horizontal displacement), and ZQC (pile top vertical displacement), respectively.

[0039] (1) Data collection and processing of surface subsidence.

[0040] The study focused on the measured data from monitoring point DBC27-1, which was collected from October 1, 2023 to April 2, 2024, with data collected once a day, for a total of 185 measurements.

[0041] Use PyCharm programming to filter and perform calculations on data.

[0042] (2) Determine the number of key feature inputs.

[0043] The impact of foundation pit excavation on the surrounding soil is multifaceted and complex. This invention comprehensively considers the collaborative working relationship between the foundation pit support structure, the surrounding soil, and adjacent buildings during the foundation pit excavation process. Pearson correlation analysis is used to perform Pearson correlation analysis between monitoring point DBC27-1 and its surrounding monitoring points. The calculation results are shown in Table 1. In this embodiment, the threshold is set to 0.5, meaning that monitoring points with a Pearson correlation coefficient greater than or equal to 0.5 are identified as key feature inputs (one monitoring point is selected for the same monitoring project). Since GXC6-6 and GXC6-7 are settlement monitoring points at different locations along the same pipeline, and GXC6-7 has a larger correlation coefficient, ZQC27, ZQS28, GXC6-7, and DBC28-1 are used as key feature inputs.

[0044] Table 1. Pearson correlation coefficients between DBC27-1 and surrounding monitoring points. monitoring points Correlation coefficient ZQC27 -0.547 ZQC28 -0.093 ZQS27 0.030 ZQS28 -0.638 GXC6-6 0.909 GXC6-7 0.943 DBC27-3 0.599 DBC28-1 0.909 .

[0045] (3) Set key feature input, GWO optimization and sliding window.

[0046] Considering the multifaceted impacts of foundation pit excavation on the soil, including surface settlement and soil unloading that could affect retaining piles and underground pipelines, the CNN-GRU-Attention model requires significant time to find optimal parameters during prediction. Therefore, six models were established: input items containing key features without GWO optimization, input items containing key features without GWO optimization, input items without key features but optimized with improved GWO optimization, input items containing key features but optimized with improved GWO optimization, input items without key features but optimized with standard GWO optimization, and input items containing key features and optimized with standard GWO optimization. Six prediction models were created: input items without key features (unoptimized by GWO), input items with key features (unoptimized by GWO), input items without key features (optimized by improved GWO), input items with key features (optimized by improved GWO), input items without key features (optimized by standard GWO), and input items with key features (optimized by standard GWO). These were named Model 1, Model 2, Model 3, Model 4, Model 5, and Model 6. The impact of whether the input items contain key features and whether improved GWO is used on the optimization of model hyperparameters and network structure on the prediction results was analyzed. The differences between the different models are shown in Table 2.

[0047] Table 2 Comparison of different models Model types Does the input contain key features? Optimization methods First Model no Unoptimized Second Model yes Unoptimized Third Model no Improved GWO optimization Fourth Model yes Improved GWO optimization Fifth Model no Standard GWO optimization Sixth Model yes Standard GWO optimization .

[0048] Based on historical monitoring data The time-based surface settlement value is used as the input, based on the historical monitoring data of the [number]th [time]. The surface subsidence value at time 1 is output. Models 1, 3, and 5 are trained and optimized using the Adam optimizer with an initial learning rate of 0.001 and 200 training epochs. An early stopping mechanism is set to stop training if the loss on the validation set does not decrease after 10 consecutive epochs. Finally, the prediction results of the optimized models 1, 3, and 5 are shown in [link to prediction results]. Figure 2 As shown.

[0049] according to Figure 2 The results in Table 3 show that the third model, with inputs containing key features, achieves the minimum mean absolute error (MAE) and root mean square error (RMSE), and the coefficient of determination R0 is also the lowest. 2 It reached a maximum of 0.8516.

[0050] Table 3. Surface subsidence prediction and evaluation index values ​​for the first, third, and fifth models. Sliding window RMSE MAE <![CDATA[R 2 ]]> First Model 0.8235 0.7926 0.6996 Third Model 0.5787 0.5269 0.8516 Fifth Model 0.6134 0.5467 0.8334 .

[0051] Using key features and land subsidence values ​​as model inputs, and land subsidence values ​​as model outputs, the second, fourth, and sixth models were trained and optimized. The Adam optimizer was used for model training with an initial learning rate of 0.001 and 200 training epochs. An early stopping mechanism was set to stop training if the loss on the validation set did not decrease after 10 consecutive epochs. Finally, the prediction results of the trained and optimized second, fourth, and sixth models are shown below. Figure 3 As shown.

[0052] according to Figure 3 The results in Table 4 show that the fourth model has the lowest mean absolute error (MAE) and root mean square error (RMSE), and the coefficient of determination R0 is also the lowest. 2 It reached a maximum of 0.9568.

[0053] Table 4. Surface subsidence prediction and evaluation index values ​​for the second, fourth, and sixth models. Sliding window RMSE MAE <![CDATA[R 2 ]]> Second Model 0.4361 0.4112 0.9158 Fourth Model 0.3125 0.3072 0.9568 Sixth Model 0.4174 0.355 0.9228 .

