A hierarchical extrapolation method for pile length of urban building complexes based on sparse partial least squares

By using the sparse partial least squares (SPLS) hierarchical extrapolation method, variables are automatically screened to identify factors influencing pile length, solving the problem of determining the depth of pile foundations in old urban areas. This enables rapid and accurate pile length inversion and risk assessment, reducing costs.

CN123116409APending Publication Date: 2026-11-17CHINA GEOLOGICAL SURVEY CHANGSHA NATURAL RESOURCES COMPREHENSIVE SURVEY CENT
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611357551.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-03
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In old urban areas and urban renewal areas, there is a lack of effective methods to quickly determine the pile foundation depth of existing buildings. Existing technologies such as traditional geotechnical investigation, empirical formula estimation, and BIM/3D geological modeling have problems such as data gaps, high costs, or complex calculations in this scenario, and cannot accurately obtain the pile length.

Method used

The Sparse Partial Least Squares (SPLS) method was adopted. By introducing an L1 penalty term to automatically screen variables, a sparse partial least squares regression model was established. Using data on building height, geological parameters, and pile foundation parameters, the pile length was extrapolated in layers, and core influencing factors were automatically identified and irrelevant variables were removed.

Benefits of technology

It enables rapid and accurate inversion of existing building pile lengths under limited sample conditions, reducing costs, improving model stability and prediction accuracy, and providing guidance for engineering surveys and risk assessments. It is applicable to old urban areas and urban renewal areas.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116409A_ABST
    Figure CN123116409A_ABST
Patent Text Reader

Abstract

This invention relates to the field of building exploration technology, specifically to a hierarchical extrapolation method for pile length of urban building clusters based on sparse partial least squares (PLS). Addressing the practical challenge of insufficient building data in old urban areas and urban renewal zones, this method provides a technical means to invert the pile length of existing buildings using a small sample size. This scheme automatically identifies core influencing factors during dimensionality reduction through the L1 penalty mechanism of the sparse PLS regression model. Unlike traditional machine learning methods, the sparse PLS regression model directly outputs variable coefficients, clearly demonstrating the direction and intensity of each factor's influence on pile length, which helps in understanding the controlling factors of pile length in the area and provides guidance for subsequent engineering surveys and risk assessments. Addressing the practical constraint of difficult sample acquisition in old urban areas, this application ensures model stability under small sample conditions through a data hierarchical mechanism, L1 sparsity penalty, and cross-validation parameter selection.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of building exploration technology, specifically to a method for stratified extrapolation of pile lengths in urban building clusters based on sparse partial least squares. Background Technology

[0002] Pile foundation depth (pile length) is a key parameter for risk assessment in urban underground space development, building foundation design, and shield tunnel construction. Accurately obtaining the pile foundation depth of existing buildings is of great significance for ensuring the safety of adjacent underground projects and controlling construction risks.

[0003] However, in old urban areas, urban renewal zones, and existing built-up areas, many buildings lack complete pile foundation as-built data. When underground space development is carried out in these areas (such as shield tunneling, deep foundation pit excavation, and underground utility tunnel construction), the pile foundation depth of surrounding existing buildings cannot be accurately determined, posing a serious threat to project safety.

[0004] Currently, there are several conventional methods for determining the depth of pile foundations for existing buildings: 1. Traditional geotechnical investigation and design methods: This method involves obtaining data such as the bearing stratum depth and soil parameters through building-by-building geological surveys, and then determining the pile length based on load calculations. This method is highly accurate and serves as the legally mandated basis for pile foundation design in new building projects; it is irreplaceable. However, for existing old urban areas and urban renewal zones, this method faces the following practical difficulties: First, there is a lack of data. Many old buildings lack complete as-built drawings and survey data for their pile foundations, making it impossible to directly obtain information on pile lengths.

[0005] Second, retrospective investigation is not possible. Since the building has already been completed, it is impossible to trace back and investigate each completed pile foundation. Traditional methods are not applicable in this scenario.

[0006] Third, the cost is unacceptable. Even if it were theoretically feasible, the cost and time required to conduct a building-by-building survey of an already developed area would be prohibitive, making it uneconomical.

