Intelligent prediction method and device for foundation pit deformation

Abstract

The application provides a method and device for intelligently predicting foundation pit deformation. The method provided by the application comprises the following steps: constructing a historical data set; obtaining the excavation depth and original geological soil layer parameters of a foundation pit to be predicted; dividing the original geological soil layer structure of the foundation pit to be predicted within the influence depth range into multiple equivalent layers according to the excavation depth; determining the spatial overlap mapping relationship between each equivalent layer and each original geological soil layer in the vertical direction, spatially weighting and equally distributing the corresponding original geological soil layer parameters, obtaining the standard layer thickness and standardized physical and mechanical parameters of each equivalent layer, and combining to construct a standardized feature vector; training multiple machine learning models in parallel based on the historical data set, matching the corresponding best prediction model for each foundation pit deformation index according to a test performance index; and inputting the standardized feature vector of the foundation pit to be predicted into each best prediction model corresponding to each foundation pit deformation index, and outputting the prediction values of multiple foundation pit deformation indexes.

Description

Technical Field

[0001] This application relates to the field of construction monitoring and safety assessment technology for foundation pit engineering, and in particular to an intelligent prediction method and device for foundation pit deformation. Background Technology

[0002] During the actual excavation of the foundation pit, the retaining structure and surrounding soil will undergo varying degrees of deformation due to soil stress release and groundwater disturbance. Among these, the deep horizontal displacement of the retaining structure, the surrounding surface settlement, and the horizontal displacement of the pile tops are the most direct and critical indicators for measuring and assessing the overall structural safety of the foundation pit. Therefore, scientifically and accurately predicting these foundation pit deformation indicators in advance is of great practical significance for identifying instability risks and ensuring project safety.

[0003] Currently, the industry mainly relies on three traditional methods to predict the aforementioned foundation pit deformation indicators. The first is the empirical formula method, which primarily relies on historical engineering experience and statistical data from specific regions for trend estimation. The second is the numerical simulation method (such as finite element analysis), which simulates the actual excavation stress process by establishing complex mechanical models and inputting soil parameters. The third is the data-driven method, which has emerged in recent years, using a single machine learning model (such as a single neural network or a single support vector machine) to establish a nonlinear mapping for deformation data prediction. However, these existing prediction methods are insufficient to meet the requirements of modern high-precision safety monitoring. First, the empirical formula method relies excessively on specific regional experience, resulting in poor generalization ability and difficulty in adapting to different and changing geological conditions. Second, the numerical simulation method suffers from high computational load and long analysis cycles, and is highly dependent on soil constitutive parameters that are difficult to obtain accurately, leading to low efficiency in engineering applications. Meanwhile, single machine learning models face problems such as unstable prediction accuracy, overfitting, and significant differences in adaptability to different types of deformation indicators. More seriously, existing methods fail to fully utilize the stratified characteristics of soil parameters, often employing simplified processing methods that result in a loss of accuracy in geological information. At the same time, the lack of a unified prediction framework for multiple deformation indices means that different indices need to be modeled independently, resulting in a huge overall workload.

[0004] Therefore, there is an urgent need for a method that can make full use of soil layer parameters to accurately predict various foundation pit deformation indices. Summary of the Invention

[0005] In view of this, this application provides an intelligent prediction method and device for foundation pit deformation, which makes full use of soil layer parameters to achieve accurate prediction of multiple foundation pit deformation indicators.

[0006] Specifically, this application is implemented through the following technical solution:

[0007] The first aspect of this application provides an intelligent prediction method for foundation pit deformation, the method comprising:

[0008] Construct a historical dataset; the historical dataset includes the geometric parameters of historical foundation pit projects, the original geological soil layer parameters, and the foundation pit deformation index, the geometric parameters including the excavation depth;

[0009] Obtain the excavation depth and original geological soil parameters of the foundation pit to be predicted;

[0010] Based on the excavation depth, the original geological soil structure of the foundation pit to be predicted within the influence depth range is divided into a preset number of equivalent layers;

[0011] Determine the spatial overlap mapping relationship between each equivalent layer and each original geological soil layer in the vertical direction. Based on the spatial overlap mapping relationship, perform spatial weighted average distribution on the corresponding original geological soil layer parameters to obtain the standard layer thickness and standardized physical and mechanical parameters of each equivalent layer. Combine these parameters to construct a standardized feature vector.

[0012] Based on the historical dataset, multiple machine learning models are trained in parallel, and the best prediction model is adaptively matched for each foundation pit deformation index according to the test performance index.

[0013] The standardized feature vectors of the foundation pit to be predicted are input into the optimal prediction model corresponding to each foundation pit deformation index, and the predicted values ​​of multiple foundation pit deformation indices are output simultaneously.

[0014] A second aspect of this application provides an intelligent prediction device for foundation pit deformation, the device comprising a construction module, an acquisition module, a division module, a matching module, and an output module;

[0015] The construction module is used to construct a historical dataset; the historical dataset includes the geometric parameters of historical foundation pit projects, the original geological soil layer parameters, and the foundation pit deformation index, and the geometric parameters include the excavation depth;

[0016] The acquisition module is used to acquire the excavation depth and original geological soil layer parameters of the foundation pit to be predicted.

[0017] The division module is used to divide the original geological soil layer structure of the foundation pit to be predicted within the influence depth range into a preset number of equivalent layers according to the excavation depth.

[0018] The construction module is also used to determine the spatial overlap mapping relationship between each equivalent layer and each original geological soil layer in the vertical direction. Based on the spatial overlap mapping relationship, the corresponding original geological soil layer parameters are spatially weighted and averaged to obtain the standard layer thickness and standardized physical and mechanical parameters of each equivalent layer, and then a standardized feature vector is constructed by combining them.

[0019] The matching module is used to train multiple machine learning models in parallel based on the historical dataset, and adaptively match the best prediction model for each foundation pit deformation index according to the test performance index.

[0020] The output module is used to input the standardized feature vector of the foundation pit to be predicted into the best prediction model corresponding to each foundation pit deformation index, and simultaneously output the predicted values ​​of multiple foundation pit deformation indices.

[0021] The intelligent prediction method and device for foundation pit deformation provided in this application construct a historical dataset containing geometric parameters, original geological soil layer parameters, and foundation pit deformation indices as the underlying data driver. This provides a highly deterministic training base for subsequent parallel training and nonlinear law mining, effectively overcoming the shortcomings of traditional mechanical empirical models in generalizing and fitting massive heterogeneous soil and rock data. Furthermore, by directly acquiring two core engineering parameters—the excavation depth of the foundation pit to be predicted and the original geological soil layer parameters—precise anchoring of spatial geometric boundaries and complex and variable natural stratification features is achieved, ensuring strong coupling between digital representation and the engineering site from the input source. Based on this, the original geological soil layer structure is rationally divided into a preset number of equivalent layers within the influence depth according to the dynamically changing excavation depth. Combined with the spatial overlap mapping relationship in the vertical direction, the natural strata parameters are spatially weighted and averaged to extract standard layer thickness and standardized physical and mechanical parameters. This innovative topological transformation and geometric averaging processing... This approach fundamentally solves the underlying technical bottleneck of misaligned machine learning input matrix dimensions caused by the non-fixed number of soil layers and extremely uneven natural layer thickness in different engineering geological surveys. Without losing valuable mechanical information such as key strata shear strength and compression characteristics, it forcibly constrains and combines standardized feature vectors, achieving efficient conversion of unstructured strata data into standard digital interfaces for models. Finally, using this normalized vector, multiple machine learning models are trained in parallel based on historical datasets. Adaptive decoupling is performed for different deformation indices according to test performance metrics, thereby accurately locking in the optimal prediction model for each. This deeply aligns with the objective reality of drastically different underlying mechanical response mechanisms for multi-dimensional deformations such as lateral displacement and surface settlement of foundation pit retaining structures. It breaks through the limitations of traditional single-algorithm prediction leading to local index accuracy distortion, significantly improving the accuracy and training convergence stability of intelligent prediction of foundation pit deformation under complex geotechnical conditions while simultaneously outputting predicted values ​​for multiple indices. Attached Figure Description

[0022] Figure 1 A flowchart of the intelligent prediction method for foundation pit deformation provided in Embodiment 1 of this application;

[0023] Figure 2 This is a schematic diagram of the intelligent prediction device for foundation pit deformation provided in Embodiment 2 of this application. Detailed Implementation

[0024] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0025] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0026] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0027] The following specific embodiments are given to illustrate the technical solution of this application in detail.

[0028] Figure 1 This is a flowchart of the intelligent prediction method for foundation pit deformation provided in Embodiment 1 of this application. Please refer to... Figure 1 The method provided in this embodiment may include:

[0029] S101. Construct historical datasets.

