An intelligent detection method and system for uneven settlement of a super-high building foundation

Abstract

The application relates to an intelligent detection method and system for uneven settlement of a super-high building foundation, and belongs to the technical field of settlement detection. The method comprises the following steps: collecting settlement monitoring data and constructing a training data set; calculating local statistics according to the spatial topological relationship of monitoring points, combining adjacent point set standardized original data to obtain standardized settlement values; amplifying the contribution of the mutation area by a second-order difference algorithm curvature mutation factor and a local variance weight factor, splicing the standardized settlement characteristic vector and the curvature mutation factor to obtain an enhanced characteristic vector; constructing an uneven settlement detection model, inputting the enhanced characteristic vector into the model to obtain an uneven settlement probability and a local anomaly heat map vector; optimizing the training process of the model through a loss function to obtain a trained model; and inputting the processed real-time settlement monitoring data collected newly into the trained model to obtain an intelligent detection result. The application can improve the detection capability of the uneven settlement of the foundation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of settlement detection, and particularly relates to an intelligent detection method and system for uneven settlement of a super high-rise building foundation. BACKGROUND

[0002] With the acceleration of urbanization and the continuous rise of building height, super high-rise buildings have become an important structural form of modern cities, and the foundation bearing thereof is complex, the load distribution is uneven, and the settlement behavior presents significant spatial coupling and nonlinear characteristics. Since the building foundation is usually formed by the joint action of multiple pile foundations or large rafts, affected by factors such as differences in soil properties, construction errors and load changes, the settlement of each part is often different. When the difference exceeds the allowable range, it will cause additional internal forces, cracks and even tilting of the structure, and other engineering safety problems. Therefore, how to accurately identify the uneven settlement of the foundation and carry out early warning has become a key technical problem in long-term safety monitoring of super high-rise buildings.

[0003] The existing building foundation settlement detection method mainly relies on the settlement data of the monitoring points obtained by periodic leveling or static leveling system, and determines whether there is an anomaly through statistical analysis, interpolation fitting or empirical threshold. Although this method can realize the overall trend evaluation of the foundation settlement, it has many limitations: in the prior art, conventional settlement detection mainly relies on statistical data of independent monitoring points, lacks modeling of spatial adjacency relationship, and thus the ability to capture the physical constraints and settlement transmission rules between adjacent points is insufficient; the existing method usually directly uses the original settlement value or the first difference feature, and cannot effectively utilize the curvature change feature, so the identification accuracy of small uneven settlement and potential structural risk is low; the traditional neural network model ignores the spatial position and distance characteristics of the monitoring points, and the weight learning cannot reflect the topological association of the foundation components, so the features learned by the model lack physical interpretability and spatial consistency; the existing uneven settlement detection model mainly has a single output, and cannot simultaneously consider global trend judgment and local anomaly positioning, so the recognition result is coarse and the engineering guidance value is limited. SUMMARY

[0004] The application provides an intelligent detection method and system for uneven settlement of a super high-rise building foundation.

[0005] To achieve the above purpose, the application realizes the following technical scheme:

[0006] The application provides an intelligent detection method for uneven settlement of a super high-rise building foundation, which comprises the following steps:

[0007] S1, collecting settlement monitoring data through monitoring points arranged at key positions of the building foundation, and constructing a training data set;

[0008] S2. Calculate local statistics by predefined spatial topological relationships between monitoring points, and standardize the original settlement data based on the set of adjacent points to obtain standardized settlement values.

[0009] Furthermore, in the task of detecting uneven settlement of foundations of super high-rise buildings, the settlement monitoring data has obvious spatial location correlation. The settlement values ​​of adjacent monitoring points are affected by physical constraints. Conventional Z-score standardization methods only process based on single-point statistical distribution, ignoring the spatial topology between monitoring points. This results in the loss of spatial correlation information in the standardized data, which cannot effectively preserve physical constraints, thereby weakening the ability of subsequent models to learn spatial patterns.

[0010] This invention predefines the spatial topological relationships between monitoring points based on the building foundation drawing, and determines the set of adjacent points for each monitoring point. Based on the set of adjacent points for each monitoring point, the original settlement values ​​of all monitoring points within the set are aggregated and calculated to obtain a local mean. At the same time, the local standard deviation is obtained by calculating the dispersion of settlement values ​​within the same set of adjacent points. For each monitoring point, the original settlement value is centered and scaled using the local mean and local standard deviation. Simultaneously, a scaling factor constructed using the local coefficient of variation is used for enhancement processing to obtain a standardized settlement value.

[0011] S3. In uneven settlement detection, abrupt changes in the curvature of the settlement curve are an essential characteristic. Conventional methods directly use the original or standardized settlement values ​​as model input, making it difficult to capture curvature change characteristics. This results in insufficient sensitivity of the model to minor uneven settlement and an inability to effectively identify potential dangerous settlement patterns. This invention calculates the curvature abrupt change factor using second-order difference and amplifies the contribution of the abrupt change region by combining it with a local variance weighting factor. The standardized settlement feature vector is then concatenated with the curvature abrupt change factor to obtain an enhanced feature vector.

[0012] Furthermore, this invention, based on the ratio of the variance of settlement values ​​within the set of adjacent points of each monitoring point to the global maximum variance, and combined with the ratio of the local standard deviation of the monitoring point to the global maximum local standard deviation, obtains a local variance weighting factor through an exponential amplification mechanism and linear superposition operation. For intermediate monitoring points, second-order difference calculation is performed based on the original settlement values ​​of the current point and its adjacent monitoring points before and after it, and the result is multiplied by the local variance weighting factor to obtain a curvature abrupt change factor. For boundary monitoring points, a one-sided difference calculation method is used, and the calculation result is multiplied by the local variance weighting factor to obtain a boundary curvature abrupt change factor. The curvature abrupt change factor vector is multiplied element-wise with the standardized settlement feature vector compressed by the hyperbolic tangent function, and then the standardized settlement feature vector is concatenated with the product result to obtain an enhanced feature vector.

[0013] S4. Construct a non-uniform settlement detection model, including a spatially perceptual fully connected layer, a dual-channel feature fusion module, and a non-uniform settlement prediction layer; input the enhanced feature vector into the non-uniform settlement detection model to obtain the non-uniform settlement probability and local anomaly heatmap vector; optimize the training process of the model through a loss function to obtain a trained model.

[0014] Furthermore, in settlement detection, traditional fully connected layers process all input features equally, but the spatial differences in the location of monitoring points lead to different contributions to the classification results. Conventional neural network layers ignore the spatial topological relationships between monitoring points and cannot effectively utilize the physical constraints between adjacent monitoring points, resulting in poor feature coupling and affecting the model's ability to learn spatial distribution patterns.