[0007] 2. Empirical Formula Estimation Method: This method establishes an empirical regression equation between pile length and building height based on engineering experience. While simple to operate, it has the following limitations: it considers only the building height, neglecting key variables such as bearing stratum conditions and soil parameters; it cannot adapt to geological differences, as pile length patterns vary significantly across different geological units (soft soil areas, karst areas), and a single empirical formula cannot characterize these differences; and it lacks a variable screening mechanism, with too many irrelevant variables interfering with prediction accuracy and failing to identify truly critical influencing factors.

[0008] 3. BIM / 3D Geological Modeling Method: In recent years, BIM technology and 3D geological modeling have been applied in pile foundation engineering. This method has the following limitations: it is highly dependent on modeling accuracy, requiring detailed geological data and accurate 3D models, which are often lacking in old urban areas; it has high computational complexity, requiring professional operations such as finite element analysis and BIM secondary development.

[0009] In summary, there is currently a lack of a method to quickly invert the pile length of existing buildings using limited sample data, especially in scenarios where data is scarce, such as old urban areas and urban renewal areas, where there is no effective technical means to solve this problem.

[0010] It should be noted that the Partial Least Squares (PLS) method has been documented in the field of pile foundations. For example, the "Partial Least Squares Regression Prediction Method for Ultimate Bearing Capacity of Complete Pile" (2003) used the PLS method to establish a prediction model between parameters such as pile length and pile diameter and the ultimate bearing capacity of the pile foundation. However, this method belongs to the "forward calculation" of bearing capacity under known parameter conditions. Its input parameters already include the pile length, and the output target is the bearing capacity. This is fundamentally different from the pile length "inversion" under the condition of missing data in this invention in terms of technical problems, technical paths, and technical effects. First, this invention solves the inversion problem when the pile length of existing buildings is unknown, rather than calculating the bearing capacity when the pile length is known. Second, the L1 sparsity penalty mechanism introduced in this invention realizes automatic variable screening, which can automatically identify core influencing factors and eliminate irrelevant variables, a function that traditional PLS methods do not have. Third, this invention achieves the technical effect of extrapolating the overall region using a small number of samples. The above differences make this invention have outstanding substantive features and significant progress compared with the prior art. Summary of the Invention

[0011] The main objective of this invention is to provide a hierarchical extrapolation method for pile length of urban building clusters based on sparse partial least squares, aiming to solve the problem of the current lack of a method that can quickly invert the pile length of existing buildings using limited sample data.

[0012] Compared to the traditional Partial Least Squares (PLS) method, the Sparse Partial Least Squares (SPLS) method of this invention introduces an L1 penalty term, which automatically compresses the regression coefficients of irrelevant variables to zero during the modeling process, thereby achieving automatic variable selection. The traditional PLS method incorporates all input variables into the model and cannot automatically identify core influencing factors, leading to overfitting and poor model interpretability in the application scenario of this invention. Therefore, SPLS is not a simple replacement for PLS, but a targeted improvement for the specific technical problem of pile length inversion under conditions of missing data.

[0013] The technical solution proposed in this invention is as follows: A hierarchical extrapolation method for pile lengths in urban building complexes based on sparse partial least squares includes: Obtain sample building data corresponding to sample buildings within the target area, and classify the sample buildings into three height types according to building height: low-rise and multi-story, high-rise and super high-rise. The sample building data includes building feature data, geological parameter data and pile foundation parameter data. Convert the sample building data corresponding to the sample building into a numerical feature matrix X and a target vector Y; Based on the numerical feature matrix X and the target vector Y, a corresponding sparse partial least squares regression model is established for each height type. The sparse partial least squares regression model introduces an L1 sparse penalty term into the objective function of the partial least squares framework. The regression coefficients of each sparse partial least squares regression model are extracted. Regression coefficients with absolute values ​​greater than a preset threshold are identified as core influencing factors, and regression coefficients with absolute values ​​close to zero are identified as irrelevant variables and removed. Obtain the actual building data of the buildings to be evaluated within the target area, determine the height type of the buildings based on their actual building height, and select the corresponding sparse partial least squares model based on the height type as the target model. The actual building data and the sample building data are of the same type. Input the actual building data into the target model to obtain the predicted pile length of the building to be evaluated, as output by the target model.

[0014] Preferably, the building feature data includes building type and total building height, the geological parameter data includes geological unit, bearing layer type and bearing layer depth, and the pile foundation parameter data includes pile diameter, standard penetration test blow value, and actual pile length.