[0030] The historical dataset contains geometric parameters of historical foundation pit projects, original geological soil layer parameters, and foundation pit deformation indices. The geometric parameters include the excavation depth.

[0031] Optionally, the original geological soil layer parameters include thickness, cohesion, internal friction angle and compression modulus, and the foundation pit deformation indicators include horizontal displacement of the retaining structure, surface settlement and horizontal displacement of the pile top.

[0032] Specifically, historical foundation pit projects refer to a collection of various underground space development or foundation pit excavation projects that have been completed, are under construction, or have accumulated phased monitoring data before the excavation of the foundation pit to be predicted. These historical foundation pit projects are usually located in the same or similar geological areas as the foundation pit to be predicted, or have comparable geotechnical engineering characteristics.

[0033] Geometric parameters refer to physical geometric characteristic parameters used to characterize the macroscopic three-dimensional morphology, boundary conditions, and support design dimensions of foundation pit engineering. In this application, geometric parameters include excavation depth (ExcDepth), which is the core controlling factor inducing stress release and deformation of the foundation pit soil.

[0034] Original geological soil layer parameters refer to the in-situ physical and mechanical properties of each natural original soil layer directly derived from the conventional geological survey report (geological investigation report) during the foundation pit design or construction stage, without undergoing artificial equivalent stratification. Specific original geological soil layer parameters include, but are not limited to: the actual thickness (Thick), cohesion (c), internal friction angle (φ), and compression modulus (Es) of each natural soil layer. Original geological soil layer parameters represent the initial, unsimplified geological information of nature, and the number and thickness of each layer vary irregularly depending on the engineering site, exhibiting an arbitrary multi-layered state.

[0035] Excavation pit deformation indices refer to various physical deformation or displacement observation variables used to dynamically measure and quantitatively assess the stability of the excavation pit's own support system and the safety status of the surrounding ecological environment during the excavation and subsequent underground structure construction lifecycle. Specific excavation pit deformation indices include, but are not limited to: horizontal displacement of the retaining structure (such as the deep horizontal displacement CX of the wall), which characterizes the lateral stress deformation of the retaining structure; surface settlement (CJ), which characterizes the degree of disturbance to the surrounding environment; and horizontal displacement of the pile top (SP), which characterizes the top constraint state of the support system.

[0036] In practice, geometric design data for each excavation stage is extracted from design drawings, construction plans, and work condition records of previously completed or ongoing historical foundation pit projects. The focus is on reading and recording the excavation depth values ​​corresponding to each work condition sample. Natural strata data at each borehole location is retrieved from the detailed geotechnical investigation reports corresponding to the aforementioned historical foundation pit projects. The thickness of each original soil layer is extracted layer by layer, along with its corresponding physical and mechanical properties such as cohesion, internal friction angle, and compression modulus. Measured deformation data corresponding to the excavation depth and geological conditions in time and space are exported from the automated on-site monitoring system or historical monitoring reports. Monitoring values ​​for the horizontal displacement of the retaining structure, surface settlement, and pile top horizontal displacement are recorded. Using specific historical foundation pit project numbers, monitoring point locations, and specific excavation construction steps as anchor tags, the collected geometric parameters, original geological soil layer parameters, and corresponding foundation pit deformation monitoring values ​​are matched spatiotemporally and aligned with the work conditions. The aligned single-set working condition data is transformed into structured data samples. The excavation depth parameter and the combination of the thickness, cohesion, internal friction angle and compression modulus parameters of all original soil layers are used as the input feature items of the sample. The corresponding horizontal displacement of the retaining structure, surface settlement and pile top horizontal displacement monitoring values ​​are used as the output label items of the sample. All structured data samples collected and aligned from historical projects are stacked vertically and assembled to form the initial historical dataset matrix.

[0037] S102. Obtain the excavation depth and original geological soil layer parameters of the foundation pit to be predicted.

[0038] Specifically, the term "predictable foundation pit" refers to a specific target underground space development or foundation pit support project that is currently undergoing scheme design, optimization demonstration, or is in the excavation and construction preparation stage, and requires an advanced quantitative assessment of its future deformation risks. Unlike historical foundation pit projects that have accumulated a large amount of deformation monitoring results, the key safety deformation indicators such as the horizontal displacement of the retaining structure and surface settlement of the predictable foundation pit are unknown under the current target excavation conditions, which is the direct target of this application.

[0039] Excavation depth refers to the vertical distance measured vertically downwards from the natural ground surface (or the relative reference zero-point plane set for the foundation pit project) to the current designed excavation bottom surface (or the bottom cross-section of the pit) under specific construction stages or design conditions. Excavation depth is the fundamental driving force causing the disruption of the initial stress balance of the foundation pit soil and the lateral and vertical deformation of the retaining structure. Excavation depth can refer to the final designed total excavation depth of the foundation pit, or it can refer to the staged excavation depth corresponding to each working condition step in a multi-step, layered excavation construction process. In this application, excavation depth is used as the first input variable of the standardized feature vector to quantitatively characterize the physical intensity of stress release in the foundation pit.

[0040] In practice, the design and construction drawings, support design specifications, or current construction plan of the foundation pit project to be predicted are consulted to obtain the excavation depth value corresponding to the target predicted working condition. A detailed geotechnical investigation report matching the spatial location of the foundation pit to be predicted is retrieved, and a typical stratigraphic column or stratigraphic profile is located within the target foundation pit area. Starting from the natural ground surface and progressing from shallow to deep, the physical and mechanical properties of each natural original soil layer within the project's influence range are extracted layer by layer. The actual thickness of each original soil layer, as well as the corresponding cohesion parameters, internal friction angle parameters, and compression modulus parameters, are read and recorded to obtain the original geological soil layer parameters.

[0041] S103. Based on the excavation depth, the original geological soil structure of the foundation pit to be predicted within the influence depth range is divided into a preset number of equivalent layers.

[0042] Specifically, the depth of influence refers to the entire vertical spatial depth range affected by the initial stress release of the soil caused by the excavation of the foundation pit, which in turn triggers stress redistribution, soil shear deformation, and groundwater disturbance. In this application, the depth of influence is defined as the entire vertical area extending vertically downwards from the natural ground surface of the foundation pit to be predicted, to the bottom interface of the total thickness of the disturbed soil. It not only covers the depth space where earthwork excavation is directly carried out, but also fully includes the deep, unexcavated soil area below the pit bottom that is significantly affected by the stress release due to the heave of the pit bottom and the deep lateral bending transmission of the retaining structure.

[0043] The original geological soil layer structure refers to the naturally generated, complex, multi-level geological soil strata distribution within the site of the excavation pit to be predicted, without artificial simplification, numerical smoothing, or geometric equivalence. The equivalent layer refers to a virtual, standardized computational space layer obtained by scientifically dividing the original geological soil layers according to geometric and mechanical rules. This overcomes the inherent technical limitation of machine learning models being unable to accept fixed-dimensional tensor inputs due to the variable number and thickness of the original geological soil layers. The equivalent layer is a geometrically defined layer with a fixed number of layers, using the current excavation depth as a dynamic geometric benchmark. It divides the aforementioned influencing depth range vertically according to a preset proportional boundary. The equivalent layer does not change the physical nature of the original soil strata; its core function is to act as a digital mapping bridge, transforming arbitrarily complex original geological soil layer structures into standardized feature vectors, thereby adapting to the parallel training and adaptive prediction of machine learning models.

[0044] It should be noted that if a fixed number of layers is used to divide the original geological soil structure, it relies too much on prior human experience and has rigid boundaries, making it impossible to make adaptive topological adjustments based on the scale of the foundation pit project itself (such as the total depth). At the same time, the soil disturbance and stress release in the vertical depth direction of the foundation pit are highly nonlinear, and fixed layers can easily smooth or mask the local mechanical abrupt changes, resulting in a decrease in the generalization prediction accuracy of subsequent machine learning models when facing heterogeneous samples such as deep and large foundation pits.

[0045] In specific implementation, based on the excavation depth, the original geological soil structure of the foundation pit to be predicted within the influence depth range is divided into a preset number of equivalent layers, including: obtaining the total soil depth of the foundation pit to be predicted; determining whether the total soil depth is greater than a preset depth threshold; if the total soil depth is not greater than the preset depth threshold, then a standard number of layers is used as the preset number, and a standard division method based on a fixed proportion of the excavation depth is executed; if the total soil depth is greater than the preset depth threshold, then the ratio of the total soil depth to the excavation depth is calculated, and based on the ratio... The thickness expansion coefficient and the total number of adaptive equivalent layers are determined by the value and the standard number of layers. The influence depth range is divided into three longitudinal intervals, wherein the first interval and the second interval are each fixedly divided into two equivalent layers, and the number of layers to be divided in the third interval is determined according to the difference between the total number of layers and the fixed number of layers. Based on the thickness expansion coefficient, the standard thickness of each equivalent layer is calculated layer by layer for the number of layers to be divided in the third interval. The longitudinal depth boundary value of each equivalent layer is determined by accumulating the standard layer thicknesses in sequence, and the influence depth range is divided into an adaptive equivalent layer structure with more than the standard number of layers.