[0015] The spatially aware fully connected layer of this invention maps spatial distance through learnable position encoding and constructs a spatial modulation matrix to modulate the basic weights. The specific steps are as follows:

[0016] The spatial distance between hidden neurons and input features is calculated by learningable positional encoding. This distance is then combined with a spatial decay coefficient and a logarithmic transformation term of the local standard deviation for weighted adjustment to construct a spatial modulation matrix. The basic weight matrix and the spatial modulation matrix are multiplied element-wise, and the weighted influence of the self-spliced ​​curvature mutation factor vector is added to obtain the modulation weight matrix. The modulation weight matrix is ​​used to perform a linear transformation on the enhanced feature vector. The local standard deviation vector is projected and combined with the bias term as a dynamic offset. The spatially perceptual features are then output through the ReLU activation function.

[0017] Furthermore, in the task of settling pattern recognition, it is necessary to capture both global trends and local abrupt changes. Traditional single feature extraction methods are difficult to take into account settlement patterns at different scales. Conventional neural network structures either overemphasize global statistical features and ignore local details, or overemphasize local abrupt changes and ignore the overall trend, resulting in insufficient ability of the model to recognize complex and uneven settlement patterns.

[0018] The dual-channel feature fusion module of this invention constructs a dual-channel feature fusion structure, utilizes the global channel to extract overall trend features, utilizes the local channel to extract significant abrupt change features, and employs a gated fusion mechanism to dynamically adjust the weights of the original features. The specific steps are as follows:

[0019] The spatial sensing features are subjected to global average pooling to obtain basic statistical features. These basic statistical features are then concatenated with the average of the local means, the maximum local standard deviation, and the mean of the curvature mutation factor of all monitoring points. A linear projection is then applied to obtain the projected global feature representation. The spatial sensing features are then subjected to global max pooling to obtain local salient features. These local salient features are then concatenated with the minimum local mean, the maximum local standard deviation, and the maximum curvature mutation factor. A linear projection is then applied to obtain the projected local feature representation. After concatenating the projected global and projected local feature representations, gated attention weights are generated using a weight matrix mapping and a sigmoid function. These gated attention weights are then used to dynamically modulate the spatial sensing features. The modulated result is then concatenated with the dual-channel features to obtain the dual-channel fused features.

[0020] Furthermore, in the detection of uneven settlement, it is necessary to consider the global trend, local abrupt changes and spatial physical constraints. Conventional output layers directly use fully connected networks for binary classification, ignoring the multi-scale spatiotemporal characteristics and engineering physical significance of uneven settlement. This results in insufficient granularity of the model for judging complex settlement patterns and an inability to provide local anomaly location information.

[0021] The uneven settlement prediction layer of this invention generates uneven settlement probability and local anomaly heatmap vectors by constructing a multi-task adaptive output module and utilizing dual-channel fusion features, local statistics, and curvature mutation factors. The specific steps are as follows:

[0022] Based on dual-channel fusion features, combined with local mean, local standard deviation and curvature mutation factor, multi-scale feature weights are calculated through gating attention mechanism to obtain enhanced discriminative feature vector; the enhanced discriminative feature vector is concatenated with local anomaly factor and mapped to latent space through multilayer perceptron, and the probability of uneven settlement is calculated by combining global statistics; using the enhanced discriminative feature vector and spatial modulation matrix, local anomaly heatmap vector is obtained through deconvolution layer and spatial interpolation technique.

[0023] S5. The newly collected real-time settlement monitoring data is processed through steps S2 and S3 and then input into the trained model to obtain intelligent detection results.

[0024] Furthermore, the loss function includes the overall uneven settlement probability loss and the local anomaly heatmap loss. The overall uneven settlement probability loss adopts the binary cross-entropy loss function to calculate the deviation between the predicted probability and the true overall label in order to optimize the global classification performance. The local anomaly heatmap loss uses the mean squared error loss function to quantify the difference between the predicted heatmap vector and the true local label, ensuring accurate localization of anomaly areas.

[0025] Furthermore, the total loss is a weighted sum of the overall uneven settlement probability loss and the local anomaly heatmap loss. The weight ratio is adjusted by hyperparameters to balance the global classification and local detection tasks.

[0026] This invention also provides an intelligent detection system for uneven settlement of foundations of super high-rise buildings, which executes the aforementioned intelligent detection method for uneven settlement of foundations of super high-rise buildings, including:

[0027] Data acquisition module: used to collect settlement monitoring data through monitoring points deployed at key locations on the building foundation;

[0028] Data standardization module: used to calculate local statistics based on the spatial topological relationship between predefined monitoring points, and to standardize the original settlement data based on the set of adjacent points to obtain standardized settlement values;

[0029] Feature enhancement module: used to calculate curvature abrupt change factor through second-order difference, amplify the contribution of abrupt change region by combining local variance weighting factor, and concatenate the standardized settlement feature vector with curvature abrupt change factor to obtain enhanced feature vector;

[0030] Model building module: used to build a non-uniform settlement detection model, including a spatially aware fully connected layer, a dual-channel feature fusion module, and a non-uniform settlement prediction layer; and optimizes the model training process through a loss function to obtain a trained model;

[0031] Uneven settlement detection module: This module processes newly acquired real-time settlement monitoring data and inputs it into the trained model to obtain intelligent detection results.

[0032] The advantages of this invention are:

[0033] This invention introduces a standardization method based on spatial adjacency constraints, which preserves the spatial topological correlation of settlement monitoring data during the standardization process. Compared with the traditional Z-score standardization method, which is based on single-point statistics, this method can more realistically reflect the physical coupling relationship between monitoring points of the building foundation, thereby improving the model's ability to learn spatial patterns. A feature enhancement mechanism based on curvature mutation factors is proposed, combining second-order difference curvature changes with local variance weights to create a dynamic amplification effect on local mutation regions. This makes the model more sensitive to small uneven settlements and can capture potential structural deformation and crack risk points. A spatially perceptive fully connected layer is constructed, using learnable position encoding and spatial attenuation mechanisms to modulate the weight matrix. This gives the feature coupling the dual characteristics of spatial distance attenuation and local fluctuation response, thus integrating building physical constraints into the deep network and achieving structured modeling of the spatial distribution pattern of settlement. A dual-channel feature fusion and multi-task output mechanism is designed to extract global trend and local mutation features respectively, and achieve dynamic weighting through gated fusion. This allows the model to adaptively balance the detection capabilities of overall settlement and local anomalies, while simultaneously outputting global settlement probability and local anomaly heatmaps for refined diagnosis. Attached Figure Description

[0034] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0035] Figure 1 This is a flowchart of the steps of the present invention;

[0036] Figure 2 This is a heatmap of the original settlement data of the present invention;

[0037] Figure 3 This is a standardized data heatmap of the present invention;

[0038] Figure 4 This represents the local standard deviation distribution of all samples in this invention.