[0015] Preferably, the numerical feature matrix X consists of seven feature components: building type value, total building height, geological unit value, bearing stratum type value, bearing stratum depth, pile diameter, and standard penetration test blow value, and the target vector Y is the actual pile length.

[0016] Preferably, the number of sample buildings is not less than a preset number of groups, and the proportion of sample buildings collected according to building height is as follows: low-rise and multi-story buildings occupy the first preset interval, high-rise buildings occupy the second preset interval, and super high-rise buildings occupy the third preset interval.

[0017] Preferably, the building type values ​​are: low-rise and multi-story = 1, high-rise = 2, super high-rise = 3; the geological unit values ​​are: soft soil area = 1, clay area = 2, karst area = 3; and the bearing layer type values ​​are: cohesive soil = 1, sandy soil = 2, gravel = 3, rock = 4.

[0018] Preferably, the step of acquiring sample building data corresponding to sample buildings within the target area and classifying the sample buildings into three height types—low-rise / multi-story, high-rise, and super high-rise—according to their height includes: Buildings with a height less than or equal to a first preset height value are classified as low-rise and multi-story buildings. Buildings with a height greater than the first preset height value and less than or equal to the second preset height value are classified as high-rise buildings. Buildings with a height greater than the second preset height value are classified as super high-rise buildings.

[0019] Preferably, the step of establishing a corresponding sparse partial least squares regression model for each height type based on the numerical feature matrix X and the target vector Y includes: Minimizing the prediction error is used as the objective function of the sparse partial least squares regression model. An L1 penalty term is introduced into the objective function to control the sparsity of the feature matrix X and the objective vector Y. The L1 penalty term forms a diamond-shaped feasible region so that the optimal solution falls on the coordinate axis.

[0020] Preferably, the objective function of the sparse partial least squares regression model is to minimize the prediction error, and an L1 penalty term is introduced into the objective function to control the sparsity of the feature matrix X and the target vector Y. The model further includes: By selecting the optimal sparse parameters and number of latent variables for the sparse partial least squares regression model through cross-validation, sparse partial least squares regression models corresponding to different height types can be trained.

[0021] Preferably, the step of extracting regression coefficients from each sparse partial least squares regression model, identifying regression coefficients with absolute values ​​greater than a preset threshold as core influencing factors, and determining regression coefficients with absolute values ​​close to zero as irrelevant variables and removing them, includes: The sample building data for each height type are randomly divided into a training set and a test set according to a preset ratio; Calculate the root mean square error and mean absolute error of the sparse partial least squares regression model on the test set; When the root mean square error and mean absolute error of the test set are both controlled within the preset accuracy range, the sparse partial least squares regression model is determined to meet the requirements of the early stage of engineering evaluation.

[0022] Preferably, when both the root mean square error and the mean absolute error of the test set are controlled within a preset accuracy range, the sparse partial least squares regression model is determined to meet the requirements for pre-project evaluation. This further includes: Calculate the root mean square error and mean absolute error of the sparse partial least squares regression model on the training set; Calculate the ratio between the root mean square error of the training set and the root mean square error of the test set, and use it as the first ratio. Calculate the absolute value of the difference between the first ratio and the benchmark value, and use it as the first difference value. Calculate the ratio between the mean absolute error of the training set and the mean absolute error of the test set, and use it as the second ratio. Calculate the absolute value of the difference between the second ratio and the benchmark value, and use it as the second difference value. When both the first and second difference values ​​are less than the preset difference value, it is determined that the sparse partial least squares regression model has not undergone significant overfitting.

[0023] The above technical solution can achieve the following beneficial effects: This application proposes a hierarchical extrapolation method for pile length in urban building clusters based on sparse partial least squares (PLLS). Addressing the challenge of insufficient building data in existing urban areas and urban renewal zones, this method provides a technical means to invert the pile length of existing buildings using a small sample size. The scheme automatically identifies core influencing factors during dimensionality reduction through the L1 penalty mechanism of the sparse PLS regression model. Unlike traditional machine learning methods, the sparse PLS regression model directly outputs variable coefficients, clearly demonstrating the direction and intensity of each factor's influence on pile length. This helps in understanding the controlling factors of pile length in the area and provides guidance for subsequent engineering surveys and risk assessments. Addressing the practical constraint of difficult sample acquisition in old urban areas, this application ensures model stability under small sample conditions through a data hierarchical mechanism, L1 sparsity penalty, and cross-validation parameter selection. Under limited sample conditions, the hierarchical modeling method outperforms the mixed modeling method in predictive accuracy within each type.