[0046] Specifically, the total depth of the soil in the foundation pit to be predicted is obtained, and a preset depth threshold is retrieved for comparison. If the total soil depth is not greater than the preset depth threshold, the affected depth range is directly divided into N_std layers based on a fixed ratio of the excavation depth. N_std is the number of standard layers, with a value range of 5 ≤ N_std ≤ 10. In this embodiment, N_std = 5 is preferred, locking the standard layer thickness and longitudinal boundary of each layer. If the total soil depth is greater than the preset depth threshold, the ratio r of the total soil depth to the excavation depth under the current working condition is calculated. The total number of layers N is determined based on the difference between the ratio r and the preset ratio threshold r_th. The calculation formula is N = N_std + floor(( The formula is r-r_th)×k), where N_std is the standard number of layers, r_th is the preset ratio threshold, k is the layer expansion coefficient and k>0, and floor(·) is the floor function for rounding down. The total number of adaptive equivalent layers N ranges from N_std≤N≤N_max, where N_max is the maximum number of equivalent layers allowed by the system, and its range is 10≤N_max≤20. When the N value calculated by the above formula exceeds N_max, N_max is taken as the final value of the total number of layers N. The layer thickness expansion coefficient λ is determined based on the ratio r, and the calculation formula is λ=1+(r-1)×μ, where μ is the preset expansion rate parameter and μ>0, and λ increases with the increase of the ratio r. Furthermore, from top to bottom, the entire vertical space of the foundation pit from the ground surface to the total depth of the soil is decoupled and divided into three longitudinally continuous and interconnected longitudinal intervals with different mechanical response characteristics. The first interval is the active stress interval between the ground surface and the excavation face. This interval is fixedly and evenly divided into the first and second equivalent layers, and the thickness of each layer is directly equal to half the excavation depth. The second interval is the passive core disturbance interval from the excavation face to twice the excavation depth. This interval is fixedly and evenly divided into the third and fourth equivalent layers, and the thickness of each layer is also directly equal to half the excavation depth. The third interval is the deep stress attenuation interval from twice the excavation depth to the total depth of the soil. Apart from the N_fix equivalent layers that are fixedly occupied by the first two intervals, N_fix is ​​a fixed number of layers with a value of 4. That is, the first and second intervals are each fixedly divided into 2 equivalent layers. The remaining equivalent layers that are dynamically increased due to the scale are all assigned to this deep stress attenuation interval. The remaining total thickness H_remain of this interval is equal to the total depth of the soil minus twice the excavation depth. The number of layers to be divided is n = N - N_fix, where the value of n is 1 ≤ n ≤ N_max - N_fix, and n is a positive integer. When calculating the longitudinal boundaries of each equivalent layer within the third interval, considering the nonlinear weakening of excavation disturbance and soil forces in the depth direction, a gradient logic of increasing layer thickness from top to bottom is adopted, that is, a shallow-dense and deep-sparse structure where the layer thickness increases as you go deeper.First, calculate the initial standard layer thickness h_(N_fix+1) of the topmost equivalent layer in the third interval, i.e., the (N_fix+1)th equivalent layer in the entire configuration. The calculation formula is h_(N_fix+1)=H_remain / (1+λ+λ²+…+λ^(n-1)), where the denominator is the sum of n terms of the layer thickness expansion coefficient. Then, by sequentially multiplying the standard layer thickness calculated for the previous layer by the layer thickness expansion coefficient λ, the subsequent layers in this interval can be recursively derived. The actual standard layer thickness of the equivalent layer is h_(N_fix+1+i)=h_(N_fix+1)×λ^i, where i=0,1,2,…,n-1; Finally, using the vertical coordinate of twice the excavation depth as the starting reference, the layer thickness of each equivalent layer in the calculated third interval is continuously accumulated downwards, and the upper and lower boundary coordinates of the longitudinal depth of each layer are output sequentially, and the lower boundary coordinate of the last layer is exactly aligned and closed with the bottom boundary of the total depth of the soil in the foundation pit to be predicted.

[0047] The method provided in this embodiment breaks through the fundamental technical bottleneck of rigid spatial geometric boundary division in traditional geotechnical engineering characteristic engineering, which cannot perceive the abrupt changes in the scale of the foundation pit itself, by introducing a threshold-based determination of the total soil depth and a method for calculating the layer thickness expansion coefficient based on the ratio of total depth to excavation depth and determining the longitudinal boundary layer by layer-by-layer recursion. For foundation pits of conventional scale, maintaining a five-layer standard layer with minimal computational overhead and focused physical representation can efficiently anchor the sensitive areas of active and passive earth pressure. For high-risk deep foundation pits, by calculating the total number of layers and the layer thickness expansion coefficient based on the ratio, and using geometric series summation and layer-by-layer recursion to determine the longitudinal depth boundary of each equivalent layer, the longitudinal depth boundary of the virtual layer can deterministically fit the nonlinear stress release and transmission gradient in the deep soil. This fundamentally prevents the problem of stress abrupt change surface in the passive zone at the bottom of the pit and the problem of deep base constraint features being crudely smoothed or diluted and contaminated when spatially overlapping and distributed in ultra-deep foundation pits, greatly improving the mechanical representation accuracy of the finally constructed standardized feature vector when facing extremely heterogeneous engineering samples.

[0048] Optionally, when the standard number of layers is 5, the standard division method is as follows: the original geological soil structure is divided into five standard equivalent layers in the vertical direction. The depth range of the first equivalent layer is from the ground surface to half of the excavation depth, and the thickness is equal to half of the excavation depth; the depth range of the second equivalent layer is from half of the excavation depth to the excavation face, and the thickness is equal to half of the excavation depth; the depth range of the third equivalent layer is from the excavation face to twice the excavation depth, and the thickness is equal to the excavation depth; the depth range of the fourth equivalent layer is from twice the excavation depth to twice the excavation depth plus half of the total thickness of the remaining soil, and the thickness is equal to half of the total thickness of the remaining soil; the depth range of the fifth equivalent layer is from twice the excavation depth plus half of the total thickness of the remaining soil to the total thickness of the soil, and the thickness is equal to half of the total thickness of the remaining soil; wherein, the total thickness of the remaining soil is the thickness value obtained by subtracting twice the excavation depth from the total soil thickness.

[0049] Specifically, the system retrieves the excavation depth value corresponding to the current working condition and the total soil thickness value set by the site survey from the system's cache. The remaining total soil thickness value is calculated by subtracting twice the excavation depth value from the total soil thickness value. Starting from the vertical coordinates of the ground surface and ending at half the excavation depth, the depth range of the first equivalent layer is defined, and the standard layer thickness of this layer is directly assigned to half the excavation depth. Starting from half the excavation depth and ending at the vertical coordinates of the excavation face, the depth range of the second equivalent layer is defined, and the standard layer thickness of this layer is directly assigned to half the excavation depth. Starting from the vertical coordinates of the excavation face and ending at twice the excavation depth, the depth range of the third equivalent layer is defined, and the standard layer thickness of this layer is directly assigned to the excavation depth. Starting from twice the excavation depth and ending at twice the excavation depth plus half the total thickness of the remaining soil, the depth range of the fourth equivalent layer is defined, and the standard layer thickness of this layer is directly assigned to half the total thickness of the remaining soil. Starting from twice the excavation depth plus half the total thickness of the remaining soil and ending at the total thickness of the soil, the depth range of the fifth equivalent layer is defined, and the standard layer thickness of this layer is directly assigned to half the total thickness of the remaining soil.

[0050] For example, in one embodiment, the five equivalent layers are as follows:

[0051] First equivalent layer (Up1): Depth range [0, ExcDepth / 2], thickness = ExcDepth / 2;

[0052] Second equivalent layer (Up2): Depth range [ExcDepth / 2, ExcDepth], thickness = ExcDepth / 2;

[0053] Third equivalent layer (Down1): Depth range [ExcDepth, 2×ExcDepth], thickness = ExcDepth;

[0054] Fourth equivalent layer (Down2): Depth range [2×ExcDepth, 2×ExcDepth + H_remain / 2, thickness = H_remain / 2;

[0055] Fifth equivalent layer (Down3): Depth range [2×ExcDepth + H_remain / 2, total thickness], thickness = H_remain / 2;

[0056] Where H_remain = total thickness - 2 × ExcDepth, and ExcDepth is the excavation depth.