[0039] Figure 5 This represents the distribution of curvature change intensity at each monitoring point for all samples in this invention.

[0040] Figure 6 This describes the actual layout and spatial connection of the monitoring points on the building foundation plane according to the present invention. Detailed Implementation

[0041] 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.

[0042] Example 1

[0043] In this embodiment, as Figure 1 As shown, this invention provides an intelligent detection method for uneven settlement of foundations of super high-rise buildings, the specific steps of which include:

[0044] S1. Settlement monitoring data acquisition and construction of training dataset

[0045] Settlement monitoring data was collected by setting up monitoring points at key locations on the building foundation, and a training dataset was constructed based on engineering practice and expert knowledge.

[0046] Specifically, based on the building foundation drawings (such as raft foundation layout drawings or pile foundation positioning drawings), monitoring points are set up at the actual locations of foundation components such as the center point of the pile foundation and the corner point of the pile cap to form a monitoring network covering the entire foundation;

[0047] Settlement monitoring data is collected periodically using a high-precision level or hydrostatic leveling system, recording the settlement value at each monitoring point to form time series data.

[0048] To construct the training dataset, the original settlement data needs to be labeled. The labeling method is based on engineering measured data and expert evaluation, defining overall uneven settlement labels and local anomaly labels.

[0049] The overall uneven settlement label is a binary variable, indicating whether there is significant uneven settlement in the entire foundation. It is determined based on the degree of curvature change of the settlement curve and whether the settlement difference exceeds the safety threshold.

[0050] The local anomaly label is a binary variable for each monitoring point, indicating whether the point belongs to an abnormal settlement area. It is calculated based on the difference between the local settlement value and the values ​​of adjacent points, as well as the curvature mutation factor.

[0051] The labeling categories include positive samples (with uneven sedimentation) and negative samples (without significant uneven sedimentation). Meanwhile, local anomaly labels are used to train local anomaly heatmaps to provide fine-grained localization information.

[0052] S2, Data Standardization for Enhanced Spatial Correlation

[0053] In the task of detecting uneven settlement of foundations of super high-rise buildings, the settlement monitoring data has obvious spatial location correlation. The settlement values ​​of adjacent monitoring points are affected by physical constraints. Conventional Z-score standardization methods only process based on single-point statistical distribution, ignoring the spatial topology between monitoring points. This results in the loss of spatial correlation information in the standardized data, which cannot effectively preserve physical constraints and thus weakens the ability of subsequent models to learn spatial patterns.

[0054] The standardization method for spatial adjacency constraints in this invention calculates local statistics by predefining the spatial topological relationships between monitoring points, and standardizes the original settlement data based on the set of adjacent points. This ensures that the standardized data retains spatial correlation and provides physically meaningful structured input for subsequent models. The specific steps are as follows:

[0055] 1) Constructing a spatial adjacency topology

[0056] Based on the predefined spatial topological relationships between monitoring points according to the building foundation drawing, the set of adjacent points for each monitoring point is determined. Specifically, the first... The set of neighboring points of each monitoring point is , For the first The size of the set of neighboring points of the nth monitoring point, that is, the size of the set of neighboring points of the nth monitoring point. The number of adjacent points of each monitoring point For monitoring point index.

[0057] It should be noted that the building foundation drawing is the structural design drawing of a super high-rise building, such as a raft foundation layout drawing or a pile foundation positioning drawing. When the spatial topological relationship between monitoring points is predefined according to the building foundation drawing, the monitoring points are first arranged according to the actual positions of the foundation components such as the center point of the pile foundation and the corner point of the pile cap in the drawing. Then, the set of adjacent points is determined by calculating the Euclidean distance between the monitoring points. For example, in a rectangular raft foundation, the monitoring points are arranged in a grid, and adjacent points can be defined as the set of nearest neighbors within the Euclidean distance threshold.

[0058] 2) Calculate the local mean and local standard deviation.

[0059] Based on the set of neighboring points for each monitoring point, the original settlement values ​​of all monitoring points within this set are aggregated to obtain a local mean reflecting the settlement trend of the local area. Simultaneously, based on the dispersion of settlement values ​​within the same set of neighboring points, a local standard deviation measuring the settlement fluctuation of the local area is obtained, expressed as:

[0060] ,

[0061] ,

[0062] In the formula, Indicates the first The local mean of each monitoring point is used to capture the settlement trend of a local area; Indicates the first The local standard deviation of each monitoring point is used to measure the settlement fluctuation in a local area. Indicates the first The original settlement values ​​of each monitoring point; For monitoring point index.

[0063] 3) Generate standardized output

[0064] For each monitoring point, the original settlement value is centered and scaled using its local mean and local standard deviation. Simultaneously, a scaling factor constructed from the local coefficient of variation is used for enhancement processing to obtain standardized settlement values ​​that retain spatial correlation. This standardized value highlights the degree of local regional heterogeneity while maintaining spatial correlation, as expressed below:

[0065] ,

[0066] In the formula, Indicates the first The standardized settlement values ​​of each monitoring point not only preserve the spatial correlation of settlement data, but also enhance the local relative fluctuation information through local variation coefficient scaling, thereby highlighting the unevenness of local areas of the building foundation. To prevent division by zero in the denominator, it is preferably set to 0.00001; Indicates the first The original settlement values ​​of each monitoring point.

[0067] It should be noted that the ratio of the local standard deviation to the local mean, used as the local coefficient of variation, incorporates spatial topological information into the standardization process. This term utilizes local coefficients of variation to enhance information about local relative fluctuations. The hyperbolic tangent function maps the coefficients of variation to... The range can avoid the influence of extreme values ​​and highlight the unevenness of local areas. It not only preserves the spatial correlation of the data, but also amplifies the signal of areas with significant fluctuations through the scaling factor, making the model more sensitive to local anomalies of uneven settlement of the foundation.

[0068] S3, Feature Enhancement Based on Curvature Mutation Factor

[0069] In the detection of uneven settlement, the abrupt change in curvature of the settlement curve is an essential feature. Conventional methods directly use the original settlement value or the standardized settlement value as the model input, which makes it difficult to capture the curvature change feature. This results in insufficient sensitivity of the model to small uneven settlement and an inability to effectively identify potential dangerous settlement patterns.