[0024] It should be noted that the SPLS model in this invention is not a purely mathematical algorithm, but rather takes geotechnical engineering parameters with clear physical meaning (total building height, bearing stratum depth, standard penetration test blow count, etc.) as input and outputs predicted pile lengths with engineering application value. The variable selection and coefficient extraction steps in the model building process all serve the specific engineering technical objective of "existing building pile length inversion." The algorithmic features and technical features functionally support each other and interact with each other, forming a comprehensive technical solution. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0026] Figure 1This is a flowchart illustrating the first embodiment of a stratified extrapolation method for pile lengths of urban building complexes based on sparse partial least squares proposed in this invention. Detailed Implementation

[0027] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0028] This invention proposes a hierarchical extrapolation method for pile length of urban building clusters based on sparse partial least squares.

[0029] As attached Figure 1 As shown, in the first embodiment of the hierarchical extrapolation method for pile length of urban building clusters based on sparse partial least squares proposed in this invention, this embodiment includes the following steps: Step S110: Obtain sample building data corresponding to the sample buildings in the target area, and classify the sample buildings into three height types according to their height: low-rise and multi-story, high-rise and super high-rise. The sample building data includes building feature data, geological parameter data and pile foundation parameter data.

[0030] Specifically, 120 representative sample buildings were selected within the target area of ​​the city, including 30 low-rise and multi-story buildings (25%), 66 high-rise buildings (55%), and 24 super high-rise buildings (20%).

[0031] Step S120: Convert the sample building data corresponding to the sample building into a numerical feature matrix X and a target vector Y.

[0032] Specifically, the constructed numerical feature matrix X has a dimension of 120×7, and the constructed target vector Y has a dimension of 120×1.

[0033] Step S130: Based on the numerical feature matrix X and the target vector Y, establish a corresponding sparse partial least squares regression model (SPLS model) for each height type. The sparse partial least squares regression model introduces an L1 sparse penalty term into the objective function of the partial least squares framework.

[0034] Specifically, the 120 samples were divided into three categories based on building height: low-rise / multi-story (height ≤ 27m, 30 groups), high-rise (27m < height ≤ 100m, 66 groups), and super high-rise (height > 100m, 24 groups). For each height type, the training and test sets were randomly divided in an 8:2 ratio.

[0035] For the sample building data of each height type, 10-fold cross-validation was used to search for the optimal parameter combination within the range of 0.3-0.9 for the sparsity parameter and 1-5 for the number of latent variables. Taking high-rise buildings as an example, the optimal sparsity parameter was 0.6 and the number of latent variables was 3, and an SPLS regression model was established.

[0036] Step S140: Extract the regression coefficients of each sparse partial least squares regression model, identify the regression coefficients with absolute values ​​greater than a preset threshold (e.g., 0.0001) as core influencing factors, and remove the regression coefficients with absolute values ​​close to zero as irrelevant variables. Then, verify the accuracy of the sparse partial least squares regression model on the test set.

[0037] Specifically, the regression coefficients of each sparse partial least squares regression model were extracted. In the application validation in the target area, the SPLS model automatically identified the total building height, bearing stratum depth, and standard penetration test (SPT) blow values ​​as the core control variables (with relatively large absolute values), while the coefficients of geological units approached 0. This result indicates that, under the geological conditions of this area, the influence of macro-geological zoning on pile length has been explained by the bearing stratum depth and SPT blow values.

[0038] Regarding the method for determining the burial depth of the bearing layer: When obtaining the bearing stratum depth from adjacent borehole data, the following rules apply: (a) If the building has no basement: the depth of the bearing layer is directly taken from the “layer replacement depth” value corresponding to the top surface of the bearing layer in the borehole record.

[0039] (b) If the building has a basement: the depth of the bearing stratum shall be corrected according to the following formula: Effective bearing layer depth = max(0, layer replacement depth - basement floor depth) The depth of the basement floor slab can be determined by referring to the architectural drawings or by estimating the depth based on the number of underground floors (approximately 3.5-5 meters for the first underground floor and approximately 7-10 meters for the second underground floor).

