A method and system for capturing and predicting micro-deformation of deep foundation pit support pile by fusing physical mechanism constraints
By constructing a micro-deformation prediction method for deep foundation pit support piles that integrates physical mechanisms, the problems of micro-feature annihilation and lack of physical consistency in existing technologies are solved, achieving high-precision micro-deformation capture and prediction, and providing reliable safety early warning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHENGDU GUANGSHU INVESTIGATION BASIC CO
- Filing Date
- 2026-02-12
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies for predicting deformation of support piles in deep foundation pits suffer from problems such as micro-feature annihilation, lack of physical consistency, and blind attention, resulting in inaccurate prediction results and a lack of interpretability, making it difficult to meet the requirements for high-precision and high-reliability safety early warning.
A method for capturing and predicting micro-deformation of deep foundation pit support piles that integrates physical mechanism constraints is constructed. By acquiring and fusing static physical fields and dynamic displacement fields, an enhanced spatiotemporal field map is constructed. The model is trained using void convolution and residual connection units combined with physical information loss function to extract the micro-strain characteristics of the support piles. A mechanism mask attention mechanism is introduced for feature focusing.
It achieves high-fidelity capture of micron-level deformation, and the prediction results are consistent with the common sense of mechanics, which improves the accuracy and robustness of early warning, enables timely identification of high-risk areas, and provides reliable basis for safety early warning.
Smart Images
Figure CN121706609B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent cross-disciplinary technology of precision monitoring and calculation in geotechnical engineering, specifically involving a method and system for capturing and predicting micro-deformation of deep foundation pit support piles that integrates physical mechanism constraints. Background Technology
[0002] In recent years, with the increasing intensity of underground space development, safety monitoring of deep foundation pit support engineering has become a key focus of urban disaster prevention and mitigation. As the core component of the foundation pit retaining system, the deformation behavior of support piles exhibits significant nonlinear spatiotemporal coupling characteristics. Especially in the precursory stage of structural failure, it often manifests as cumulative micro-deformation at the micrometer to millimeter level. The ability to timely capture and accurately predict these extremely weak "quantitative change signals" is crucial for preventing the "qualitative disaster" of foundation pit collapse.
[0003] Currently, deformation prediction of retaining piles mainly relies on time series analysis methods or general deep learning models. Although these data-driven methods improve prediction efficiency to some extent, when directly applied to the specific scenario of precision monitoring in geotechnical and foundation pit engineering, three unavoidable structural technical defects exist: 1) The "micro-feature annihilation" problem caused by the general downsampling architecture: Existing convolutional neural networks generally use pooling layers or large stride convolutions for downsampling dimensionality reduction to reduce computational load and extract high-level semantic features. This mechanism is effective in the field of computer vision, but it is fatal in high-precision displacement monitoring. Pooling operation is essentially a lossy compression, which smooths out high-frequency micro-deformation signals at locations such as deep embedded sections as "noise." This makes the model insensitive to micron-level displacements in key hidden areas such as depths of -25m, which can easily lead to missed detections of safety hazards. 2) The problem of "lack of physical consistency" caused by pure data-driven models: Existing algorithms such as LSTM or Transformer are essentially probabilistic statistical models. Their training process relies entirely on fitting historical data, ignoring the principles of continuum mechanics that support structures must strictly follow (such as the fourth-order differential equation of beam elements and the pile-soil constitutive relationship). Under small sample sizes or noise interference, these models are prone to outputting predictions that violate physical common sense (such as the predicted pile deflection curve oscillating violently in the absence of external force, or failing to meet the boundary conditions for bottom embedment). This black-box model lacks physical interpretability, making it impossible for engineers to determine whether the prediction results represent a real danger or an algorithm error. 3) The problem of "blind attention" due to lack of prior knowledge: The failure mode of deep foundation pit support piles has a clear location sensitivity (mainly controlled by the maximum bending moment point and weak interlayers). However, general self-attention mechanisms lack this geological prior knowledge. They require a large amount of data to "blindly explore" feature weights, making it difficult to quickly focus on the "most dangerous section" of the support pile with limited engineering samples. This makes the model less robust when faced with geological upheavals (such as weak lenses) as experienced engineers can apply targeted risk weighting. This indicates that there is an urgent need for a dedicated prediction method that can preserve full-resolution signals of minute deformations while incorporating constraints from geotechnical mechanics mechanisms to meet the stringent requirements of high-precision, high-reliability safety early warning in deep foundation pit engineering. Summary of the Invention
[0004] This invention aims to address the shortcomings of existing technologies and provides the following solutions:
[0005] A method for capturing and predicting micro-deformation of deep foundation pit support piles that integrates physical mechanism constraints includes the following steps:
[0006] Acquire and fuse the static physical field and dynamic displacement field of deep foundation pit support piles to construct an enhanced spatiotemporal field map containing physical boundary conditions;
[0007] A micro-deformation prediction model is constructed, and a physical information loss function is constructed based on the fourth-order differential equation of the support pile beam element. The micro-deformation prediction model is then trained based on the physical information loss function to obtain the trained model.