[0070] This invention calculates the curvature abrupt change factor using second-order difference and amplifies the contribution of abrupt change regions by combining it with a local variance weighting factor. The standardized features are then concatenated with the curvature abrupt change factor to form enhanced features, thereby strengthening local abrupt change information and improving the model's sensitivity to minor uneven settlement. The specific steps are as follows:

[0071] 1) Calculate the local variance weighting factor

[0072] Based on the ratio of the variance of settlement values ​​within the set of adjacent points of each monitoring point to the global maximum variance, and combined with the ratio of the local standard deviation of that monitoring point to the global maximum local standard deviation, a local variance weighting factor is generated to enhance the contribution of high-fluctuation areas through an exponential amplification mechanism and linear superposition operation, expressed as:

[0073] ,

[0074] In the formula, Indicates the first The local variance weighting factor for each monitoring point is used to adjust the weight of the curvature abrupt change factor; Indicates the first The set of adjacent points of each monitoring point The variance of the internal settlement value characterizes the degree of settlement fluctuation in a local area; Represents the natural exponential function; This represents the maximum local standard deviation across all monitoring points; This represents the maximum local variance across all monitoring points; This represents the set of local variance values ​​for all monitoring points, obtained based on the local standard deviation of all monitoring points.

[0075] 2) Calculate the curvature mutation factor

[0076] For intermediate monitoring points, second-order difference calculations are performed based on the original settlement values ​​of the current point and its adjacent monitoring points before and after it. The result is then multiplied by the local variance weighting factor to obtain the curvature abrupt change factor that quantifies the local curvature change. For boundary monitoring points, a one-sided difference calculation method is used, and the result is also multiplied by the corresponding weighting factor to generate the boundary curvature abrupt change factor, expressed as:

[0077] For intermediate monitoring points, i.e. The calculation method is as follows ;

[0078] For boundary monitoring points, i.e. The calculation method is as follows , .

[0079] in, Indicates the first The curvature abrupt change factor at each monitoring point is used to quantify the degree of local curvature change. This represents the curvature mutation factor at the first monitoring point. This represents the curvature mutation factor at the 20th monitoring point; This represents the local variance weighting factor for the first monitoring point. This represents the local variance weighting factor for the 20th monitoring point; Indicates the first The original settlement values ​​of each monitoring point Indicates the first The original settlement values ​​of each monitoring point This represents the initial settlement value at the first monitoring point. This represents the original settlement value at the second monitoring point. This represents the original settlement value at the 20th monitoring point. This represents the original settlement value at the 19th monitoring point.

[0080] 3) Construct enhanced feature vectors

[0081] The curvature abrupt change factor vector is multiplied element-wise with the standardized settlement feature vector compressed by the hyperbolic tangent function. The standardized settlement feature vector is then concatenated with the product result to form an enhanced feature vector that simultaneously contains spatial standardization information and curvature abrupt change intensity information, expressed as:

[0082] ,

[0083] In the formula, This represents the enhanced eigenvector, which characterizes the combined features of static relative deviation and dynamic change intensity in the settlement field; Represents the standardized settlement feature vector. The The standardized settlement value of each monitoring point is ; Represents the curvature mutation factor vector, i.e. ; This represents a vector concatenation operation; It is the hyperbolic tangent function; This represents the Hadamard product, which is an element-wise multiplication.

[0084] It should be noted that, The term is used to amplify the contribution in regions of high variance, while The method utilizes local statistics, enabling the weighting factors to not only reflect local fluctuations but also to work synergistically with the spatial standardization process. This results in a synergistic amplification effect of the curvature mutation factor in boundary regions and areas of intense fluctuation, enhancing the model's sensitivity to potential hazardous settlement patterns and thus more accurately characterizing local anomalies in uneven settlement.

[0085] It should also be noted that in enhancing feature vectors During the construction process, the curvature abrupt change factor is interacted with the normalized settlement characteristics through the Hadamard product, and combined with... The term further compresses the feature range to avoid numerical instability. By splicing, it not only retains the standardized spatial distribution information, but also enhances the dynamic change features of curvature mutation. This makes the enhanced feature vector contain both static relative deviation and dynamic change intensity, so that the curvature mutation factor and the standardized features are coupled, and the interactive representation between features is enhanced.

[0086] S4. Construction of Uneven Settlement Detection Model

[0087] S401, Spatial Awareness Fully Connected Layer

[0088] In settlement detection, traditional fully connected layers process all input features equally. However, the spatial differences in the location of monitoring points lead to varying contributions to the classification results. Conventional neural network layers ignore the spatial topological relationships between monitoring points and cannot effectively utilize the physical constraints between adjacent monitoring points, resulting in poor feature coupling and affecting the model's ability to learn spatial distribution patterns. This invention maps spatial distance through learnable location encoding and constructs a spatial modulation matrix to modulate the basic weights, thereby enabling stronger feature coupling between adjacent monitoring points and improving the model's ability to model spatial correlations. The specific steps are as follows:

[0089] 1) Constructing the spatial modulation matrix

[0090] The spatial distance between hidden neurons and input features is calculated using learnable positional encoding. This distance is then combined with a spatial attenuation coefficient and a weighted adjustment using a logarithmic transformation term of the local standard deviation. This results in a spatial modulation matrix that integrates the spatial distance attenuation effect and the local fluctuation response, expressed as follows:

[0091] ,

[0092] In the formula, Represents the spatial modulation matrix. The first element of the spatial modulation matrix Line number Column elements are used to modulate the base weights; Indicates the first The nth hidden layer neuron corresponds to the nth... Spatial distance of each input feature; Indicates the first The nth hidden layer neuron corresponds to the nth... Spatial distance of each input feature; For the first Local standard deviation of each monitoring point; Index of hidden layer neurons; Input feature index; This represents the spatial attenuation coefficient, which controls the intensity of distance attenuation. The preferred setting is 0.95.

[0093] In specific implementation, the first The nth hidden layer neuron corresponds to the nth... Spatial distance of each input feature It is obtained through a learnable positional encoding mapping, specifically the base. Hidden layer neurons and bottom Each input feature is assigned a learnable coordinate vector. and ,in Here, represents the spatial coordinate vector of the hidden layer neurons, which is 2-dimensional and represents their position in a two-dimensional plane. The input feature is a spatial coordinate vector with a dimension of 2, and then it is calculated. and The Euclidean distance between them is obtained Learnable location encoding allows the model to adaptively learn the spatial distribution of monitoring points, enhancing the model's flexibility.

[0094] It should be noted that, through the spatial attenuation coefficient Control the degree of distance attenuation and combine it with local standard deviation. logarithmic terms This enhances the modulation weights in high-fluctuation regions, enabling the modulation matrix to not only reflect the attenuation effect of spatial distance but also respond to local subsidence fluctuations, thus achieving dual modulation of weights based on both spatial distance and the degree of local fluctuations.

