Intelligent early warning system and method for dynamic settlement of soft soil foundation based on AI edge computing

Abstract

This invention discloses an intelligent early warning system and method for dynamic settlement of soft soil foundations based on AI edge computing, belonging to the field of soft soil foundation monitoring technology. The method acquires multi-dimensional dynamic monitoring data of soft soil foundations, constructs a first intelligent analytical model, a second intelligent decoupling model, and a third intelligent prediction model, obtains a set of key influencing parameters for foundation settlement, a set of settlement trend adaptation parameters, and generates dynamic settlement prediction curves and risk level coefficients, which are then processed in real time through AI edge computing nodes. This invention integrates multi-dimensional core monitoring data, combines a nonlinear rheological constitutive model of soft soil with feature extraction technology to remove parameter interference, and uses a fractional derivative prediction model to improve the reliability of settlement prediction. It relies on AI edge computing to achieve localized rapid processing and dual-model computation, ensuring continuous and efficient early warning, constructing a fully automated closed-loop system, and combining remote monitoring visualization and traceability functions to improve the level of intelligent management and control. The system has strong adaptability and is convenient to deploy and maintain.

Description

Technical Field

[0001] This invention relates to the field of soft soil foundation monitoring technology, specifically to an intelligent early warning system and method for dynamic settlement of soft soil foundations based on AI edge computing. Background Technology

[0002] In the construction and operation of soft soil foundation projects, foundation settlement is a core risk factor affecting project safety. Accurate monitoring, efficient prediction and timely early warning have become core industry requirements. With the expansion of project construction scale and the increasing complexity of geological conditions, traditional soft soil foundation settlement monitoring and early warning technologies have gradually exposed many limitations and are difficult to meet the intelligent management and control requirements of modern projects.

[0003] Chinese patent (publication number: CN116753905A) discloses an automated monitoring system and method for ultra-soft soil. This patent uses devices such as a graduated rod, intelligent monitor, GPS positioning chip, vacuum sensor, and pore water pressure gauge to automatically collect settlement data and parameters such as vacuum degree and pore water pressure. It processes the data with a computer terminal and draws settlement cloud maps. When the monitored values ​​reach the warning threshold, an early warning is issued. However, this technology mainly focuses on data collection and simple threshold warning. It does not effectively remove the coupled interference factors in the monitoring data. Moreover, it relies on centralized processing by a computer terminal. Under complex geological conditions, it is difficult to accurately capture key factors affecting settlement. The early warning response efficiency and prediction accuracy are limited, and it cannot achieve early prediction and dynamic early warning of settlement trends.

[0004] Chinese patent (publication number: CN119669790A) discloses an artificial intelligence-based method for predicting settlement in shield tunnels. This patent improves the accuracy of identifying and predicting complex settlement patterns by cleaning data, constructing an AI prediction model, and dynamically optimizing model parameters, providing an intelligent approach to settlement management. However, this technology is primarily designed for shield tunnel scenarios and does not incorporate the nonlinear rheological characteristics of soft soil to construct a suitable analytical and decoupled model. Furthermore, it does not employ an edge computing architecture, which can lead to data processing delays and discontinuous early warnings when the network is unstable or interrupted at the engineering site. This makes it difficult to adapt to the complex environment and real-time management requirements of soft soil foundation engineering sites.

[0005] Existing technologies still have room for improvement in terms of the accuracy of settlement monitoring in soft soil foundations, the adaptability of predictions, the real-time nature of early warnings, and the stability of the system. Therefore, there is an urgent need for an integrated intelligent early warning technology solution that combines multi-dimensional data analysis, intelligent decoupling, accurate prediction, and edge computing to meet the actual needs of safety management in soft soil foundation engineering. Summary of the Invention

[0006] To address the aforementioned technical issues, this application discloses an intelligent early warning method for dynamic settlement of soft soil foundations based on AI edge computing, specifically including:

[0007] Acquire multi-dimensional dynamic monitoring data of soft soil foundation, wherein the multi-dimensional dynamic monitoring data includes first core data and second core data;

[0008] A first intelligent analytical model is constructed based on the nonlinear rheological constitutive model of soft soil and the spatiotemporal sequence feature extraction algorithm. The first core data and the second core data are input into the first intelligent analytical model to obtain the set of key influencing parameters of foundation settlement.