[0040] (c) For cases of strata tilt: the "contour map of the top surface of the bearing stratum" in the exploration report should be used first. The coordinates of the building should be placed on the map, the elevation of the top surface of the bearing stratum at that point should be read, and then the following formula should be used for calculation: Bearing stratum depth = Ground elevation - Bearing stratum top surface elevation (d) For samples with moderately weathered rock as the bearing layer: the standard penetration test blow value (SPT-N) usually exceeds 100 blows, which loses its distinguishability. In this case, the value should be marked as ">100" or a special identifier (such as 999) should be used, and the model should be distinguished according to the bearing layer type factor during model training.

[0041] Step S150: Obtain the actual building data of the building to be evaluated within the target area, determine the height type of the building based on its actual building height, and select the corresponding sparse partial least squares model based on the height type as the target model. The actual building data and the sample building data are of the same type.

[0042] Step S160: Input the actual building data into the target model to obtain the predicted pile length and prediction range (e.g., 95% prediction range) of the building to be evaluated output by the target model.

[0043] For example, the building to be evaluated (actual building data: height 68 meters, located in a clay area, bearing layer is sand, bearing layer depth 32 meters, pile diameter 0.8 meters, standard penetration test blow value is 25) after collecting the above seven characteristic data, the building is automatically determined to belong to the high-rise category based on its building height (68 meters). The SPLS model corresponding to the high-rise category is called, the actual building data is input, and the predicted pile length is 28.5 meters, with a 95% prediction range of [25.8m, 31.2m]. This result can be used as the basis for risk assessment when the shield tunnel passes under the building.

[0044] This application proposes a hierarchical extrapolation method for pile length in urban building clusters based on sparse partial least squares (PLLS). Addressing the challenge of insufficient building data in existing urban areas and urban renewal zones, this method provides a technical means to invert the pile length of existing buildings using a small sample size. The scheme automatically identifies core influencing factors during dimensionality reduction through the L1 penalty mechanism of the sparse PLS regression model. Unlike traditional machine learning methods, the sparse PLS regression model directly outputs variable coefficients, clearly demonstrating the direction and intensity of each factor's influence on pile length. This helps in understanding the controlling factors of pile length in the area and provides guidance for subsequent engineering surveys and risk assessments. Addressing the practical constraint of difficult sample acquisition in old urban areas, this application ensures model stability under small sample conditions through a data hierarchical mechanism, L1 sparsity penalty, and cross-validation parameter selection. Under limited sample conditions, the hierarchical modeling method outperforms the mixed modeling method in predictive accuracy within each type.

[0045] Furthermore, this application establishes separate models for low-rise / multi-story, high-rise, and super high-rise buildings, addressing the issue of differences in pile length patterns across different building types. This stratified strategy originates from general engineering principles of pile foundation design and has nationwide applicability. From sample building data collection to regional inversion, this scheme requires only a few steps. It eliminates the need for complex numerical simulations, finite element analysis, or extensive geological exploration, significantly reducing the technical threshold and cost of evaluating existing building pile lengths. This method can serve as an auxiliary tool for rapid assessment in the early stages of urban underground space development.

[0046] In the second embodiment of the stratified extrapolation method for pile length of urban building clusters based on sparse partial least squares proposed in this invention, based on the first embodiment, the building characteristic data includes building type and total building height, the geological parameter data includes geological unit, bearing layer type and bearing layer burial depth, and the pile foundation parameter data includes pile diameter, standard penetration test blow value, and actual pile length.

[0047] Specifically, building characteristic data includes: building type (low-rise / multi-story / high-rise / super high-rise) and total building height, which can be obtained through on-site surveys or design drawings; Geological parameter data includes: geological unit (soft soil area / clay area / karst area), bearing layer type, bearing layer burial depth, which can be obtained through adjacent borehole data or regional geological maps; The pile foundation parameter data includes: pile diameter, standard penetration test blow count (SPT-N value), and actual pile length, which can be obtained through local data collection or on-site survey.

[0048] In the third embodiment of the method for stratified extrapolation of pile length in urban building complexes based on sparse partial least squares proposed in this invention, based on the first embodiment, the numerical feature matrix X consists of seven feature components: building type value, total building height, geological unit value, bearing layer type value, bearing layer burial depth, pile diameter, and standard penetration test blow value, and the target vector Y is the actual pile length.