[0095] 2) Calculate the modulated weight matrix

[0096] The basic weight matrix and the spatial modulation matrix are multiplied element-wise, and then the weighted influence of the self-spliced ​​and expanded curvature abrupt change factor vector is added to generate a modulation weight matrix that integrates spatial topological relationships, local wave characteristics, and curvature abrupt change information, as follows:

[0097] ,

[0098] In the formula, The modulation weight matrix incorporates spatial topological relationships and curvature abrupt change information, representing the weight modulation result based on spatial distance, local fluctuations, and curvature abrupt changes; These are weighting coefficients used to adjust the influence of the curvature mutation vector on weight modulation. The preferred value is set to ; This represents the basic weight matrix, with dimensions of . , are trainable parameters that are randomly initialized and optimized during training using gradient descent, and are used to learn the global linear mapping relationship between features; This refers to the hidden layer dimension, which is the number of neurons in the spatially perceptual fully connected layer.

[0099] It should be noted that, The term will be the curvature mutation factor vector. splicing itself to match The dimension of the curvature mutation factor vector After being spliced ​​together, it expands to 40 dimensions and uses a self-referential reinforcement mechanism to double the contribution of curvature abrupt change regions, promote feature coupling, and thus allow the model to more sensitively identify boundary points such as raft corners and areas of violent fluctuation, thereby improving the ability to capture patterns of small uneven settlement.

[0100] It should be noted that the base weight matrix differs from that of traditional fully connected layers. It is not used directly for calculation, but rather through the spatial modulation matrix. Modulate the curvature mutation factor vector to preserve the basic feature interaction capability.

[0101] 3) Calculate the output of the fully connected layer

[0102] The enhanced feature vector is linearly transformed using a modulation weight matrix, and the local standard deviation vector is projected and combined with the bias term as a dynamic offset. The ReLU activation function is then used to output a spatially aware feature representation, expressed as:

[0103] ,

[0104] In the formula, This represents the output features of the first fully connected layer, characterizing the spatially sensed modulated feature vector. The output is a linear mapping of the local standard deviation vector, dimension matching. Specifically, this is achieved through projection matrix mapping, represented as... ;

[0105] Let be the local standard deviation vector of all monitoring points, the th The local standard deviation of each monitoring point is ; The mappable weight matrix consists of trainable parameters used to map the local standard deviation vector. Linear projection onto the bias term of the fully connected layer Same dimensions; The mapping bias term is a trainable parameter. Represents the ReLU activation function; The modulation weight matrix represents trainable parameters that enhance the feature vector. Perform a linear transformation; This represents the bias term of the fully connected layer, which is a trainable parameter used to dynamically adjust the feature offset.

[0106] It should be noted that the settlement behavior of adjacent points is constrained by the physical structure, such as the physical structural constraints of raft continuity, in the spatial modulation matrix. In the calculation process, Euclidean distance is calculated through learnable location codes, and combined with... The weights are implemented using a dual modulation of spatial distance attenuation and local fluctuations. The former enhances the weights of adjacent points, mimicking the connection strength of physical constraints, while the latter amplifies the contribution of high subsidence fluctuation zones, responding to potential hazardous modes. Furthermore... After the item is self-assembled, it is connected with... The superposition produces a synergistic amplification effect, significantly increasing the weights of boundary points such as the raft edge and areas of abrupt curvature changes. The multimodal modulation mechanism combining distance, fluctuation, and curvature is not a simple weighting, but rather a deep fusion in the feature space through the Hadamard product. This model not only captures static settlement distribution but also enhances sensitivity to dynamic curvature changes, such as strengthening the sensitivity to settlement abrupt changes at micro-cracks. In the context of settlement detection, the term and self-splicing method transforms building foundation constraints into learnable parameters. This kind of modulation based on physical mechanisms cannot be easily achieved by conventional neural network design because the weights of conventional methods are not related to spatial correlation and cannot adaptively form local feature clusters.

[0107] S402, Dual-channel feature fusion module

[0108] In settlement pattern recognition tasks, it is necessary to simultaneously capture global trends and local abrupt changes. Traditional single-feature extraction methods struggle to handle settlement patterns at different scales. Conventional neural network structures either overemphasize global statistical features while neglecting local details, or overemphasize local abrupt changes while ignoring the overall trend, resulting in insufficient recognition capabilities for complex and uneven settlement patterns. This invention constructs a dual-channel feature fusion structure, utilizing a global channel to extract overall trend features and a local channel to extract significant abrupt changes. A gated fusion mechanism is employed to dynamically adjust the weights of the original features, enabling the model to adaptively fuse global statistical features with local salient features. The specific steps are as follows:

[0109] 1) Global channel feature extraction

[0110] The basic statistical features are obtained by global average pooling of the spatial sensing features. These features are then concatenated with the average of the local means, the maximum local standard deviation, and the mean of the curvature mutation factor for all monitoring points. A global feature representation integrating global trends and key physical quantities is generated through linear projection, as follows:

[0111] ,

[0112] In the formula, This represents the average of the local mean values ​​of all monitoring points, used to capture the overall settlement trend; This represents the maximum local standard deviation across all monitoring points, used to amplify the contribution of global fluctuations. This represents the average curvature abrupt change factor across all monitoring points, used to incorporate global curvature change information; This represents the pooled output of the global channel, obtained by analyzing the output features of the first fully connected layer. Obtained by global average pooling; The projection global feature representation is in vector format, which characterizes the overall trend of the settlement field and integrates global statistical information such as local mean, local standard deviation and curvature change factor to enhance the physical meaning and discriminative ability of the global features. The projection weight matrix representing the global channel is a trainable parameter that linearly maps the concatenated global features to a high-dimensional space to extract richer global patterns. The bias term representing the global channel is a trainable parameter.

[0113] 2) Local channel feature extraction

[0114] Global max pooling is applied to spatially perceived features to obtain locally salient features. These features are then concatenated with the local minimum mean, the local maximum standard deviation, and the maximum curvature abrupt change factor. Linear projection is then used to generate a local feature representation that enhances local anomaly and extreme value information, as follows:

[0115] ,

[0116] In the formula, This represents the pooled output of a local channel, obtained by analyzing the output features of the first fully connected layer. Obtained by performing global max pooling; The projection local feature representation is a vector format that characterizes local abrupt change points and anomalous regions in the settlement field. It integrates the local mean minimum, local standard deviation maximum, and curvature abrupt change factor maximum to highlight the local details of uneven settlement. The projection weight matrix, representing the local channel, is a trainable parameter that linearly maps the spliced ​​local features to a high-dimensional space to enhance sensitivity to local mutations. The bias term representing the local channel is a trainable parameter; This represents the minimum local mean value of all monitoring points, used to capture areas of minimal local settlement. This represents the maximum local standard deviation across all monitoring points, used to emphasize areas of high volatility. This represents the maximum value of the curvature abrupt change factor across all monitoring points, used to amplify the contribution of local curvature abrupt changes.