[0057] The method provided in this embodiment deeply binds the geometric boundaries of the virtual equivalent layer to the dynamically changing excavation depth and total soil thickness, and precisely divides it into five standard layers with a specific ratio. First, dynamic segmentation based on the excavation depth directly solves the underlying technical bottleneck of the inability of machine learning models to accept fixed-dimensional input tensors due to the variable number and uneven thickness of natural strata, achieving efficient conversion of unstructured geological data to standard digital interfaces. Simultaneously, this specific layered geometric configuration deeply aligns with the mechanical response mechanism during foundation pit excavation. The fine and uniform division of the active soil zone from the surface to the excavation face into first and second equivalent layers enables high-resolution capture of the nonlinear active earth pressure distribution behind the retaining structure and the peak characteristics of wall lateral bending deformation, while the core disturbed zone from below the excavation face to twice the excavation depth is further subdivided. By independently designating the area as the third equivalent layer, the key mechanically sensitive areas of pit bottom heave, passive zone soil resistance, and deep horizontal displacement transmission are precisely identified. Furthermore, by dividing the deep soil below twice the excavation depth into the fourth and fifth equivalent layers based on the remaining thickness, the gradient attenuation effect and the constraint state of the deep foundation during the downward transmission of excavation stress are perfectly simulated. Thus, without losing the accuracy of geological stratification information, the input feature vector is endowed with extremely strong geotechnical engineering mechanical characterization capabilities, significantly improving the generalization accuracy and training stability of the machine learning model in predicting multi-dimensional indicators such as deep horizontal displacement of the retaining structure, surface settlement, and pile top horizontal displacement.

[0058] S104. Determine the spatial overlap mapping relationship between each equivalent layer and each original geological soil layer in the vertical direction. Based on the spatial overlap mapping relationship, perform spatial weighted average distribution on the corresponding original geological soil layer parameters to obtain the standard layer thickness and standardized physical and mechanical parameters of each equivalent layer, and combine them to construct a standardized feature vector.

[0059] Optionally, the standardized feature vector includes the excavation depth parameter of the pit to be predicted, and the standard layer thickness parameter, cohesion parameter, internal friction angle parameter, and compressibility modulus parameter corresponding to each equivalent layer according to the physical topological order of vertical space; wherein, the standardized feature vector is constructed by taking the excavation depth parameter as the starting dimension and sequentially spatially cascading and splicing the standard layer thickness parameter, cohesion parameter, internal friction angle parameter, and compressibility modulus parameter corresponding to each of the divided equivalent layers, and combining them together to form the standardized feature vector.

[0060] Specifically, the original geological soil layer refers to the original, undisturbed soil or rock layers that exist naturally within the site of the foundation pit project to be predicted, without any artificial simplification or mathematical smoothing. These soil layers directly correspond to the natural strata (such as miscellaneous fill, silty clay, silty soil, strongly weathered sandstone, etc.) identified in the geotechnical engineering investigation report. Spatially, each original geological soil layer has irregular physical boundaries, a natural thickness that varies with the borehole location, and its own independent initial mechanical response characteristics.

[0061] Spatial overlap mapping refers to the geometric overlap and topological inclusion quantity relationship between an artificially divided virtual standardized equivalent layer and naturally existing original geological soil layers in vertical depth space. Spatial overlap mapping is used to quantitatively describe which original geological soil layers are contained within a specific target equivalent layer, and how many meters of spatial thickness each original geological soil layer occupies within the vertical span of the equivalent layer.

[0062] Standard layer thickness refers to the standard total span or nominal thickness of each equivalent layer in vertical space. This thickness is not the actual natural thickness of the soil, but a standardized spatial scale calculated strictly according to a preset proportional relationship based on the excavation depth under the current working conditions.

[0063] Standardized physical and mechanical parameters refer to a set of standardized indices that represent the comprehensive engineering mechanical properties of a specific virtual equivalent layer after overall equivalent transformation through a spatial weighted averaging algorithm. Standardized physical and mechanical parameters specifically include, but are not limited to, cohesion, internal friction angle, and compression modulus. It is a composite index formed by aggregating the parameters of all overlapping original geological soil layers within the vertical span of the equivalent layer according to their actual overlap thickness. This preserves the accuracy of layered geological information while eliminating the disturbance to the input structure caused by the variable number of original soil layers.

[0064] A standardized feature vector refers to a structured one-dimensional array or low-dimensional matrix tensor formed by cascading and splicing all geometric parameters and spatial equivalent parameters characterizing the current foundation pit engineering properties in a strict and fixed topological order. In this application, the standardized feature vector uses the excavation depth parameter of the foundation pit to be predicted as the starting dimension. Then, in accordance with the physical order from shallow to deep in vertical space, the four parameters (standard layer thickness, cohesion, internal friction angle, and compressive modulus) corresponding to each equivalent layer are cascaded and spliced ​​together to form the standardized feature vector.

[0065] In specific implementation, for each equivalent layer, the top and bottom vertical coordinates of the equivalent layer are introduced, and the span of the equivalent layer in vertical space is calculated. The span is used as the corresponding standard layer thickness. The vertical depth range of any target equivalent layer is compared with the natural depth range of each original geological soil layer. Based on the spatial overlap mapping relationship, it is divided into inclusion relationship, partial overlap relationship, and no overlap relationship. For target equivalent layers that are determined to be inclusion relationship or partial overlap relationship, the actual overlap thickness between the target equivalent layer and the corresponding original geological soil layer is calculated as a spatial weighting weight. The physical and mechanical parameters of each overlapping original geological soil layer are multiplied by the corresponding actual overlap thickness to obtain a parameter thickness product term. All parameter thickness product terms corresponding to the target equivalent layer are summed and then divided by the standard layer thickness to obtain the corresponding standardized physical and mechanical parameters. For target equivalent layers that are determined to be no overlap relationship, when the geometric overlap thickness is zero, the segmentation coordinates of the spatially adjacent actual original geological soil layers are extracted. The parameters of the adjacent original geological soil layers are interpolated and reconstructed using the neighbor distance inverse weighting algorithm to obtain the standardized physical and mechanical parameters of the target equivalent layer.

[0066] Specifically, for each defined equivalent layer, the top and bottom vertical coordinates of the equivalent layer are automatically imported and retrieved. The top vertical coordinate is subtracted from the bottom vertical coordinate to calculate the total span of the equivalent layer in vertical space, which is then used as the corresponding standard layer thickness. Next, the vertical depth range of any target equivalent layer used as the current calculation objective is compared and aligned with the natural depth ranges of each naturally existing original geological soil layer in the site. Based on the spatial overlap, their topological mapping relationship is strictly divided into three categories: inclusion relationship, partial overlap relationship, and no overlap relationship. For target equivalent layers determined to be inclusion or partial overlap relationships, the actual overlap thickness between the target equivalent layer and each overlapping original geological soil layer in the vertical depth direction is calculated, and the actual overlap thickness is used as the spatial weighting weight. The corresponding material of each overlapping original geological soil layer is then compared and aligned. The physical and mechanical parameters are multiplied by the corresponding actual overlap thickness to obtain the product terms of each parameter thickness. Then, all parameter thickness product terms corresponding to the same target equivalent layer are summed to output the weighted total value of the parameters. Finally, the weighted total value of the parameters is divided by the standard layer thickness to obtain the standardized physical and mechanical parameters of the target equivalent layer. For target equivalent layers that are determined to be completely non-overlapping, in the extreme case where the geometric overlap thickness is zero, the system automatically extracts the segmented coordinates of the nearest adjacent actual original geological soil layer in vertical space, calculates the vertical distance between the target equivalent layer and each adjacent actual original geological soil layer, and uses the proximity distance inverse weighting algorithm to interpolate and reconstruct the physical and mechanical parameters of the adjacent actual original geological soil layers according to the ratio of greater weight for closer distances and smaller weights for farther distances, and calculates and outputs the standardized physical and mechanical parameters of the completely non-overlapping target equivalent layer.

[0067] The method provided in this embodiment decouples the geometric overlap between the virtual equivalent layer and the actual natural layer into three types of topological relationships for classification and discussion. It also adopts a spatial weighted amortization based on the actual overlap thickness and a neighbor distance inverse weighted interpolation algorithm to overcome the severe loss of mechanical accuracy caused by the use of coarse pure geometric volume amortization in traditional geotechnical feature engineering, as well as the underlying technical defects of singular divergence of the input matrix under critical heterogeneous samples. On the one hand, the physical and mechanical parameters of each overlapping original geological soil layer are multiplied by the corresponding actual overlap thickness, summed, and then divided by the standard layer thickness. The actual overlap thickness is used as the spatial weighting weight for amortization calculation, which deeply matches the volume ratio and stress transmission contribution of different soil physical quantities in the spatial vertical stratification. This greatly improves the engineering characterization accuracy of the input feature vector for the actual stress state of the retaining structure at the foundation pit engineering site. On the other hand, for completely non-overlapping data faults with zero geometric thickness, a stable singular value lossless data compensation mechanism is constructed at the input end of the machine learning algorithm by dynamically extracting the segmentation coordinates of adjacent layers and implementing distance inverse weighted interpolation. This eliminates singular noise such as zero or null values ​​in the data feature matrix due to spatial misalignment from the source.