[0049] In the fourth embodiment of the urban building cluster pile length stratification method based on sparse partial least squares proposed in this invention, based on the second embodiment, the number of sample buildings is not less than a preset number of groups (100 groups), and the proportion of sample buildings collected according to building height is as follows: low-rise and multi-story buildings account for the first preset interval (20%-30%), high-rise buildings account for the second preset interval (50%-60%), and super high-rise buildings account for the third preset interval (15%-20%).

[0050] It should be noted that the number of sample buildings is far less than the total number of buildings in the target area. The overall building situation in the target area can be inferred from a small number of samples.

[0051] In the fifth embodiment of the method for stratified derivation of pile length of urban building groups based on sparse partial least squares proposed in this invention, based on the third embodiment, the building type values ​​are: low-rise multi-story = 1, high-rise = 2, super high-rise = 3; the geological unit values ​​are: soft soil area = 1, clay area = 2, karst area = 3; and the bearing layer type values ​​are: cohesive soil = 1, sandy soil = 2, gravel = 3, rock = 4.

[0052] In the sixth embodiment of the method for hierarchical extrapolation of pile length in urban building clusters based on sparse partial least squares proposed in this invention, based on the first embodiment, step S110 includes the following steps: Step S610: Classify buildings with a height less than or equal to the first preset height value (27 meters) into low-rise and multi-story buildings.

[0053] Specifically, low-rise and multi-story buildings correspond to Class C pile foundation design. Since the upper load is relatively small, the natural foundation can usually meet the bearing capacity and deformation requirements. Therefore, the foundation form is mainly shallow foundation or short pile.

[0054] Step S620: Classify buildings whose height is greater than the first preset height value (27 meters) and less than or equal to the second preset height value (100 meters) as high-rise buildings.

[0055] Specifically, high-rise buildings correspond to Class B pile foundation design. As building height increases, the load increases significantly. To effectively control settlement and meet bearing capacity requirements, conventional pile foundations are the primary form of foundation for such buildings.

[0056] Step S630: Classify buildings with a height greater than the second preset height value (100 meters) as super high-rise buildings, corresponding to Class A pile foundation design.

[0057] Specifically, super high-rise buildings correspond to Class A pile foundation design. These buildings bear enormous loads and have extremely high requirements for settlement control. According to the design grade classification of the "Code for Design of Building Pile Foundations" JGJ 94-2008, their foundations typically adopt ultra-long piles or rock-socketed piles with higher bearing capacity.

[0058] Specifically, SPLS models are established in layers according to building type. This layering strategy is based on the design grade classification principle of the "Code for Design of Building Pile Foundations" JGJ 94-2008: different building types have different load magnitudes and settlement control requirements, and the corresponding pile foundation design grades and foundation forms have systematic differences.

[0059] In the seventh embodiment of the method for hierarchical extrapolation of pile length in urban building clusters based on sparse partial least squares proposed in this invention, based on the first embodiment, step S130 includes the following steps: Step S710: Minimize the prediction error as the objective function of the sparse partial least squares regression model, and introduce an L1 penalty term into the objective function to control the sparsity of the numerical feature matrix X and the objective vector Y. The L1 penalty term forms a diamond-shaped feasible region so that the optimal solution falls on the coordinate axis.

[0060] Specifically, let the dimension of the standardized numerical feature matrix be n×7 (n is the number of samples, 7 feature components), and the dimension of the target vector Y be n×1 (n is the number of samples, 1 feature component); the sparse partial least squares regression model, based on traditional partial least squares, introduces an L1 penalty term to construct the following objective function: , In the formula, , It is an information matrix that measures the covariance between features and the target. It is the direction vector (i.e., load vector) of the sparse partial least squares regression model. It is the sparse direction vector after applying sparse constraints; It is a trade-off parameter used to balance maximizing covariance with regression fit, and it satisfies: ; It is the L1 sparsity penalty term parameter, used to control... The sparsity of the features makes the unimportant feature coefficients zero; It is the L2 penalty term parameter, used to handle feature collinearity and enhance stability; , It is the norm of the L1 penalty term. It is the j-th original input feature (i.e., the j-th feature component in the numerical feature matrix); It is the square of the norm of the L2 penalty term.

[0061] Specifically, the above steps compress unimportant regression coefficients to zero, achieving variable selection simultaneously during dimensionality reduction. With the goal of minimizing prediction error, an L1 penalty term is introduced into the objective function, and optimal parameters are selected through cross-validation (λ1 controls the sparsity of the feature matrix X, and λ2 controls the sparsity of the target vector Y).