[0008] The enhanced spatiotemporal field map is input into the trained model to extract the spatial morphological features of the micro-strain of the pile body in the enhanced spatiotemporal field map and obtain the deformation prediction results of the support pile.
[0009] Preferably, the method for obtaining the enhanced spatiotemporal field map includes:
[0010] Based on the geometric dimensions of the support piles, they are discretized into N spatial nodes along the depth direction, and the dynamic displacement field of each spatial node during the monitoring period is obtained.
[0011] Based on geological survey data, a heterogeneous rock and soil physical property matrix is constructed that maps one-to-one with the spatial nodes to obtain the static physical field;
[0012] Perform spatiotemporal alignment and channel expansion operations to broadcast and copy the static physical field along the time dimension, and stitch it together with the dynamic displacement field in the depth dimension to obtain the enhanced spatiotemporal field map.
[0013] Preferably, the dynamic displacement field is:
[0014] ;
[0015] Among them, X dis Represents the dynamic displacement field, where T represents time, N represents the node, and u T,N R represents the horizontal displacement monitoring value at time T and depth node N, where R represents the set of real numbers.
[0016] The static physical field is:
[0017] ;
[0018] Among them, X stat Represents a static physical field, e N c represents the soil void ratio at node N. N k represents the cohesion at node N. N The value represents the horizontal resistance coefficient of the soil at node N, and M represents the number of physical and mechanical parameters.
[0019] Preferably, the enhanced spatiotemporal field map is:
[0020] ;
[0021] Among them, X inputThis represents an enhanced spatiotemporal field map, Concat represents channel splicing operation, Reshape represents tensor dimension transformation, and Tile represents copy expansion.
[0022] Preferably, the micro-deformation prediction model includes: a first dilated convolutional unit, a second dilated convolutional unit, a third dilated convolutional unit, a feature fusion unit, and a residual connection unit;
[0023] The first dilated convolutional unit uses a 3×3 convolutional kernel with an expansion rate of 1 and a stride of 1 to extract the local micro-strain abrupt change features between adjacent depth nodes in the enhanced spatiotemporal field map.
[0024] The second dilated convolutional unit is set in parallel with the first dilated convolutional unit, using a 3×3 convolutional kernel, an expansion rate of 2, and a stride of 1, to capture the stress transfer and transition characteristics in the cross-node region of the enhanced spatiotemporal field map;
[0025] The third dilated convolutional unit is set in parallel with the first dilated convolutional unit and the second dilated convolutional unit. It uses a 3×3 convolutional kernel, an expansion rate of 3, and a stride of 1. It is used to expand the receptive field without reducing the resolution and extract the overall elastic flexural modal features of the support pile.
[0026] The feature fusion unit is connected to the outputs of the first dilated convolution unit, the second dilated convolution unit, and the third dilated convolution unit. It is used to concatenate the features output by the three dilated convolution units and perform cross-channel weighted fusion of features extracted with different dilation rates.
[0027] One end of the residual connection unit is connected to the model input, and the other end is connected to the output of the feature fusion unit. It is used to fuse the enhanced spatiotemporal field map with the weighted fusion features to obtain the spatial morphological features.
[0028] The present invention also provides a system for capturing and predicting micro-deformation of deep foundation pit support piles that integrates physical mechanism constraints. The system applies the above-mentioned method and includes: a map construction module, a model construction module, and a prediction module.
[0029] The map construction module is used to acquire and fuse the static physical field and dynamic displacement field of deep foundation pit support piles to construct an enhanced spatiotemporal field map containing physical boundary conditions.
[0030] The model building module is used to build a micro-deformation prediction model, and to build a physical information loss function based on the fourth-order differential equation of the support pile beam element. The micro-deformation prediction model is then trained based on the physical information loss function to obtain the trained model.
[0031] The prediction module is used to input the enhanced spatiotemporal field map into the trained model, extract the spatial morphological features of the micro-strain of the pile body in the enhanced spatiotemporal field map, and obtain the deformation prediction results of the support pile.
[0032] Preferably, the method for obtaining the enhanced spatiotemporal field map includes:
[0033] Based on the geometric dimensions of the support piles, they are discretized into N spatial nodes along the depth direction, and the dynamic displacement field of each spatial node during the monitoring period is obtained.
[0034] Based on geological survey data, a heterogeneous rock and soil physical property matrix is constructed that maps one-to-one with the spatial nodes to obtain the static physical field;
[0035] Perform spatiotemporal alignment and channel expansion operations to broadcast and copy the static physical field along the time dimension, and stitch it together with the dynamic displacement field in the depth dimension to obtain the enhanced spatiotemporal field map.
[0036] Preferably, the dynamic displacement field is:
[0037] ;
[0038] Among them, X dis Represents the dynamic displacement field, where T represents time, N represents the node, and u T,N R represents the horizontal displacement monitoring value at time T and depth node N, where R represents the set of real numbers.