[0068] Optionally, in addition to the calculation method based on spatial overlap mapping relationship described above, a thickness-weighted average method can also be used: Based on the vertical depth boundaries of each equivalent layer, the span of each equivalent layer in vertical space is determined as the corresponding standard layer thickness; for any target equivalent layer, each original geological soil layer overlapping with the target equivalent layer is identified in vertical space; the actual overlap thickness between the target equivalent layer and each overlapping original geological soil layer in the vertical direction is calculated as the spatial weighting weight; the physical and mechanical parameters of each overlapping original geological soil layer are multiplied by the corresponding actual overlap thickness to obtain the corresponding parameter thickness product terms; all parameter thickness product terms corresponding to the target equivalent layer are summed to obtain the parameter weighted total value; the parameter weighted total value is divided by the standard layer thickness to obtain the standardized physical and mechanical parameters of the target equivalent layer; the standardized physical and mechanical parameters include cohesion, internal friction angle, and compression modulus.

[0069] Specifically, the upper and lower boundary coordinates of the vertical depth of each of the five equivalent layers are read. The upper boundary coordinate is subtracted from the lower boundary coordinate of each equivalent layer to calculate the corresponding vertical spatial span. This span value is determined as the standard layer thickness of the equivalent layer. The natural depth boundary data of each original geological soil layer in the geological exploration report are imported. Any equivalent layer is selected as the target equivalent layer, and its vertical depth range is compared with the natural depth ranges of all original geological soil layers. Each original geological soil layer that intersects with the depth range of the target equivalent layer is selected, and an overlapping stratum index list is established. For each original geological soil layer in the overlapping stratum index list, the vertical start and end coordinates that coincide with the target equivalent layer are extracted. The difference between the end coordinate and the start coordinate is calculated to obtain the corresponding actual overlap thickness. The actual overlap thickness is used as the spatial weighting weight for subsequent calculations. Retrieve the original mechanical parameters of each overlapping original soil layer, including original cohesion, original internal friction angle, and original compression modulus. Multiply the original cohesion, original internal friction angle, and original compression modulus of each overlapping original soil layer by the actual overlap thickness calculated in the previous steps to obtain the corresponding cohesion-thickness product, internal friction angle-thickness product, and compression modulus-thickness product terms. Sum the cohesion-thickness products of all original soil layers corresponding to the target equivalent layer to obtain the weighted total cohesion value. Similarly, sum the internal friction angle-thickness product and the compression modulus-thickness product terms to obtain the corresponding weighted total internal friction angle and weighted total compression modulus values. Divide the weighted total cohesion value, weighted total internal friction angle value, and weighted total compression modulus value by the standard layer thickness of the target equivalent layer calculated in the previous steps to output the standardized cohesion, standardized internal friction angle, and standardized compression modulus of the target equivalent layer. Repeat the above steps to complete the standardized parameter calculation for all five equivalent layers.

[0070] For example, in one embodiment, the standardized physical and mechanical parameters can be expressed as:

[0071] ;

[0072] in, Standardized physical and mechanical parameters; This represents the actual overlap thickness. The physical and mechanical parameters of the original geological soil layers that overlap; This represents the number of equivalent layers.

[0073] The method provided in this embodiment identifies the vertical spatial overlap boundary between the virtual equivalent layer and the natural original soil layer, and uses the actual overlap thickness as the spatial weighting weight for amortization calculation, ultimately constructing a standardized feature vector through cascading. First, by topologically mapping irregular multi-layered structures to a fixed five-layer structure, a fixed 21-dimensional standardized feature vector is forcibly constrained and output, achieving efficient conversion of unstructured geological data into a standard digital interface for algorithm models. This directly solves the underlying technical bottleneck in geotechnical engineering where the number of soil layers and the extremely uneven layer thickness in geological survey reports for different sites cannot be aligned with the dimensions of the machine learning input matrix. Furthermore, by using the amortization process of multiplying the actual overlap thickness by physical and mechanical parameters and dividing by the standard layer thickness, it overcomes the limitations of traditional numerical simulations or empirical formulas that simplify strata or use simple arithmetic averages to obtain high-value geological information. Despite the limitations of heavy loss, this method can faithfully reproduce the spatial volume ratio and stress transmission contribution of each natural original soil layer within a specific equivalent depth range. This results in the calculated standardized cohesion, internal friction angle, and compression modulus having strong geophysical realism, thereby endowing the standardized feature vector with extremely high accuracy in characterizing the stratum shear strength and deformation resistance. This significantly reduces the risk of model overfitting caused by geological uncertainties, significantly enhances the convergence stability of multiple machine learning algorithms during parallel training, and significantly improves the synchronous intelligent prediction accuracy of multi-dimensional indicators such as the horizontal displacement of the retaining structure, surface settlement, and pile top horizontal displacement.

[0074] S105. Based on the historical dataset, perform parallel training on multiple machine learning models, and adaptively match the best prediction model for each foundation pit deformation index according to the test performance index.

[0075] Specifically, machine learning models can automatically learn the complex nonlinear mapping relationship between geological geometric features and foundation pit deformation response from input historical data samples through mathematical optimization algorithms. In this application, in order to cover the fitting requirements of different mechanical features, the machine learning models specifically include multiple regression algorithm models with different mathematical principles, including but not limited to: linear regression models as linear benchmarks, random forest models using bagged ensemble methods and containing a preset number (e.g., one hundred) decision trees, support vector regression models (SVR) using radial basis function (RBF) kernels, and Gaussian process regression models (GPR) using squared exponential kernels.

[0076] Performance metrics refer to mathematical indicators used to quantitatively evaluate and verify the generalization prediction accuracy, nonlinear fit goodness, and error control level of various machine learning models for unknown foundation pit samples on independent test sets after model training is completed. In this application, in order to accurately measure the model's ability to capture multivariate deformation trends, the coefficient of determination (R²) is used. 2The Coefficient of Determination (CDE) is the core performance metric used to characterize the proportion of variance of the actual monitored values ​​explained by the predicted values. Optional performance metrics may also include, but are not limited to, root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and residual standard deviation.

[0077] In specific implementation, the historical dataset is standardized by standard deviation and randomly divided into training and test sets according to a preset partitioning ratio. For different foundation pit deformation indices, multiple different regression algorithm models are trained in parallel using the training set. These regression algorithm models include a linear regression model as a linear baseline, a random forest model using a bagged ensemble method and containing one hundred decision trees, a support vector regression model using a radial basis function kernel, and a Gaussian process regression model using a squared exponential kernel. The test set is input into each of the regression algorithm models trained for each foundation pit deformation index, and the determination coefficient of each regression algorithm model under the corresponding foundation pit deformation index is calculated. The determination coefficient is used as a test performance index. For any foundation pit deformation index, the magnitude of the determination coefficients output by each regression algorithm model is compared, and the regression algorithm model with the highest determination coefficient is determined as the best prediction model.

[0078] Specifically, the acquired historical dataset is standardized using standard deviation (Z-score standardization) to eliminate the influence of different physical dimensions on model training. Following a preset partitioning ratio (e.g., 7:3), the standardized historical dataset is randomly divided into independent training and test sets. Independent parallel computing channels are established for different foundation pit deformation indices. Within each deformation index's corresponding computing channel, the training set is input into four different regression algorithm models for parallel iterative training. After all regression algorithm models are trained, the input features of the test set are input into the respective models trained for each foundation pit deformation index, outputting the corresponding predicted value sequence. The coefficient of determination (R²) of each regression algorithm model under the corresponding foundation pit deformation index is calculated by comparing it with the actual monitoring label values ​​in the test set. 2 The calculated coefficients of determination are used as performance metrics for the model. For any excavation pit deformation index, the coefficients of determination of four models trained under that index—linear regression, random forest, support vector regression, and Gaussian process regression—are compared. The regression algorithm model with the highest coefficient of determination is selected and identified as the best prediction model for that specific excavation pit deformation index.

[0079] For example, in one embodiment, the standard deviation standardization process can be expressed as:

[0080] X_norm = (X – mean(X)) / std(X);

[0081] Where X_norm is the data after standardization; X is the data in the historical dataset before standardization; mean(X) is the average value of the data in the historical dataset before standardization; and std(X) is the number of samples in the historical dataset.