[0117] 3) Dual-channel feature fusion

[0118] After concatenating the projected global feature representation and the projected local feature representation, gated attention weights are generated through weight matrix mapping and the Sigmoid function. These weights are then used to dynamically modulate the spatial perception features, and the modulation result is concatenated with the dual-channel features to output a dual-channel fused feature that integrates global trends and local mutations, as shown below:

[0119] ,

[0120] In the formula, This represents the output characteristics of the second fully connected layer, which is also the final output characteristics of the dual-channel fusion module; This represents the Sigmoid activation function, which maps the input to... interval; The fusion weight matrix represents trainable parameters that map the concatenated global and local features to attention weights through linear transformation, thereby dynamically adjusting the contribution of the original features.

[0121] It should be noted that, considering the physical characteristics of uneven settlement of foundations in super high-rise buildings, there may be an overall trend of raft foundation settlement or a local abrupt change in differential settlement of pile foundations. Therefore, the settlement mode needs to consider both the overall trend and the local abrupt change. By separating global and local feature extraction through a dual-channel structure and using a gating fusion mechanism to achieve adaptive weighting, the projections of global and local channels incorporate local statistics that already contain spatial topology and curvature information. This makes the dual-channel features not only based on neural network outputs but also directly introduce intermediate features with physical constraints, thereby enhancing the physical interpretability of the features.

[0122] It should also be noted that, The dynamic generation of attention weights allows the model to adjust the fusion ratio of global and local features based on the input data, when the average curvature abrupt change factor of all monitoring points is... When the gating weights are high, the gating weights automatically amplify the contribution of local features, making the model pay more attention to cracks or edge regions. Conversely, when the global trend dominates, the gating strengthens the weights of global features to maintain overall consistency. This adaptive fusion is not a simple weighted average, but rather a fusion that combines the Hadamard product with the output features of the first fully connected layer. Interaction allows the output features of the second fully connected layer to be displayed. It retains both the original spatial perception features and enhanced multi-scale information.

[0123] It should also be noted that uneven settlement often manifests as a superposition of slow global changes and rapid local abrupt changes. The combination of terms and gating modulation ensures that the fused features contain both compressed global / local abstract representations and retain the original features rich in detail, enabling the model to adaptively capture complex settlement patterns from the center to the edge of the raft. The deep integration of physical priors of spatial statistics and curvature representations with data-driven learning can improve the model's generalization ability in real-world scenarios.

[0124] S403, Uneven Settlement Prediction Layer

[0125] In the detection of uneven settlement, it is necessary to consider the global trend, local abrupt changes and spatial physical constraints. Conventional output layers directly use fully connected networks for binary classification, ignoring the multi-scale spatiotemporal characteristics and engineering physical significance of uneven settlement. This results in insufficient granularity of the model in discriminating complex settlement patterns and an inability to provide information on the local anomaly location.

[0126] This invention constructs a multi-task adaptive output module, utilizing dual-channel fusion features, local statistics, and curvature mutation factors to generate uneven settlement probability distributions and local anomaly heatmap vectors. Furthermore, it dynamically fuses multi-source features through a gated residual mechanism, achieving collaborative optimization of high-precision global classification and local anomaly detection. The specific steps are as follows:

[0127] 1) Adaptive weighting of multi-scale features

[0128] Based on dual-channel fusion features, and combining local mean, local standard deviation, and curvature mutation factor, a gated attention mechanism is used to calculate multi-scale feature weights and dynamically modulate the features to generate enhanced discriminative feature representations. This achieves adaptive weighting of multi-scale features, allowing the enhanced features to simultaneously retain global distribution patterns and local anomaly details, thereby improving feature discriminative power. This is represented as follows:

[0129] ,

[0130] In the formula, This represents an enhanced discriminative feature vector that integrates multi-scale physical statistics and curvature abrupt change information. The gating weight matrix represents trainable parameters with dimension 1. ; For gated hidden layer dimensions; for The dimension; Number of monitoring points; This represents the gated bias term, which is a trainable parameter; It is a logarithmic function, with the default base being the natural constant; The local mean of all monitoring points, the first... The local mean of each monitoring point is ; For the local standard deviation of all monitoring points, the first... The local standard deviation of each monitoring point is .

[0131] It should be noted that, through the gating attention mechanism, the model can adaptively weight multi-scale features, where... Provide local trend priors, The amplified fluctuation area has an impact. The term enhances the curvature abrupt change signal, thereby enhancing the discriminative feature vector. At the same time, it preserves the global distribution pattern and local anomaly details, thereby improving feature discrimination power.

[0132] 2) Calculation of uneven settlement probability

[0133] The enhanced discriminative feature vector is concatenated with local anomaly factors, mapped to the latent space via a multilayer perceptron, and combined with global statistics to calculate the probability of uneven settlement. This achieves accurate calculation of the overall foundation uneven settlement probability. By embedding physical priors, the model's ability to discriminate complex settlement patterns is enhanced, as expressed below:

[0134] ,

[0135] In the formula, It represents the probability of uneven settlement of the overall foundation, and is used to determine whether there is significant uneven settlement; This represents the output weight vector, which consists of trainable parameters with dimension 1. ; This represents the output bias term, which is a trainable parameter; The hidden layer weight matrix represents trainable parameters with dimension 1. ; For hidden layer dimensions; To enhance the discriminative feature vector The dimension; This represents the hidden layer bias term, which is a trainable parameter.

[0136] It should be noted that, The local anomaly factor constructed by the term characterization is used to quantify the coupling effect between curvature abrupt change and local fluctuation, amplifying the anomaly contribution of boundary region and high variance region, while the hidden layer transformation captures the overall non-uniformity through global statistics, so that probability calculation not only depends on feature interaction, but also embeds physical priors.

[0137] 3) Generation of local anomaly heatmap vectors

[0138] By utilizing enhanced discriminative feature vectors and spatial modulation matrices, local anomaly heatmap vectors are generated through deconvolution layers and spatial interpolation techniques. Anomaly intensity is then calculated using the spatial gradient of the curvature abrupt change factor, outputting anomaly scores for each monitoring point. This enables precise local anomaly region localization, providing detailed local anomaly information for engineering decision-making, as represented by:

[0139] ,

[0140] In the formula, This represents a local anomaly heatmap vector, where each element corresponds to the anomaly score of a monitoring point. This represents the deconvolution weight matrix, which is a trainable parameter used for feature upsampling; This represents the deconvolution bias term, which is a trainable parameter; This indicates that the upsampling operation will enhance the discriminative feature vector. Mapped to the same dimension as the number of monitoring points; Represents the spatial modulation matrix The row mean vector is used to introduce spatial topological weights; Represents the curvature mutation factor vector The gradient magnitude is used, and the boundary points are quantized using one-sided difference to quantify the spatial drasticness of curvature changes.