[0039] The static physical field is:
[0040] ;
[0041] Among them, X stat Represents a static physical field, e N c represents the soil void ratio at node N. N k represents the cohesion at node N. N The value represents the horizontal resistance coefficient of the soil at node N, and M represents the number of physical and mechanical parameters.
[0042] Preferably, the enhanced spatiotemporal field map is:
[0043] ;
[0044] Among them, X input This represents an enhanced spatiotemporal field map, Concat represents channel splicing operation, Reshape represents tensor dimension transformation, and Tile represents copy expansion.
[0045] Preferably, the micro-deformation prediction model includes: a first dilated convolutional unit, a second dilated convolutional unit, a third dilated convolutional unit, a feature fusion unit, and a residual connection unit;
[0046] The first dilated convolutional unit uses a 3×3 convolutional kernel with an expansion rate of 1 and a stride of 1 to extract the local micro-strain abrupt change features between adjacent depth nodes in the enhanced spatiotemporal field map.
[0047] The second dilated convolutional unit is set in parallel with the first dilated convolutional unit, using a 3×3 convolutional kernel, an expansion rate of 2, and a stride of 1, to capture the stress transfer and transition characteristics in the cross-node region of the enhanced spatiotemporal field map;
[0048] The third dilated convolutional unit is set in parallel with the first dilated convolutional unit and the second dilated convolutional unit. It uses a 3×3 convolutional kernel, an expansion rate of 3, and a stride of 1. It is used to expand the receptive field without reducing the resolution and extract the overall elastic flexural modal features of the support pile.
[0049] The feature fusion unit is connected to the outputs of the first dilated convolution unit, the second dilated convolution unit, and the third dilated convolution unit. It is used to concatenate the features output by the three dilated convolution units and perform cross-channel weighted fusion of features extracted with different dilation rates.
[0050] One end of the residual connection unit is connected to the model input, and the other end is connected to the output of the feature fusion unit. It is used to fuse the enhanced spatiotemporal field map with the weighted fusion features to obtain the spatial morphological features.
[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0052] This invention overcomes the technical bottleneck of "micro-feature annihilation" in general models, achieving high-fidelity capture of micron-level deformation. Addressing the deficiency of existing convolutional networks in losing minute signals due to pooling dimensionality reduction, this invention innovatively constructs a full-resolution multi-scale void residual architecture that eliminates pooling layers. While maintaining the feature map spatial size, this design, through parallel convolutional kernels with different dilation rates, preserves micron-level (0.1mm-level) initial displacement signals in hidden regions such as deeply embedded sections, while also capturing the overall deflection morphology of the pile body. This results in extremely high sensitivity of the system to safety hazards during the "quantitative change" stage of deep foundation pits. It endows deep learning models with rigorous "physical interpretability," eliminating non-physical oscillation errors. Unlike the "black box" prediction results output by purely data-driven models (such as LSTM), this invention establishes a deep fusion mechanism between data-driven approaches and mechanical mechanisms. By introducing a physical loss function based on the fourth-order differential equation of beam elements, this invention forces the model's prediction trajectory to strictly follow the principle of continuous medium mechanical equilibrium, effectively eliminating the sawtooth non-physical oscillations common in traditional methods. This allows the predicted curvature distribution to accurately reflect the abrupt stiffness changes at weak interlayers, providing engineers with a reliable decision-making basis that conforms to mechanical common sense. It achieves "high-value area focusing" guided by engineering mechanisms, improving robustness under complex geological conditions. Abandoning the inefficient "blind search" weighting mode of general attention mechanisms, this invention explicitly injects prior geotechnical engineering knowledge (such as the depth of the maximum bending moment point and the location of weak soil layers) into the model through mechanism masking. This hybrid driving mode of "expert system + neural network" guides the model to automatically focus on the spatiotemporal evolution of the highest-risk section, not only accelerating the model's convergence speed with small sample data but also significantly improving its anti-interference ability and early warning accuracy under heterogeneous and complex geological conditions. Attached Figure Description
[0053] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0054] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;
[0055] Figure 2 The predicted curvature curve of the present invention at a depth of -12m (weak interlayer) is shown in this embodiment of the present invention. Detailed Implementation
[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0057] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0058] Specifically, to verify the superiority of this invention in capturing minute deformations, this embodiment selects a typical deep foundation pit project as the verification object. The excavation depth of this foundation pit reaches 26m, which falls into the category of ultra-deep foundation pits, with a safety level of Level 1. The support system adopts a pile + internal bracing structure, with support piles 35m long, 1.2m in diameter, and 2.0m apart. The engineering geological conditions have a typical "soft on top, hard on the bottom" binary structure: the strata from top to bottom are mainly miscellaneous fill, silty clay (weak layer), sand and gravel layer, and strongly weathered mudstone (impregnated layer). This complex combination of strata leads to abrupt changes in the stress mode of the support piles at different depths, and micron-level elastic deformation, which is extremely difficult to capture, often exists in the impregnated section below 20m depth. This embodiment collects and constructs a dataset containing the complete construction cycle by deploying high-precision automated monitoring elements (including fixed inclinometers, rebar gauges, etc.), thereby training and verifying the "physical information enhancement + full-resolution network" model proposed in this invention.