[0082] Optionally, before training multiple machine learning models in parallel based on the historical dataset, the method further includes: performing univariate linear fitting with the excavation depth as the independent variable and the monitoring value of the corresponding foundation pit deformation index as the dependent variable, calculating the fitting residuals for each sample, and removing samples whose absolute value of the fitting residuals exceeds 1.2 times the standard deviation of the residuals; for the retained samples, calculating the lower quartile, upper quartile, and interquartile range of the corresponding foundation pit deformation index monitoring value; removing samples whose values ​​are lower than the lower quartile minus 1.5 times the interquartile range, or whose values ​​are higher than the upper quartile plus 1.5 times the interquartile range, to obtain the filtered historical dataset.

[0083] Specifically, the excavation depth of each sample is extracted from the historical dataset as the independent variable, and the corresponding pit deformation index monitoring value is extracted as the dependent variable. A univariate linear regression is performed on the independent and dependent variables to determine the linear regression trend line. For each data sample, the difference between the actual pit deformation index monitoring value and the predicted value of the linear regression trend line is calculated to obtain the fitting residual for each sample. The standard deviation of the fitting residual set corresponding to all data samples is calculated, and the residual standard deviation is output. The residual standard deviation is multiplied by 1.2 as the upper limit threshold for the absolute value of the residuals in the initial filtering. Each data sample in the historical dataset is traversed, its fitting residual is extracted, and its absolute value is calculated. It is determined whether the absolute value of the fitting residual exceeds the upper limit threshold. If it does, the entire row of the sample is deleted from the historical dataset; if it does not exceed the threshold, the sample is temporarily retained. The samples retained after the first stage of outlier extraction are sorted in ascending order of their corresponding pit deformation index monitoring values ​​to construct a one-dimensional deformation numerical sequence. A quantile statistical function is called to calculate the lower quartile at the 25th percentile and the upper quartile at the 75th percentile. The interquartile range is calculated by subtracting the lower quartile from the upper quartile. The lower threshold is obtained by subtracting 1.5 times the interquartile range from the lower quartile; the upper threshold is obtained by adding 1.5 times the interquartile range to the upper quartile. The pit deformation index monitoring values ​​of each sample in the retained samples are read one by one and compared with the lower and upper thresholds. If the deformation index monitoring value of a sample is lower than the lower threshold or higher than the upper threshold, the data sample is removed from the set. All the structured data samples remaining after two rounds of outlier filtering are reassembled, output, and stored as the filtered historical dataset.

[0084] S106. Input the standardized feature vector of the foundation pit to be predicted into the best prediction model corresponding to each foundation pit deformation index, and output the predicted values ​​of multiple foundation pit deformation indices simultaneously.

[0085] Specifically, the predicted value refers to the scientifically estimated value that each adaptive matching locked best prediction model calculates and outputs synchronously after receiving the standardized feature vector of the foundation pit to be predicted and performing nonlinear mapping inference through the internal regression algorithm. This value is used to quantitatively characterize the various safety deformation states of the foundation pit under the current target excavation conditions.

[0086] In specific implementation, the planned excavation depth values ​​corresponding to the foundation pit to be predicted under multiple consecutive construction steps are obtained, and a time series sequence of planned excavation depths is constructed. For each planned excavation depth value in the time series sequence, the spatial geometric partitioning and parameter space weighting and amortization steps are repeatedly executed to generate standardized feature vectors, which are then combined according to the physical topological order of the consecutive construction steps to construct a time series matrix of feature vectors to be predicted. The time series matrix of feature vectors to be predicted is input into the best prediction model corresponding to each foundation pit deformation index for parallel inference, and the predicted value sequence of multiple foundation pit deformation indices under multiple consecutive construction steps is output synchronously to form a continuous time series prediction change curve that dynamically evolves with the construction steps. Based on the continuous time series prediction change curve, the maximum predicted value of each foundation pit deformation index is extracted, and it is determined whether the maximum predicted value exceeds the preset safety control threshold of the corresponding foundation pit deformation index. If it exceeds, the corresponding safety warning response is triggered according to the severity of the exceedance.

[0087] Specifically, the planned excavation depth values ​​for the foundation pit under multiple consecutive construction steps are obtained from the design scheme or construction organization plan. These planned excavation depth values ​​are then arranged in chronological order to form a time-series sequence of planned excavation depths. A cyclical calculation architecture is established, sequentially extracting each planned excavation depth value from the time-series sequence as the target depth for the current cycle. For each target depth, the spatial geometric partitioning logic and parameter space weighted amortization algorithm from the previous steps are called to calculate the boundaries and standardized mechanical parameters of the five equivalent layers corresponding to the current depth, generating a standardized feature vector with a fixed 21-dimensional dimension. The series of standardized feature vectors generated for each consecutive construction step are then vertically concatenated according to the physical topological order (chronological order) of the corresponding consecutive construction steps. This outputs a time-series matrix of the predicted feature vectors, consisting of multiple 21-dimensional vectors arranged horizontally or vertically with a fixed dimension. The time series matrix of the feature vector to be predicted is simultaneously input into the optimal prediction model, which is independently configured and locked for each foundation pit deformation index. The matrix inference program of each optimal prediction model is run in parallel, and the predicted value sequence of each foundation pit deformation index under multiple consecutive construction steps is output synchronously. The output predicted value sequence is mapped with the construction step as the abscissa and the deformation prediction value as the ordinate, and a continuous time series prediction change curve that dynamically evolves with the construction step is automatically generated. All discrete data points on the continuous time series prediction change curve are traversed and read, and the maximum predicted value of each foundation pit deformation index in the whole construction cycle is selected and extracted. The preset safety control threshold corresponding to each foundation pit deformation index is retrieved from the system configuration. The maximum predicted value is compared with the preset safety control threshold to determine whether it exceeds the threshold. If the result is that it exceeds the threshold, the absolute deviation or percentage of the maximum predicted value exceeding the safety control threshold is calculated, and the severity level is classified accordingly. The corresponding level of safety warning response signal is triggered and output to the on-site engineering safety management terminal.

[0088] Optionally, after simultaneously outputting the predicted values ​​of multiple foundation pit deformation indices, the method further includes: acquiring real-time on-site deformation monitoring data of the foundation pit to be predicted during the excavation stage; comparing the real-site deformation monitoring data with the predicted values ​​of the corresponding construction stage in real time to calculate the excavation prediction residual; determining whether the absolute value of the excavation prediction residual exceeds the set safety allowable threshold; if the absolute value of the excavation prediction residual exceeds the safety allowable threshold, inputting the real-site deformation monitoring data as a boundary feedback constraint into the inversion correction algorithm to inversely correct the standardized physical and mechanical parameters of each equivalent layer, or dynamically adjusting the internal weight coefficients of the optimal prediction model; reconstructing the standardized feature vector using the inversely corrected standardized physical and mechanical parameters or internal weight coefficients and performing prediction to obtain updated predicted values ​​of multiple foundation pit deformation indices.

[0089] Specifically, by connecting to automated monitoring hardware devices deployed at the foundation pit site (including deep horizontal displacement gauges, inclinometers, hydrostatic levels, and total stations), real-time online data on the actual deformation of the foundation pit during the actual excavation stage is collected. The predicted values ​​of various foundation pit deformation indicators output in the preceding steps, corresponding to the current construction stage, are retrieved. The predicted values ​​for the corresponding construction stage are subtracted from the actual deformation monitoring data to calculate and output the excavation prediction residuals corresponding to each deformation indicator. The absolute value of each excavation prediction residual is calculated. The absolute value of the excavation prediction residual is compared with the system's built-in preset safety allowable threshold to determine if it exceeds the limit. If the result is within the limit, the existing model structure is maintained and prediction continues. If the result is exceeded, the inversion correction algorithm module is immediately triggered in the system. The actual deformation monitoring data that triggers the exceedance is used as boundary feedback constraints and input into the inversion correction algorithm configured in the system. The inversion correction algorithm is run to optimize the parameter space and inversely correct the values ​​of the standardized physical and mechanical parameters of each equivalent layer defined in the previous steps. Alternatively, the inversion correction algorithm is run to adjust the internal weights of the currently used best prediction model and dynamically update the internal weight coefficients of the best prediction model. The standardized physical and mechanical parameters after inverse correction are substituted back in, or the updated internal weight coefficients are directly applied, and feature concatenation is performed again to reconstruct the corrected standardized feature vector. The reconstructed standardized feature vector is then input back into the corresponding prediction model to perform regression inference, and the updated predicted values ​​of various foundation pit deformation indices, which have been dynamically corrected in real time, are output simultaneously.