[0062] The L1 penalty term is a key mechanism for automatic variable selection. Its mathematical principle lies in the fact that the L1 penalty term forms a diamond-shaped feasible region, making it easier for the optimal solution to fall on the coordinate axes during the optimization process, thus generating sparse solutions, i.e., some coefficients are zero. This mechanism compresses the coefficients of unimportant variables to zero, achieving variable selection simultaneously during dimensionality reduction.

[0063] In the eighth embodiment of the method for hierarchical extrapolation of pile lengths of urban building clusters based on sparse partial least squares proposed in this invention, based on the seventh embodiment, step S710 is followed by the following steps: Step S810: Select the optimal sparse parameters and number of latent variables for the sparse partial least squares regression model through cross-validation, thereby training sparse partial least squares regression models corresponding to different height types.

[0064] Specifically, the cross-validation adopts 10-fold cross-validation, the search range of sparse parameters is 0.3 to 0.9, the search range of the number of latent variables is 1 to 5, and the optimal parameter combination is selected with the goal of minimizing the prediction error.

[0065] In the ninth embodiment of the method for hierarchical extrapolation of pile length in urban building clusters based on sparse partial least squares proposed in this invention, based on the first embodiment, step S140 includes the following steps: Step S910: Randomly divide the sample building data of each height type into a training set and a test set according to a preset ratio (8:2).

[0066] Specifically, 80% of the training set is used for model training, and 20% of the test set is used to independently verify the model's accuracy. A fixed random seed is used to ensure that the results are reproducible.

[0067] Step S920: Calculate the root mean square error and mean absolute error of the sparse partial least squares regression model on the test set.

[0068] Step S930: When the root mean square error and mean absolute error of the test set are both controlled within the preset accuracy range (meter-level range, for example, 1 meter) and are significantly better than the estimation error of the traditional empirical formula, the sparse partial least squares regression model is determined to meet the requirements of the early stage of engineering evaluation.

[0069] Specifically, in terms of model accuracy verification, the root mean square error (RMSE) and mean absolute error (MAE) are calculated on independent test sets. The RMSE reflects the overall deviation level of the predicted values, while the MAE reflects the average error magnitude of the predicted values.

[0070] When the root mean square error and mean absolute error of the test set are both controlled within a reasonable range (usually at the meter level) and are significantly better than the estimation error of traditional empirical formulas, the model is considered to meet the requirements for rapid evaluation in the early stage of engineering.

[0071] In the tenth embodiment of the method for hierarchical extrapolation of pile length in urban building clusters based on sparse partial least squares proposed in this invention, based on the ninth embodiment, after step S930, the following steps are further included: Step S1010: Calculate the root mean square error and mean absolute error of the sparse partial least squares regression model on the training set.

[0072] Step S1020: Calculate the ratio between the root mean square error of the training set and the root mean square error of the test set, and use it as the first ratio. Calculate the absolute value of the difference between the first ratio and the benchmark value (e.g., 1), and use it as the first difference value.

[0073] Step S1030: Calculate the ratio between the mean absolute error of the training set and the mean absolute error of the test set, and use it as the second ratio. Calculate the absolute value of the difference between the second ratio and the benchmark value (e.g., 1), and use it as the second difference value.

[0074] Step S1040: When both the first difference value and the second difference value are less than the preset difference value (e.g., 0.1), it is determined that the sparse partial least squares regression model has not undergone significant overfitting.

[0075] Specifically, by comparing the accuracy of the training set with the accuracy of the test set, if the difference between the two is less than 10%, the model is considered not to have obviously overfitted.

[0076] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0077] 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. All of these forms are within the protection scope of the present invention.

Claims

1. A sparse partial least squares-based hierarchical deduction method for pile length of urban building groups, characterized in that, include: Obtain sample building data corresponding to sample buildings within the target area, and classify the sample buildings into three height types according to building height: low-rise and multi-story, high-rise and super high-rise. The sample building data includes building feature data, geological parameter data and pile foundation parameter data. Convert the sample building data corresponding to the sample building into a numerical feature matrix X and a target vector Y; Based on the numerical feature matrix X and the target vector Y, a corresponding sparse partial least squares regression model is established for each height type. The sparse partial least squares regression model introduces an L1 sparse penalty term into the objective function of the partial least squares framework. The regression coefficients of each sparse partial least squares regression model are extracted. Regression coefficients with absolute values ​​greater than a preset threshold are identified as core influencing factors, and regression coefficients with absolute values ​​close to zero are identified as irrelevant variables and removed. Obtain the actual building data of the buildings to be evaluated within the target area, determine the height type of the buildings based on their actual building height, and select the corresponding sparse partial least squares model based on the height type as the target model. The actual building data and the sample building data are of the same type. Input the actual building data into the target model to obtain the predicted pile length of the building to be evaluated, as output by the target model.