[0059] Example 1:
[0060] In this embodiment, as Figure 1 As shown, a method for capturing and predicting micro-deformation of deep foundation pit support piles that integrates physical mechanism constraints includes the following steps:
[0061] S1. Acquire and fuse the static physical field and dynamic displacement field of the deep foundation pit support piles to construct an enhanced spatiotemporal field map containing physical boundary conditions.
[0062] The method for obtaining the enhanced spatiotemporal field map includes: discretizing the support pile into N spatial nodes along the depth direction based on the geometric dimensions of the support pile, and obtaining the dynamic displacement field of each spatial node during the monitoring period; constructing a heterogeneous soil and rock physical property matrix that maps one-to-one with the spatial nodes based on geological survey data to obtain the static physical field; performing spatiotemporal dimension alignment and channel expansion operations to broadcast and copy the static physical field along the time dimension, and stitching it with the dynamic displacement field through the depth dimension to obtain the enhanced spatiotemporal field map.
[0063] In this embodiment, the support piles are discretized into N spatial nodes along the depth direction based on their geometric dimensions, and the dynamic displacement field of each spatial node during the monitoring period is obtained. Specifically, a fixed inclinometer embedded inside the pile is used, with a monitoring node set at every 0.5m, for a total of 70 nodes. The data is collected every 2 hours, continuously collecting horizontal displacement data throughout the entire excavation process (a total of 120 days), and constructing a dynamic displacement field X. dis As shown in the following formula:
[0064] ;
[0065] Where T represents time, N represents node, and u T,N R represents the horizontal displacement monitoring value at time T and depth node N, and R represents the set of real numbers.
[0066] Based on geological survey data, a heterogeneous soil and rock physical property matrix is constructed with a one-to-one mapping to spatial nodes to obtain a static physical field. The static physical field includes at least the void ratio characterizing soil compressibility, the cohesion characterizing shear strength, and the horizontal resistance coefficient characterizing soil-pile interaction stiffness. Specifically, a static physical field representation X corresponding to each spatial node is constructed. stat Select M physical and mechanical parameters that significantly affect micro-deformation (e.g., when M=3, these are soil void ratio e, cohesion c, and soil horizontal resistance coefficient k), then the static physical field X stat As shown in the following formula:
[0067] ;
[0068] Among them, e N c represents the soil void ratio at node N. N k represents the cohesion at node N. N The value represents the horizontal resistance coefficient of the soil at node N, and M represents the number of physical and mechanical parameters.
[0069] The spatiotemporal alignment and channel expansion operations are performed, broadcasting and copying the static physical property matrix along the time dimension, and then concatenating it with the dynamic displacement field data in the depth dimension to generate a high-dimensional multi-channel physical input tensor that combines time-varying deformation characteristics with time-invariant material constitutive characteristics, resulting in an enhanced spatiotemporal field map. Specifically, based on the geological survey report, the geological physical and mechanical parameters are mapped to the corresponding pile depth nodes. The displacement sequence X is then... dis ∈R T×70 With physical parameter matrix X stat ∈R 1×70×30 (Including stiffness, void ratio, and cohesion) Tensor combination is performed to generate a physically enhanced spacetime map X. input The specific calculation formula is as follows:
[0070] ;
[0071] Among them, X input This represents an enhanced spatiotemporal field map, where Concat represents the channel stitching operation, Reshape represents the tensor dimension transformation, and Tile represents copy expansion. During this process, the static physical field X... stat The data will be copied and expanded along the time dimension (Tile) to align with the spatiotemporal dimension of the displacement data, thereby forming a multidimensional feature tensor containing one displacement channel and M physical mechanism channels. It should be noted that the "map" mentioned in this embodiment refers to a multidimensional feature tensor that integrates spatial, temporal, and physical attribute dimensions. In terms of data structure, it is represented in a matrix form similar to a multi-channel image to facilitate processing by convolutional neural networks, rather than referring to a topological graph structure.
[0072] S2. Construct a micro-deformation prediction model, and build a physical information loss function based on the fourth-order differential equation of the support pile beam element. Then, train the micro-deformation prediction model based on the physical information loss function to obtain the trained model.
[0073] The micro-deformation prediction model includes: a first dilated convolutional unit, a second dilated convolutional unit, a third dilated convolutional unit, a feature fusion unit, and a residual connection unit.
[0074] First dilated convolutional unit (high-frequency feature branch): This unit is configured to receive the physically enhanced spatiotemporal map features, using a 3×3 convolution kernel with an inflation rate of r1=1 and a stride of s=1. This unit maintains the feature map spatial size unchanged through zero padding (Padding=1) and is configured to extract local micro-strain abrupt change features between adjacent depth nodes in the enhanced spatiotemporal field map.