[0141] It should be noted that, through deconvolution layers and spatial interpolation, high-dimensional features are decoded into heatmap vectors consistent with the spatial distribution of monitoring points, and combined with... Strengthen the physical constraints between adjacent points, and This method captures local abrupt changes in curvature, enabling the thermal map vectors to accurately locate key areas prone to uneven settlement, such as the edges of the raft foundation and the junctions of pile foundations. The final output includes the uneven settlement probability of the entire foundation. and local anomaly thermal map vector This provides dual support for engineering decision-making, offering both global assessment and local positioning.

[0142] S404, Loss Function Calculation and Model Training

[0143] During model training, the loss function is designed as a multi-task function, including two parts: overall uneven settlement probability loss and local anomaly heatmap loss.

[0144] Specifically, the overall uneven settlement probability loss adopts the binary cross-entropy loss function to calculate the deviation between the predicted probability and the true overall label in order to optimize the global classification performance;

[0145] For local anomaly heatmap loss, the mean squared error loss function is used to quantify the difference between the predicted heatmap vector and the true local label, ensuring accurate location of anomaly areas.

[0146] The total loss is a weighted sum of the overall uneven settlement probability loss and the local anomaly heatmap loss. The weight ratio is adjusted by hyperparameters to balance the global classification and local detection tasks.

[0147] The model training uses the Adam optimizer to update parameters and iteratively optimizes through the backpropagation algorithm. During training, the mini-batch gradient descent method is used to calculate the loss and update the weights on the training set.

[0148] Meanwhile, the loss changes are monitored on the validation set, and an early stopping mechanism is used as the criterion for stopping iteration. That is, when the validation set loss no longer decreases significantly or begins to increase within several consecutive iteration cycles (such as 10 iteration cycles), training is stopped to avoid overfitting. In addition, a maximum number of iterations (such as 100 iteration cycles) is set as an auxiliary stopping condition to ensure efficient convergence of training.

[0149] After training, the model parameters are saved for subsequent non-uniform settlement detection tasks.

[0150] S5. Uneven Settlement Detection of Foundations for Super High-Rise Buildings

[0151] After the model is trained, the process of detecting uneven settlement of the foundation of super high-rise buildings includes data input, feature processing and result output.

[0152] First, for the target super high-rise building, monitoring points are set up at key foundation locations such as the center point of the pile foundation and the corner point of the pile cap, following the data acquisition method in S1, and real-time settlement monitoring data is collected using sensors. The original settlement values ​​are processed using the spatial correlation enhancement data standardization method in S2 to preserve spatial topological correlation. Then, the feature enhancement step based on curvature mutation factor in S3 is executed to generate enhanced feature vectors to strengthen local mutation information. The processed data is input into the trained model constructed in S4, and the model outputs the overall uneven settlement probability and local anomaly heat map vectors.

[0153] The overall uneven settlement probability is used to determine whether there is significant uneven settlement in the foundation. If the probability exceeds 0.5, it is judged as abnormal. The local abnormal heat map vector provides an abnormal score for each monitoring point, and high-risk areas are located through visual heat map.

[0154] Example 2

[0155] In this embodiment, the original settlement data heatmap is as follows: Figure 2 As shown, taking the original settlement value distribution of the first 10 samples across 20 monitoring points as an example, the color intensity of each cell represents the magnitude of the settlement value. Rows represent different samples, and columns represent different monitoring points, showing the spatial distribution pattern of settlement values ​​at different locations. The visualized spatially standardized data heatmap is shown below. Figure 3As shown, the data is standardized after spatial adjacency constraints. Red indicates positive deviation (relatively large settlement), blue indicates negative deviation (relatively small settlement), and white indicates close to the local mean, reflecting the effect of preserving spatial correlation.

[0156] Example 3

[0157] In this embodiment, the local standard deviation distribution of all samples is as follows: Figure 4 As shown, the local standard deviation distribution plot of all samples shows the statistical distribution of the local fluctuation degree of 50 samples at each monitoring point. The local standard deviation distribution range of each of the 20 monitoring points is displayed in the form of a box plot. The box reflects the distribution range of the middle 50% of the data, the whiskers show the overall data range, and outliers indicate abnormal fluctuations. According to the distribution, it can reflect the stability characteristics of different locations of the building foundation. Monitoring points with larger fluctuations usually correspond to areas with weak structures or complex stress, providing important clues for identifying spatial fluctuation patterns of potential uneven settlement.

[0158] Example 4

[0159] In this embodiment, as Figure 5 As shown, the distribution of curvature abrupt change factors is presented in the form of a violin diagram, illustrating the intensity distribution of curvature changes for all samples at each monitoring point. The width of the violin represents the density of the data distribution, with wider sections corresponding to curvature values ​​occurring more frequently. This diagram demonstrates the statistical regularity of curvature changes on the building foundation surface. Areas with higher peak values ​​often correspond to inflection points or abrupt changes in the settlement curve. These locations are typically high-risk areas for crack formation or structural deformation in engineering, and are of significant indicative value for early warning of localized abnormal settlement.

[0160] Example 5

[0161] In this embodiment, as Figure 6 As shown, the actual layout and spatial connection of 20 monitoring points on the building foundation plane are simulated. Red dots represent the monitoring point locations, and gray lines represent spatial adjacency relationships determined by Euclidean distance. The topology directly reflects the physical constraints of the building foundation, simulating a grid-based monitoring network in actual engineering. The connection relationships between adjacent points ensure accurate capture of the influence range of local areas during spatial standardization, providing a realistic spatial constraint basis for subsequent feature processing.

[0162] Example 6

[0163] This embodiment provides an intelligent detection system for uneven settlement of foundations of super high-rise buildings, which implements the intelligent detection method for uneven settlement of foundations of super high-rise buildings described in Embodiment 1, including:

[0164] Data acquisition module: used to collect settlement monitoring data through monitoring points deployed at key locations on the building foundation;

[0165] Data standardization module: used to calculate local statistics based on the spatial topological relationship between predefined monitoring points, and to standardize the original settlement data based on the set of adjacent points to obtain standardized settlement values;

[0166] Feature enhancement module: used to calculate curvature abrupt change factor through second-order difference, amplify the contribution of abrupt change region by combining local variance weighting factor, and concatenate the standardized settlement feature vector with curvature abrupt change factor to obtain enhanced feature vector;

[0167] Model building module: used to build a non-uniform settlement detection model, including a spatially aware fully connected layer, a dual-channel feature fusion module, and a non-uniform settlement prediction layer; and optimizes the model training process through a loss function to obtain a trained model;

[0168] Uneven settlement detection module: This module processes newly acquired real-time settlement monitoring data and inputs it into the trained model to obtain intelligent detection results.