[0054] The above analysis shows that the fourth model, combining key feature inputs and improved GWO optimization of model hyperparameters and network structure, has sufficient data for the model to learn the sudden fluctuation trends of sedimentation data, capture long-term dependencies and overall trends in the sequence, and, under the action of the Grey Wolf optimization algorithm, determines the optimal training parameters of the model, the number of convolutional layers and the size and number of filters in the CNN, the number of GRU layers and the number of hidden units in each layer, and the type and weight of the attention mechanism, exhibiting the highest prediction accuracy. Figure 4 The results in Table 5 show that the model with the fourth model set as the key feature input and the improved GWO optimization model hyperparameters and network structure has the best evaluation index and the best prediction result.

[0055] Table 5. Evaluation index values ​​for surface subsidence prediction of six models. category RMSE MAE <![CDATA[R 2 ]]> First Model 0.8235 0.7926 0.6996 Second Model 0.4361 0.4112 0.9158 Third Model 0.5787 0.5269 0.8516 Fourth Model 0.3125 0.3072 0.9568 Fifth Model 0.6134 0.5467 0.8334 Sixth Model 0.4174 0.355 0.9228 .

[0056] (4) Model validation

[0057] Following the aforementioned method for predicting land subsidence based on feature selection and an improved GWO-optimized CNN-GRU-Attention model, the fourth model was used to continuously predict test set data. The prediction results are shown below. Figure 5 .according to Figure 5 It can be seen that the prediction results under the model conditions are close to the actual surface subsidence values, and the prediction accuracy is high, thus verifying the effectiveness of the surface subsidence prediction method based on feature selection and the improved GWO optimized CNN-GRU-Attention model of the present invention.

[0058] Example 3: This example provides a computer system, which can be a server or a terminal, and its internal structure diagram can be as follows. Figure 6 As shown, the computer system includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores video tag processing data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the land subsidence prediction method of Embodiment 1.

[0059] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0060] The databases involved in the various embodiments provided by this invention may include at least one of relational databases and non-relational databases. Non-relational databases may include, but are not limited to, distributed databases based on blockchain. The processors involved in the various embodiments provided by this invention may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these. The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably.

Claims

1. A method for predicting land subsidence based on a hybrid model of feature selection and GWO optimization, characterized in that, include: Obtain monitoring data for the target foundation pit project; Pearson correlation analysis was performed on the monitoring data, and monitoring points of other monitoring items whose correlation with the target predicted surface subsidence point was greater than a set threshold were selected as key feature inputs. The monitoring points for the other monitoring items include at least: vertical displacement monitoring points at the pile top, settlement monitoring points for surrounding buildings, and settlement monitoring points for underground pipelines around the foundation pit; Key features from the monitoring data are input into a CNN-GRU-Attention model based on an improved GWO optimization to predict the surface subsidence value of the target surface subsidence monitoring point. The improved GWO employs a hierarchical concatenated coding scheme, representing the individual position of a "gray wolf" as a hybrid vector, which is composed of three sequentially concatenated parts: (1) CNN encoding segment: length is ;in This is the maximum allowed number of convolutional layers: First dimension: a dimension in Integers within the range represent the actual number of convolutional layers used. ; Follow-up dimension Each dimension is an integer representing the number of convolutional kernels at the corresponding level; Further Dimension: Each dimension is an integer mapped from a discrete set, representing the size of the convolutional kernel at the corresponding level; (2) GRU encoded segment: length is ;in This is the maximum allowed number of GRU layers; First dimension: Integer, representing the number of GRU layers. ; Follow-up Dimension: An integer representing the number of hidden units in each layer of GRU; (3) Attention encoding segment: length 2; The first dimension: a discretized integer value used to index a predefined set of attention types; The second dimension: an integer representing the attention weight dimension. .

2. The land subsidence prediction method based on a hybrid model of feature selection and GWO optimization according to claim 1, characterized in that, The improved GWO also employs a probability-based perturbation update algorithm for position updates, specifically including: (1) Leadership probability calculation: based on , , Wolves in Dimensions The value on the table is used to calculate the individual. The attractiveness probability and random probability of learning from each leader wolf: ； ； ； ； in, and It is a coefficient vector. It is a function that maps distance and coefficients to probabilities; (2) Probability sampling update: based on the calculated probability distribution Random sampling is performed. If you are selected , or ,but Take the value of the corresponding leader wolf in this dimension; If you are selected If the parameter is selected, a new value will be randomly generated within the valid domain of the parameter.

3. The land subsidence prediction method based on a hybrid model of feature selection and GWO optimization according to claim 1, characterized in that, When updating the attention type parameter, if learning from the leader wolf is selected, its category is directly copied; if random exploration is selected, it is randomly selected uniformly from the entire category set.

4. A computer system, 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 land subsidence prediction method based on a hybrid model of feature selection and GWO optimization as described in any one of claims 1-3.

Images