[0075] The second dilated convolutional unit (mid-frequency transition branch) is set in parallel with the first dilated convolutional unit, using a 3×3 convolutional kernel, an expansion rate of r2=2, and a stride of s=1. This unit maintains the feature map spatial size unchanged through zero padding (Padding=2) and is used to capture stress transfer and transition features across node regions in the enhanced spatiotemporal field map.
[0076] The third dilated convolutional unit (low-frequency flexure branch): is set in parallel with the first and second dilated convolutional units, using a 3×3 convolutional kernel, an expansion rate of r3=3, and a stride of s=1. This unit maintains the feature map spatial size unchanged through zero padding (Padding=4), which is used to expand the receptive field without reducing the resolution and extract the overall elastic flexure modal features of the support pile.
[0077] Defined as a dilated convolution operator with an inflation rate of r, the output of the dilated convolution at any spatial location on the input feature map can be expressed as follows:
[0078] ;
[0079] Among them, W p represents the kernel weights, K represents the kernel size, and p represents the selection of each weight parameter in the kernel.
[0080] The feature fusion unit connects to the outputs of the first, second, and third dilated convolutional units. First, it performs a channel concatenation operation, concatenating the features output by the three dilated convolutional units to generate a hybrid feature tensor with three times the number of channels as the original features. Then, it connects to a 1×1 point convolutional layer to perform cross-channel weighted fusion of features extracted at different dilation rates and reduce the number of channels to match the input features.
[0081] One end of the residual connection unit is connected to the model input, and the other end is connected to the output of the feature fusion unit. It performs element-wise addition to fuse the enhanced spatiotemporal field map with the weighted fused features, obtaining spatial morphological features. This structure is configured to construct an identity mapping channel, ensuring that the original small displacement signal can be directly transmitted to deeper layers of the network, avoiding gradient vanishing or signal loss due to increased network depth. After feature concatenation, the outputs of each branch are added to the original input through residual connections, ensuring that the initial deformation signal at the micrometer level does not vanish due to increased network depth. The final spatial feature tensor H is calculated using the following formula:
[0082] ;
[0083] Where ReLU represents the linear rectified activation function, D r1 D represents the first dilated convolutional unit. r2 D represents the second dilated convolutional unit. r3 denoted as the third dilated convolution unit, dilated convolution is used to expand the receptive field without reducing resolution. The smaller dilation rate is used to capture local stress abrupt changes between adjacent nodes, and the larger dilation rate is used to capture the overall flexural deformation mode of the support pile. During the operation, the spatial resolution of the feature map is kept consistent with the input. Res represents the residual connection term used to retain the original small deformation signal. It is configured to build a direct transmission channel for the original signal and prevent the gradient vanishing or amplitude decay of the micron-level initial displacement features in the deep network operation. H is the output feature tensor that integrates multi-scale spatial information.
[0084] A mechanistic masking attention mechanism based on prior knowledge of geological stratification and stress analysis is introduced. This mechanism uses a pre-set physical mask matrix to explicitly weight and focus on key stress-bearing strata, such as the point of maximum bending moment and weak interlayers. In this embodiment, the calculation formula is:
[0085] ;
[0086] Where Q, K, and V are the query vector, key vector, and value vector, respectively, and d k Scaling factor For Hadamard; M p This is a physical mechanism mask matrix constructed based on prior knowledge of geotechnical engineering. The construction of the mask matrix is based on the estimated depth of the maximum bending moment of the support piles obtained through mechanical analysis and the depth of the interface between the weak soil layer and bedrock determined by geological survey data. In the mask matrix, the spatial coordinates corresponding to the aforementioned high-risk critical layers are assigned high-weight coefficient values, while other positions are assigned basic weight values. The model is explicitly guided to focus on the spatiotemporal evolution characteristics of potential hazard areas through Hadamard product calculation. The specific process of focusing on critical layers based on the mechanism mask attention mechanism includes: based on pile foundation stress analysis, the maximum bending moment point is predicted to be located 2-4m below the excavation surface (i.e., in the depth range of 28m-30m), and the weak interlayer is located in the silty clay layer at an excavation depth of 12m. The physical mechanism mask matrix M is constructed. p In this embodiment, the mask matrix is a static matrix preset based on prior geological survey data. However, in other engineering implementations, this mask matrix can also be dynamically updated according to drilling data or excavation progress to adapt to changes in working conditions. That is, the mask value is set to 1.5 (high-weight region) at depths of 28m-30m and 12m; the mask value is set to 0.8 (low-weight region) at other depths. The weighted feature vector is calculated using the Attention formula, forcing the model to focus on the evolutionary patterns of the aforementioned high-risk regions.