[0169] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent detection method for uneven settlement of a super high-rise building foundation, characterized in that, The method comprises the following steps: S1, collecting settlement monitoring data through monitoring points arranged at key positions of the building foundation, and constructing a training data set; S2, calculating local statistics through the spatial topological relationship between the predefined monitoring points, and performing standardization processing on the original settlement data based on the adjacent point set to obtain standardized settlement values; S3, calculating a curvature mutation factor through second-order difference, amplifying the contribution of the mutation area by combining a local variance weight factor, splicing the standardized settlement feature vector and the curvature mutation factor to obtain an enhanced feature vector; S4, constructing an uneven settlement detection model, including a spatial perception full connection layer, a double-channel feature fusion module, and an uneven settlement prediction layer; The enhanced feature vector is input into the uneven settlement detection model to obtain an uneven settlement probability and a local anomaly heat map vector; the training process of the model is optimized through a loss function to obtain a trained model; The spatial perception full connection layer maps the spatial distance through learnable position encoding, constructs a spatial modulation matrix to modulate the basic weight, and the specific steps are as follows: The spatial distance between the hidden layer neurons and the input features is calculated through the learnable position encoding, the spatial attenuation coefficient is combined, and the logarithmic transformation item of the local standard deviation is introduced for weighted adjustment to construct a spatial modulation matrix; the basic weight matrix and the spatial modulation matrix are multiplied element by element, the weighted influence of the curvature mutation factor vector expanded by self-splicing is superimposed to obtain a modulation weight matrix; The modulation weight matrix is used for linear transformation of the enhanced feature vector, the local standard deviation vector is projected and combined with a bias term as a dynamic offset, and the spatial perception feature is output through a ReLU activation function; The double-channel feature fusion module constructs a double-channel feature fusion structure, extracts overall trend features through a global channel, extracts significant mutation features through a local channel, and dynamically adjusts the original feature weight through a gating fusion mechanism, and the specific steps are as follows: The spatial perception feature is globally averaged and pooled to obtain a basic statistical feature, the basic statistical feature is spliced with the average of the local mean of all monitoring points, the maximum of the local standard deviation, and the average of the curvature mutation factor, and a projected global feature representation is obtained through linear projection; the spatial perception feature is globally maximized to obtain a local significant feature, the local significant feature is spliced with the minimum of the local mean, the maximum of the local standard deviation, and the maximum of the curvature mutation factor, and a projected local feature representation is obtained through linear projection; after the projected global feature representation and the projected local feature representation are spliced, the gating attention weight is generated through weight matrix mapping and a Sigmoid function, the spatial perception feature is dynamically modulated by using the gating attention weight, the modulation result is spliced with the double-channel feature to obtain a double-channel fusion feature; The uneven settlement prediction layer constructs a multi-task adaptive output module, uses the double-channel fusion feature, the local statistics, and the curvature mutation factor to generate an uneven settlement probability and a local anomaly heat map vector, and the specific steps are as follows: Based on the dual-channel fusion feature, combined with local mean, local standard deviation and curvature mutation factor, the multi-scale feature weight is calculated through the gating attention mechanism, and then the enhanced discriminant feature vector is obtained; the enhanced discriminant feature vector and the local anomaly factor are spliced, and are mapped to the hidden space through the multilayer perception, and the uneven settlement probability is calculated by combining the global statistics; the enhanced discriminant feature vector and the spatial modulation matrix are used to obtain the local anomaly heat map vector through the deconvolution layer and the spatial interpolation technology; S5, the real-time settlement monitoring data newly collected is input into the trained model after being processed through steps S2 and S3, and an intelligent detection result is obtained. 2.The intelligent detection method for uneven settlement of a super-high building foundation according to claim 1, characterized in that, Step S2 specifically includes: According to the spatial topological relationship between the monitoring points defined in advance according to the building foundation map, the adjacent point set of each monitoring point is determined; based on the adjacent point set of each monitoring point, the original settlement values of all monitoring points in the set are aggregated and calculated to obtain the local mean, and the local standard deviation is calculated based on the settlement value dispersion degree in the same adjacent point set; for each monitoring point, the original settlement value is centralized and scaled by using the local mean and the local standard deviation, and the standardized settlement value is obtained by enhancing the scaling factor constructed by the local coefficient of variation. 3.The intelligent detection method for uneven settlement of a super-high building foundation according to claim 2, characterized in that, Step S3 specifically includes: Based on the ratio of the variance of the settlement value in the adjacent point set of each monitoring point to the global maximum variance, combined with the ratio of the local standard deviation of the monitoring point to the global maximum local standard deviation, the local variance weight factor is obtained through the exponential amplification mechanism and the linear superposition operation; for the intermediate monitoring point, the second-order difference calculation is performed based on the original settlement values of the current point and the adjacent monitoring points before and after it, the calculation result is multiplied by the local variance weight factor to obtain the curvature mutation factor; for the boundary monitoring point, the unilateral difference calculation method is adopted, the calculation result is multiplied by the local variance weight factor to obtain the boundary curvature mutation factor; the curvature mutation factor vector and the standardized settlement feature vector compressed by the hyperbolic tangent function are multiplied element by element, and then the standardized settlement feature vector and the product result are spliced to obtain the enhanced feature vector.

4. The intelligent detection method for uneven settlement of a super-high building foundation according to claim 3, characterized in that, The loss function includes the overall uneven settlement probability loss and the local anomaly heat map loss, and the overall uneven settlement probability loss adopts the binary cross-entropy loss function; the local anomaly heat map loss uses the mean square error loss function.

5. The intelligent detection method for uneven settlement of a super-high building foundation according to claim 4, characterized in that, The total loss is the weighted sum of the overall uneven settlement probability loss and the local anomaly heat map loss.

6. An intelligent detection system for uneven settlement of a super high-rise building foundation, which executes the intelligent detection method for uneven settlement of a super high-rise building foundation according to claim 1, characterized in that, It includes: A data acquisition module for collecting settlement monitoring data through monitoring points arranged at key positions of a building foundation; A data standardization module for calculating local statistics by predefining the spatial topological relationship between monitoring points, and performing standardization processing on the original settlement data based on the adjacent point set to obtain standardized settlement values; A feature enhancement module for calculating the curvature mutation factor through second-order difference, amplifying the contribution of the mutation area by combining the local variance weight factor, splicing the standardized settlement feature vector and the curvature mutation factor to obtain an enhanced feature vector; The model construction module is configured to construct a non-uniform settlement detection model, including a spatial perception full connection layer, a double-channel feature fusion module and a non-uniform settlement prediction layer; and the training process of the model is optimized through a loss function, so as to obtain a trained model. The non-uniform settlement detection module is configured to input the processed real-time settlement monitoring data newly collected into the trained model, so as to obtain an intelligent detection result.

Images