[0090] The method provided in this embodiment, firstly, does not rigidly and forcibly adopt a single fixed spatial geometric division within the influence depth range. Instead, it introduces a threshold judgment and grading mechanism based on the total soil depth. In the case of ultra-deep large foundation pits, it calculates the layer thickness expansion coefficient and the total number of adaptive equivalent layers based on the ratio of the total soil depth to the excavation depth. Through layer-by-layer recursive calculation, the longitudinal boundary is dynamically adjusted to an adaptive equivalent layer structure of more than five layers. This not only directly solves the underlying technical bottleneck of the non-fixed number and uneven thickness of natural strata, which prevents machine learning models from accepting fixed-dimensional input tensors, and achieves efficient conversion of unstructured geological data to standard digital interfaces, but also overcomes the shortcomings of traditional rigid geometric stratification in its mechanical response mechanism. This is because rigid geometric stratification cannot perceive the rapid changes in the scale of the foundation pit itself, and the forced uniform stratification smoothing dilutes key mechanical features. It endows the feature engineering with extremely strong geotechnical engineering mechanical targeted characterization capabilities, significantly improving the generalization accuracy and training stability of the model when predicting multiple indicators under ultra-deep and extremely heterogeneous conditions.

[0091] Secondly, in the spatial mapping and data aggregation stage of the feature map, the vertical coordinates of the top and bottom of the equivalent layer are introduced to accurately lock the standard layer thickness. The geometric overlap state between the virtual equivalent layer and the actual natural layer is strictly decoupled and divided into three categories: inclusion relationship, partial overlap relationship, and no overlap relationship. For the overlapping case, the arithmetic mean of the pure geometric layer thickness is abandoned. The physical and mechanical parameters of each overlapping original geological soil layer are multiplied by the corresponding actual overlap thickness to obtain the parameter thickness product term. The sum of all parameter thickness products is divided by the standard layer thickness to obtain the standardized physical and mechanical parameters. For the extreme case of no overlap where the geometric overlap thickness is zero, the segmentation coordinates of the spatially adjacent actual layers are dynamically extracted and the neighbor distance inverse weighted interpolation algorithm is implemented to reconstruct the standardized physical and mechanical parameters in reverse. Finally, the standardized physical and mechanical parameters generated by spatial weighting of each equivalent layer and the excavation depth parameters are cascaded and spliced ​​according to the preset physical topology order to jointly construct a standardized feature vector. It overcomes the shortcomings of traditional empirical formulas or numerical simulations that lead to a serious loss of high-value mechanical accuracy due to the simplification of strata. It can faithfully reproduce the volume ratio and stress transmission contribution of different physical quantities in the vertical transmission of space. Furthermore, it has built a stable singular value lossless data compensation mechanism at the input end of the machine learning regression algorithm, eliminating singular noise such as zero values ​​or empty values ​​caused by spatial misalignment or lack of geological exploration from the source.

[0092] Thirdly, to address the need for continuous prediction, the system cleverly acquires the planned excavation depth values ​​of multiple future construction steps to construct a time series sequence. It dynamically generates a time series matrix of the feature vector to be predicted and inputs it into the best prediction model. This allows for the synchronous output of the predicted value sequence under multiple future continuous construction steps and the generation of continuous time series prediction change curves. This avoids the model dimension mismatch and crash errors caused by forcibly stuffing heterogeneous features during the model inference stage. Under the premise of ensuring that the feature dimensions are absolutely fixed, it perfectly simulates the continuous time series process of the dynamic evolution of the foundation pit deformation with the construction steps. It can accurately capture the maximum predicted value within the entire construction cycle to determine the severity of exceeding the limit, thereby triggering a graded automated safety early warning response.

[0093] Fourthly, after outputting the predicted value, the system acquires real-time deformation monitoring data from the excavation stage and calculates the excavation prediction residual. If the value exceeds the limit, the real data is used as boundary feedback constraints and input into the inversion correction algorithm to inversely correct the standardized parameters of the equivalent layer or dynamically adjust the internal weight coefficients of the model. Then, the vector is reconstructed to perform prediction and output dynamically updated predicted values. A closed-loop dynamic correction mechanism of "prediction-monitoring-residual analysis-inversion correction-rolling update" has been successfully constructed. This overcomes the cumulative error and prediction lag defects that are easily generated by pure data-driven models when the excavation stress environment fluctuates drastically. The system uses real-time feedback from the field to perform dynamic calibration and accuracy compensation on the intelligent prediction model, and finally achieves ultra-high accuracy and long-term robust prediction of multi-dimensional foundation pit deformation indicators under complex and variable geotechnical conditions.

[0094] Corresponding to the aforementioned embodiment of an intelligent prediction method for foundation pit deformation, this application also provides an embodiment of an intelligent prediction device for foundation pit deformation.

[0095] Figure 2 This is a structural schematic diagram of the intelligent prediction device for foundation pit deformation provided in Embodiment 2 of this application. Please refer to... Figure 2 The apparatus provided in this embodiment includes a construction module 210, an acquisition module 220, a division module 230, a matching module 240, and an output module 250.

[0096] The construction module 210 is used to construct a historical dataset; the historical dataset includes the geometric parameters, original geological soil layer parameters, and foundation pit deformation index of historical foundation pit projects, and the geometric parameters include the excavation depth;

[0097] The acquisition module 220 is used to acquire the excavation depth and original geological soil layer parameters of the foundation pit to be predicted;

[0098] The division module 230 is used to divide the original geological soil layer structure of the foundation pit to be predicted within the influence depth range into a preset number of equivalent layers according to the excavation depth.

[0099] The construction module 210 is also used to determine the spatial overlap mapping relationship between each equivalent layer and each original geological soil layer in the vertical direction, and to perform spatial weighted average distribution on the corresponding original geological soil layer parameters according to the spatial overlap mapping relationship to obtain the standard layer thickness and standardized physical and mechanical parameters of each equivalent layer, and to combine and construct a standardized feature vector.

[0100] The matching module 240 is used to train multiple machine learning models in parallel based on the historical dataset, and adaptively match the best prediction model for each foundation pit deformation index according to the test performance index.

[0101] The output module 250 is used to input the standardized feature vector of the foundation pit to be predicted into the best prediction model corresponding to each foundation pit deformation index, and simultaneously output the predicted values ​​of multiple foundation pit deformation indices.

[0102] The apparatus of this embodiment can be used to perform... Figure 1 The steps of the method embodiment shown are similar in principle and process, and will not be repeated here.

[0103] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0104] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0105] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A smart prediction method for foundation pit deformation, characterized in that, The method includes: Construct a historical dataset; the historical dataset contains geometric parameters, original geological soil layer parameters, and foundation pit deformation indices of historical foundation pit projects, the geometric parameters including excavation depth; Obtain the excavation depth and original geological soil parameters of the foundation pit to be predicted; Based on the excavation depth, the original geological soil structure of the foundation pit to be predicted within the influence depth range is divided into a preset number of equivalent layers; Determine the spatial overlap mapping relationship between each equivalent layer and each original geological soil layer in the vertical direction. Based on the spatial overlap mapping relationship, perform spatial weighted average distribution on the corresponding original geological soil layer parameters to obtain the standard layer thickness and standardized physical and mechanical parameters of each equivalent layer. Combine these parameters to construct a standardized feature vector. Based on the historical dataset, multiple machine learning models are trained in parallel, and the best prediction model is adaptively matched for each foundation pit deformation index according to the test performance index. The standardized feature vectors of the foundation pit to be predicted are input into the best prediction model corresponding to each foundation pit deformation index, and the predicted values ​​of multiple foundation pit deformation indices are output simultaneously. The standard layer thickness and standardized physical and mechanical parameters of each equivalent layer are obtained, including: For each equivalent layer, the top vertical coordinate and bottom vertical coordinate of the equivalent layer are introduced, the span of the equivalent layer in vertical space is calculated, and the span is used as the corresponding standard layer thickness. The vertical depth range of any target equivalent layer is compared with the natural depth range of each original geological soil layer, and the spatial overlap mapping relationship is divided into inclusion relationship, partial overlap relationship and no overlap relationship. For a target equivalent layer that is determined to have an inclusion relationship or a partial overlap relationship, the actual overlap thickness between the target equivalent layer and the corresponding original geological soil layer is calculated as a spatial weighting weight. The physical and mechanical parameters of each original geological soil layer with overlap are multiplied by the corresponding actual overlap thickness to obtain a parameter thickness product term. All parameter thickness product terms corresponding to the target equivalent layer are summed and then divided by the standard layer thickness to obtain the corresponding standardized physical and mechanical parameters. For the target equivalent layer that is determined to be completely non-overlapping, when the geometric overlap thickness is zero, the segmentation coordinates of the actual original geological soil layers that are spatially adjacent are extracted, and the parameters of the adjacent original geological soil layers are interpolated and reconstructed using the neighbor distance inverse weighted algorithm to obtain the standardized physical and mechanical parameters of the target equivalent layer.