[0087] A physical information loss function is constructed based on the fourth-order differential equation of the support pile-beam element. The physical relationship between bending stiffness and soil reaction is used as a regularization term to constrain the training process of the micro-deformation prediction model. Specifically, a total loss function containing physical constraints is constructed. During training, the Adam optimizer is used with a learning rate of 0.001, and 500 iterations are performed. By constraining the physical loss term, the curvature distribution of the predicted displacement curve y(z) after the second derivative is forced to conform to the principles of continuum mechanics, eliminating the "sawtooth" non-physical oscillations commonly found in traditional pure data models.
[0088] The total loss function is:
[0089] ;
[0090] ;
[0091] Among them, L t L represents the total loss function; MSE L represents the data-driven mean squared error loss, used to fit measured monitoring data; pRepresents the physical information loss function; λ data and λ p These are the weights for the data items and the physical items, respectively, and are dynamically adjusted with the number of training iterations. They are configured as adaptive balancing coefficients, which dynamically adjust with the number of training iterations to balance the difference in gradient contributions between data fitting and physical constraints, achieving synergistic optimization of data accuracy and physical consistency. `i` represents the index variable in the summation formula, `y(z,t)` is the horizontal displacement of the support pile predicted by the model at depth `z` and time `t`, `EI(z)` is the bending stiffness of the support pile, `k(z)` is the soil horizontal resistance coefficient varying with depth, and `q(z,t)` is the soil pressure load behind the pile, where `z` is the depth. The loss function minimizes the physical residual, eliminating non-physical high-frequency oscillations in the prediction curve, and forcing the model output to satisfy the continuum mechanical constraints.
[0092] S3. Input the enhanced spatiotemporal field map into the trained model, extract the spatial morphological features of the micro-strain of the pile body in the enhanced spatiotemporal field map, and obtain the deformation prediction results of the support pile.
[0093] Based on the analysis results, the curvature distribution q(z) at each depth of the support pile is calculated; a critical curvature threshold q is set. limit =M yield / EI, where M yield The yield bending moment of the retaining pile material; when the predicted curvature distribution q(z) > k limit At that time, a graded safety warning signal is triggered.
[0094] To verify the advanced nature of this embodiment, a key working condition of excavating the foundation pit to a depth of 20m was selected. The micro-deformation of the pile body in the next 24 hours was predicted using the model of this invention (PINN-Enhanced), the standard LSTM model, and the traditional SVM model. Specifically, error analysis was performed at three key measuring points at different depths (-5m, -15m, and -25m). The comparison between the predicted results and the measured values is shown in Table 1.
[0095] Table 1:
[0096] ;
[0097] As clearly observed in Table 1, the prediction accuracy of the model of this invention is significantly better than that of the comparative model at all depths. Particularly in the embedded section at a depth of -25m, the measured micro-deformation is only 0.42mm. Due to the extremely weak signal, the standard LSTM and SVM models misclassified it as noise or zero, predicting values of 0.15mm and 0.08mm respectively, with errors exceeding 60%, exhibiting a severe "micro-feature loss" phenomenon. In contrast, this invention, employing a multi-scale hollow residual network, successfully captured this micrometer-level deformation signal, predicting a value of 0.41mm with an error of only 0.01mm. This data strongly demonstrates the superior performance of this invention in capturing minute deformations.
[0098] Furthermore, to verify the effectiveness of the physical mechanism constraint, the predicted curvature curve at a depth of -12m (weak interlayer) was extracted, as shown below. Figure 2 As shown, the displacement curve predicted by the standard model exhibits abrupt discontinuities at the weak interlayer, and the derived curvature curve displays severe non-physical oscillations, which does not conform to the assumption of continuous deformation of the beam. The curve predicted by this invention is smooth and continuous, and accurately reflects the increase in curvature caused by the sudden drop in stiffness of the weak soil layer.
[0099] This embodiment, by introducing physical information constraints and micro-feature enhancement mechanisms, not only reduces the average error of micro-deformation prediction by an order of magnitude (from 0.283mm to 0.027mm), but also ensures the physical interpretability of the prediction results, providing a more reliable basis for safety early warning in deep foundation pit engineering.
[0100] Example 2:
[0101] In this embodiment, a deep foundation pit support pile micro-deformation capture and prediction system integrating physical mechanism constraints includes: a map construction module, a model construction module, and a prediction module.
[0102] The map construction module is used to acquire and fuse the static physical field and dynamic displacement field of deep foundation pit support piles to construct an enhanced spatiotemporal field map that includes physical boundary conditions.
[0103] The method for obtaining the enhanced spatiotemporal field map includes: discretizing the support piles along the depth direction into N spatial nodes based on their geometric dimensions, and obtaining the dynamic displacement field of each spatial node during the monitoring period; the dynamic displacement field is:
[0104] ;
[0105] Among them, X dis Represents the dynamic displacement field, where T represents time, N represents the node, and u T,NLet R represent the horizontal displacement monitoring value at time T and depth node N, and R represent the set of real numbers. Based on geological survey data, a heterogeneous soil and rock physical property matrix with a one-to-one mapping to spatial nodes is constructed to obtain the static physical field; the static physical field is:
[0106] ;
[0107] Among them, X stat Represents a static physical field, e N c represents the soil void ratio at node N. N k represents the cohesion at node N. N represents the horizontal resistance coefficient of the soil at node N, and M represents the number of physical and mechanical parameters. Spatiotemporal alignment and channel expansion operations are performed to broadcast and copy the static physical field along the time dimension, and then stitch it together with the dynamic displacement field in the depth dimension to obtain an enhanced spatiotemporal field map; the enhanced spatiotemporal field map is as follows:
[0108] ;
[0109] Among them, X input This represents an enhanced spatiotemporal field map, Concat represents channel splicing operation, Reshape represents tensor dimension transformation, and Tile represents copy expansion.