2. The method according to claim 1, wherein, The building feature data includes building type and total building height; the geological parameter data includes geological unit, bearing layer type and bearing layer depth; and the pile foundation parameter data includes pile diameter, standard penetration test blow value, and actual pile length.

3. The method according to claim 1, wherein, The numerical feature matrix X consists of seven feature components: building type value, total building height, geological unit value, bearing stratum type value, bearing stratum depth, pile diameter, and standard penetration test blow value. The target vector Y is the actual pile length.

4. The method according to claim 2, wherein, The number of sample buildings is no less than a preset number of groups, and the proportion of sample buildings collected according to building height is as follows: low-rise and multi-story buildings occupy the first preset interval, high-rise buildings occupy the second preset interval, and super high-rise buildings occupy the third preset interval.

5. The method according to claim 3, wherein, The building type values ​​are: low-rise and multi-story = 1, high-rise = 2, super high-rise = 3; the geological unit values ​​are: soft soil area = 1, clay area = 2, karst area = 3; the bearing layer type values ​​are: cohesive soil = 1, sandy soil = 2, gravel = 3, rock = 4.

6. The method according to claim 1, wherein, The process involves acquiring sample building data corresponding to sample buildings within the target area, and classifying the sample buildings into three height types based on their height: low-rise / multi-story, high-rise, and super high-rise. Buildings with a height less than or equal to a first preset height value are classified as low-rise and multi-story buildings. Buildings with a height greater than the first preset height value and less than or equal to the second preset height value are classified as high-rise buildings. Buildings with a height greater than the second preset height value are classified as super high-rise buildings.

7. The method according to claim 1, wherein, The step of establishing a corresponding sparse partial least squares regression model for each height type based on the numerical feature matrix X and the target vector Y includes: Minimizing the prediction error is used as the objective function of the sparse partial least squares regression model. An L1 penalty term is introduced into the objective function to control the sparsity of the numerical feature matrix X and the objective vector Y. The L1 penalty term forms a diamond-shaped feasible region so that the optimal solution falls on the coordinate axis.

8. The method according to claim 7, wherein, The objective function of the sparse partial least squares regression model is to minimize the prediction error, and an L1 penalty term is introduced into the objective function to control the sparsity of the feature matrix X and the target vector Y. The model also includes: By selecting the optimal sparse parameters and number of latent variables for the sparse partial least squares regression model through cross-validation, sparse partial least squares regression models corresponding to different height types can be trained.

9. The method according to claim 1, wherein, The process of extracting regression coefficients from each sparse partial least squares regression model, identifying regression coefficients with absolute values ​​greater than a preset threshold as core influencing factors, and removing regression coefficients with absolute values ​​approaching zero as irrelevant variables includes: The sample building data for each height type are randomly divided into a training set and a test set according to a preset ratio; Calculate the root mean square error and mean absolute error of the sparse partial least squares regression model on the test set; When the root mean square error and mean absolute error of the test set are both controlled within the preset accuracy range, the sparse partial least squares regression model is determined to meet the requirements of the early stage of engineering evaluation.

10. The method according to claim 9, wherein, When both the root mean square error and the mean absolute error of the test set are controlled within a preset accuracy range, the sparse partial least squares regression model is determined to meet the requirements for pre-project evaluation. This is followed by: Calculate the root mean square error and mean absolute error of the sparse partial least squares regression model on the training set; Calculate the ratio between the root mean square error of the training set and the root mean square error of the test set, and use it as the first ratio. Calculate the absolute value of the difference between the first ratio and the benchmark value, and use it as the first difference value. Calculate the ratio between the mean absolute error of the training set and the mean absolute error of the test set, and use it as the second ratio. Calculate the absolute value of the difference between the second ratio and the benchmark value, and use it as the second difference value. When both the first and second difference values ​​are less than the preset difference value, it is determined that the sparse partial least squares regression model has not undergone significant overfitting.