2. The method according to claim 1, characterized in that, The original geological soil layer parameters include thickness, cohesion, internal friction angle and compression modulus, and the foundation pit deformation indicators include horizontal displacement of the retaining structure, surface settlement and horizontal displacement of the pile top.

3. The method according to claim 1, characterized in that, Based on the excavation depth, the original geological soil structure of the foundation pit to be predicted within the influence depth range is divided into a predetermined number of equivalent layers, including: Obtain the total depth of the soil in the foundation pit to be predicted, and determine whether the total depth of the soil is greater than a preset depth threshold. If the total depth of the soil is not greater than the preset depth threshold, then the standard number of layers is used as the preset quantity, and a standard division method based on a fixed proportion of the excavation depth is executed. If the total depth of the soil is greater than the preset depth threshold, the ratio of the total depth of the soil to the excavation depth is calculated. Based on the ratio and the standard number of layers, the layer thickness expansion coefficient and the total number of adaptive equivalent layers are determined. The influence depth range is divided into three longitudinal intervals, wherein the first interval and the second interval are each fixedly divided into two equivalent layers, and the number of layers to be divided in the third interval is determined according to the difference between the total number of layers and the fixed number of layers. Based on the layer thickness expansion coefficient, the standard layer thickness of each equivalent layer is calculated layer by layer for the number of layers to be divided in the third interval. The longitudinal depth boundary value of each equivalent layer is determined by accumulating the standard layer thicknesses sequentially. The influence depth range is divided into an adaptive equivalent layer structure with more than the standard number of layers.

4. The method according to claim 1, characterized in that, The standardized feature vector includes the excavation depth parameter of the foundation pit to be predicted, and the standard layer thickness parameter, cohesion parameter, internal friction angle parameter, and compression modulus parameter corresponding to each equivalent layer according to the physical topological order of vertical space. The standardized feature vector is constructed by taking the excavation depth parameter as the starting dimension and sequentially spatially cascading and splicing the standard layer thickness parameter, cohesion parameter, internal friction angle parameter, and compression modulus parameter corresponding to each equivalent layer to form a standardized feature vector.

5. The method according to claim 1, characterized in that, Based on the historical dataset, multiple machine learning models are trained in parallel. The optimal prediction model is adaptively matched for each foundation pit deformation index according to the test performance metrics, including: The historical dataset is standardized by standard deviation and randomly divided into training and test sets according to a preset division ratio. For different excavation pit deformation indices, various regression algorithm models are trained in parallel using the training set. The regression algorithm models include a linear regression model as the linear baseline, a random forest model using the bagging ensemble method and containing one hundred decision trees, a support vector regression model using a radial basis function kernel, and a Gaussian process regression model using a squared exponential kernel. The test set is input into each regression algorithm model trained for each foundation pit deformation index, and the determination coefficient of each regression algorithm model under the corresponding foundation pit deformation index is calculated. The determination coefficient is used as the test performance index. For any excavation pit deformation index, the magnitude of the coefficient of determination output by each regression algorithm model is compared, and the regression algorithm model with the highest coefficient of determination is determined as the best prediction model.

6. The method according to claim 1, characterized in that, Before performing parallel training of multiple machine learning models based on the historical dataset, the method further includes: Using the excavation depth as the independent variable and the monitoring value of the corresponding foundation pit deformation index as the dependent variable, a univariate linear fit is performed, the fitting residuals corresponding to each sample are calculated, and samples whose absolute value of the fitting residuals exceeds 1.2 times the standard deviation of the residuals are removed. For the retained samples, calculate the lower quartile, upper quartile, and interquartile range of the corresponding foundation pit deformation index monitoring values; The filtered historical dataset is obtained by removing samples whose values ​​are lower than the lower quartile minus 1.5 times the interquartile range, or whose values ​​are higher than the upper quartile plus 1.5 times the interquartile range.

7. The method according to claim 1, characterized in that, The standardized feature vectors of the foundation pits to be predicted are input into the optimal prediction models corresponding to each foundation pit deformation index, and the predicted values ​​of multiple foundation pit deformation indices are output simultaneously, including: Obtain the planned excavation depth values ​​corresponding to the foundation pit to be predicted under multiple consecutive construction steps in the future, and construct a time series sequence of planned excavation depth; For each planned excavation depth value in the planned excavation depth time series, spatial geometric partitioning and parameter space weighted amortization are repeatedly performed to generate standardized feature vectors, which are then combined according to the physical topological order of the continuous construction steps to construct a time series matrix of feature vectors to be predicted. The time series matrix of the feature vector to be predicted is input into the best prediction model corresponding to each foundation pit deformation index for parallel inference, and the predicted value sequence of multiple foundation pit deformation indices under multiple consecutive construction steps in the future is output simultaneously to form a continuous time series prediction change curve that dynamically evolves with the construction steps. Based on the continuous time-series predicted change curve, the maximum predicted value of each foundation pit deformation index is extracted, and it is determined whether the maximum predicted value exceeds the preset safety control threshold of the corresponding foundation pit deformation index; if it exceeds, the corresponding safety warning response is triggered according to the severity of the exceedance.

8. The method according to claim 1, characterized in that, After simultaneously outputting predicted values ​​of multiple foundation pit deformation indices, the method further includes: Real-time acquisition of on-site deformation monitoring data of the foundation pit to be predicted during the excavation stage; The actual deformation monitoring data at the site is compared with the predicted values ​​at the corresponding construction stage in real time to calculate the excavation prediction residual. Determine whether the absolute value of the excavation prediction residual exceeds the set safety threshold. If the absolute value of the excavation prediction residual exceeds the safety threshold, the actual deformation monitoring data on site is used as a boundary feedback constraint and input into the inversion correction algorithm to reverse correct the standardized physical and mechanical parameters of each equivalent layer, or dynamically adjust the internal weight coefficients of the best prediction model. By reconstructing the standardized eigenvectors using the inversely corrected standardized physical and mechanical parameters or internal weighting coefficients and performing predictions, updated predicted values ​​for various foundation pit deformation indices are obtained.

9. An intelligent prediction device for foundation pit deformation, characterized in that, The device includes a construction module, an acquisition module, a division module, a matching module, and an output module; The construction module is used to construct a historical dataset; the historical dataset includes the geometric parameters of historical foundation pit projects, the original geological soil layer parameters, and the foundation pit deformation index, and the geometric parameters include the excavation depth; The acquisition module is used to acquire the excavation depth and original geological soil layer parameters of the foundation pit to be predicted. The division module is used to divide the original geological soil layer structure of the foundation pit to be predicted within the influence depth range into a preset number of equivalent layers according to the excavation depth. The construction module is also used to determine the spatial overlap mapping relationship between each equivalent layer and each original geological soil layer in the vertical direction. Based on the spatial overlap mapping relationship, the corresponding original geological soil layer parameters are spatially weighted and averaged to obtain the standard layer thickness and standardized physical and mechanical parameters of each equivalent layer, and then a standardized feature vector is constructed by combining them. The matching module is used to train multiple machine learning models in parallel based on the historical dataset, and adaptively match the best prediction model for each foundation pit deformation index according to the test performance index. The output module is used to input the standardized feature vector of the foundation pit to be predicted into the best prediction model corresponding to each foundation pit deformation index, and simultaneously output the predicted values ​​of multiple foundation pit deformation indices. The standard layer thickness and standardized physical and mechanical parameters of each equivalent layer are obtained, including: For each equivalent layer, the top vertical coordinate and bottom vertical coordinate of the equivalent layer are introduced, the span of the equivalent layer in vertical space is calculated, and the span is used as the corresponding standard layer thickness. The vertical depth range of any target equivalent layer is compared with the natural depth range of each original geological soil layer, and the spatial overlap mapping relationship is divided into inclusion relationship, partial overlap relationship and no overlap relationship. For a target equivalent layer that is determined to have an inclusion relationship or a partial overlap relationship, the actual overlap thickness between the target equivalent layer and the corresponding original geological soil layer is calculated as a spatial weighting weight. The physical and mechanical parameters of each original geological soil layer with overlap are multiplied by the corresponding actual overlap thickness to obtain a parameter thickness product term. All parameter thickness product terms corresponding to the target equivalent layer are summed and then divided by the standard layer thickness to obtain the corresponding standardized physical and mechanical parameters. For the target equivalent layer that is determined to be completely non-overlapping, when the geometric overlap thickness is zero, the segmentation coordinates of the actual original geological soil layers that are spatially adjacent are extracted, and the parameters of the adjacent original geological soil layers are interpolated and reconstructed using the neighbor distance inverse weighted algorithm to obtain the standardized physical and mechanical parameters of the target equivalent layer.

Images