[0110] The model building module is used to construct a micro-deformation prediction model, and to construct a physical information loss function based on the fourth-order differential equation of the support pile beam element. The micro-deformation prediction model is then trained based on the physical information loss function to obtain the trained model.
[0111] The micro-deformation prediction model includes: a first dilated convolutional unit, a second dilated convolutional unit, a third dilated convolutional unit, a feature fusion unit, and a residual connection unit. The first dilated convolutional unit uses a 3×3 convolutional kernel with a dilation rate of 1 and a stride of 1 to extract local micro-strain abrupt changes between adjacent depth nodes in the enhanced spatiotemporal field map. The second dilated convolutional unit is set in parallel with the first dilated convolutional unit, using a 3×3 convolutional kernel with a dilation rate of 2 and a stride of 1, to capture stress transfer and transition features across node regions in the enhanced spatiotemporal field map. The third dilated convolutional unit is set in parallel with the first and second dilated convolutional units, using a 3×3 convolutional kernel with a dilation rate of 3 and a stride of 1, to extract local micro-strain abrupt changes between adjacent depth nodes in the enhanced spatiotemporal field map. To expand the receptive field without reducing resolution, the elastic flexural modal features of the entire support pile are extracted. The feature fusion unit is connected to the outputs of the first, second, and third dilated convolutional units, and is used to concatenate the features output by the three dilated convolutional units and perform cross-channel weighted fusion of features extracted at different expansion rates. One end of the residual connection unit is connected to the model input, and the other end is connected to the output of the feature fusion unit, and is used to fuse the enhanced spatiotemporal field map with the weighted fused features to obtain spatial morphological features.
[0112] The prediction module is used to input the enhanced spatiotemporal field map into the trained model, extract the spatial morphological features of the micro-strain of the pile body in the enhanced spatiotemporal field map, and obtain the deformation prediction results of the support pile.
[0113] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for capturing and predicting micro-deformation of deep foundation pit support piles that integrates physical mechanism constraints, characterized in that, Includes the following steps: Acquire and fuse the static physical field and dynamic displacement field of deep foundation pit support piles to construct an enhanced spatiotemporal field map containing physical boundary conditions; A micro-deformation prediction model is constructed, and a physical information loss function is constructed based on the fourth-order differential equation of the support pile beam element. The micro-deformation prediction model is then trained based on the physical information loss function to obtain the trained model. The enhanced spatiotemporal field map is input into the trained model to extract the spatial morphological features of the micro-strain of the pile body in the enhanced spatiotemporal field map and obtain the deformation prediction results of the support pile. The method for obtaining the enhanced spatiotemporal field map includes: Based on the geometric dimensions of the support piles, they are discretized into N spatial nodes along the depth direction, and the dynamic displacement field of each spatial node during the monitoring period is obtained. Based on geological survey data, a heterogeneous rock and soil physical property matrix is constructed that maps one-to-one with the spatial nodes to obtain the static physical field; Perform spatiotemporal alignment and channel expansion operations to broadcast and copy the static physical field along the time dimension, and stitch it with the dynamic displacement field in the depth dimension to obtain the enhanced spatiotemporal field map; The micro-deformation prediction model includes: a first dilated convolution unit, a second dilated convolution unit, a third dilated convolution unit, a feature fusion unit, and a residual connection unit; The first dilated convolutional unit uses a 3×3 convolutional kernel with an expansion rate of 1 and a stride of 1 to extract the local micro-strain abrupt change features between adjacent depth nodes in the enhanced spatiotemporal field map. The second dilated convolutional unit is set in parallel with the first dilated convolutional unit, using a 3×3 convolutional kernel, an expansion rate of 2, and a stride of 1, to capture the stress transfer and transition characteristics in the cross-node region of the enhanced spatiotemporal field map; The third dilated convolutional unit is set in parallel with the first dilated convolutional unit and the second dilated convolutional unit. It uses a 3×3 convolutional kernel, an expansion rate of 3, and a stride of 1. It is used to expand the receptive field without reducing the resolution and extract the overall elastic flexural modal features of the support pile. The feature fusion unit is connected to the outputs of the first dilated convolution unit, the second dilated convolution unit, and the third dilated convolution unit. It is used to concatenate the features output by the three dilated convolution units and perform cross-channel weighted fusion of features extracted with different dilation rates. One end of the residual connection unit is connected to the model input, and the other end is connected to the output of the feature fusion unit. It is used to fuse the enhanced spatiotemporal field map with the weighted fusion features to obtain the spatial morphological features.
2. The method for capturing and predicting micro-deformation of deep foundation pit support piles based on physical mechanism constraints as described in claim 1, is characterized in that, The dynamic displacement field is: ； Where Xdis represents the dynamic displacement field. T Indicates time, N Represents a node. uT , N Indicates at time T Deep nodes N Horizontal displacement monitoring value at the location, R Represents the set of real numbers; The static physical field is: ； Where Xstat represents the static physical field, eN Represents a node N The soil porosity at that location cN Represents a node N Cohesion at the point, kN Represents a node N The horizontal resistance coefficient of the soil at that location, M It represents the number of physical and mechanical parameters.
3. The method for capturing and predicting micro-deformation of deep foundation pit support piles based on physical mechanism constraints as described in claim 2, is characterized in that... The enhanced spatiotemporal field map is as follows: ； Among them, Xinput represents the enhanced spatiotemporal field map, Concat represents the channel splicing operation, Reshape represents the tensor dimension transformation, and Tile represents copy expansion.
4. A system for capturing and predicting micro-deformation of deep foundation pit support piles that integrates physical mechanism constraints, wherein the system applies the method described in any one of claims 1-3, characterized in that, include: The system comprises a graph construction module, a model construction module, and a prediction module. The map construction module is used to acquire and fuse the static physical field and dynamic displacement field of deep foundation pit support piles to construct an enhanced spatiotemporal field map containing physical boundary conditions. The model building module is used to build a micro-deformation prediction model, and to build a physical information loss function based on the fourth-order differential equation of the support pile beam element. The micro-deformation prediction model is then trained based on the physical information loss function to obtain the trained model. The prediction module is used to input the enhanced spatiotemporal field map into the trained model, extract the spatial morphological features of the micro-strain of the pile body in the enhanced spatiotemporal field map, and obtain the deformation prediction results of the support pile. The method for obtaining the enhanced spatiotemporal field map includes: Based on the geometric dimensions of the support piles, they are discretized into N spatial nodes along the depth direction, and the dynamic displacement field of each spatial node during the monitoring period is obtained. Based on geological survey data, a heterogeneous rock and soil physical property matrix is constructed that maps one-to-one with the spatial nodes to obtain the static physical field; Perform spatiotemporal alignment and channel expansion operations to broadcast and copy the static physical field along the time dimension, and stitch it with the dynamic displacement field in the depth dimension to obtain the enhanced spatiotemporal field map; The micro-deformation prediction model includes: a first dilated convolution unit, a second dilated convolution unit, a third dilated convolution unit, a feature fusion unit, and a residual connection unit; The first dilated convolutional unit uses a 3×3 convolutional kernel with an expansion rate of 1 and a stride of 1 to extract the local micro-strain abrupt change features between adjacent depth nodes in the enhanced spatiotemporal field map. The second dilated convolutional unit is set in parallel with the first dilated convolutional unit, using a 3×3 convolutional kernel, an expansion rate of 2, and a stride of 1, to capture the stress transfer and transition characteristics in the cross-node region of the enhanced spatiotemporal field map; The third dilated convolutional unit is set in parallel with the first dilated convolutional unit and the second dilated convolutional unit. It uses a 3×3 convolutional kernel, an expansion rate of 3, and a stride of 1. It is used to expand the receptive field without reducing the resolution and extract the overall elastic flexural modal features of the support pile. The feature fusion unit is connected to the outputs of the first dilated convolution unit, the second dilated convolution unit, and the third dilated convolution unit. It is used to concatenate the features output by the three dilated convolution units and perform cross-channel weighted fusion of features extracted with different dilation rates. One end of the residual connection unit is connected to the model input, and the other end is connected to the output of the feature fusion unit. It is used to fuse the enhanced spatiotemporal field map with the weighted fusion features to obtain the spatial morphological features.
5. The deep foundation pit support pile micro-deformation capture and prediction system integrating physical mechanism constraints according to claim 4, characterized in that, The dynamic displacement field is: ； Where Xdis represents the dynamic displacement field. T Indicates time, N Represents a node. uT , N Indicates at time T Deep nodes N Horizontal displacement monitoring value at the location, R Represents the set of real numbers; The static physical field is: ； Where Xstat represents the static physical field, eN Represents a node N The soil porosity at that location cN Represents a node N Cohesion at the point, kN Represents a node N The horizontal resistance coefficient of the soil at that location, M It represents the number of physical and mechanical parameters.
6. The deep foundation pit support pile micro-deformation capture and prediction system integrating physical mechanism constraints as described in claim 5, is characterized in that, The enhanced spatiotemporal field map is as follows: ； Among them, Xinput represents the enhanced spatiotemporal field map, Concat represents the channel splicing operation, Reshape represents the tensor dimension transformation, and Tile represents copy expansion.