[0009] The key influencing parameter set of foundation settlement is input into the second intelligent decoupling model constructed based on the attention mechanism, and the coupling interference components between parameters are stripped to obtain the settlement trend adaptation parameter set.

[0010] The set of settlement trend adaptation parameters is input into the third intelligent prediction model constructed based on fractional derivative theory, and combined with the preset settlement early warning threshold range, to generate a dynamic settlement prediction curve and risk level coefficient.

[0011] The dynamic settlement prediction curve and risk level coefficient are calculated in real time by AI edge computing nodes. When the risk level coefficient exceeds the preset safety threshold or the predicted settlement approaches the warning threshold, a graded warning instruction and foundation reinforcement control suggestion are generated.

[0012] The tiered early warning instructions are sent to the on-site early warning terminals in real time through the edge gateway. At the same time, dynamic monitoring data, prediction results and early warning information are uploaded to the remote monitoring platform to realize intelligent early warning and collaborative management of dynamic settlement of soft soil foundation.

[0013] Preferably, the first core data is the time series data of stress and strain in the deep soil of the foundation, which is collected by a distributed fiber optic grating sensor array; the second core data is the data on the distribution of pore water pressure and the dynamic change of groundwater level in the foundation, which is collected synchronously by a pore water pressure sensor and a groundwater level monitor.

[0014] Preferably, the construction of the first intelligent analytical model includes: based on the nonlinear rheological constitutive equation of soft soil:

[0015]

[0016] in, Instantaneous strain rate For effective stress, For instantaneous elastic modulus, , The viscosity coefficient, For elastic modulus, For time, As an integral variable, a spatiotemporal sequence feature extraction algorithm is used; data segments are divided by a sliding time window, and statistical and trend features of the data in each window are extracted to construct a multidimensional feature vector; kernel principal component analysis (KPCA) is used to reduce the dimensionality of the feature vector, and principal components whose cumulative contribution rate meets the preset requirements are retained to form a set of key influencing parameters of foundation settlement.

[0017] Preferably, the statistical features include mean, variance, peak factor, and kurtosis, and the trend features include the trend term coefficient. The dimension of the multidimensional feature vector is determined according to the number of feature types extracted. The kernel function of the kernel principal component analysis is the Gaussian kernel function. After the feature vector is mapped to a high-dimensional feature space through the kernel function, principal component extraction is performed.

[0018] Preferably, the second intelligent decoupling model is constructed based on an attention mechanism, specifically adopting a multi-head self-attention architecture. The model structure includes an embedding layer, a multi-head attention layer, a feedforward neural network layer, and an output layer. The embedding layer maps the set of key influencing parameters of foundation settlement into a fixed-dimensional embedding vector, as shown in the formula:

[0019]

[0020] in, For embedding vectors, This is the weight matrix. For the key influencing parameter set vector, The bias vector is used; the multi-head attention layer sets up multiple attention heads, and the attention weight of each attention head is calculated using the following formula:

[0021]

[0022] in, These are query, key, and value matrices, respectively. Key matrix transpose, Using the key vector dimension, the coupling interference components between parameters are accurately removed; the feedforward neural network layer uses a non-linear activation function to enhance the mapping of the decoupled features and output a set of parameters that adapt to the settling trend.

[0023] Preferably, the nonlinear activation function of the feedforward neural network layer is the ReLU activation function or the GELU activation function. The formula for the ReLU activation function is:

[0024]

[0025] The formula for the GELU activation function is:

[0026]

[0027] in, The input vector for the ReLU activation function is... The input vector for the GELU activation function is... , This is the weight matrix. , For bias vectors, The cumulative distribution function is the standard normal distribution. The activation function enhances the ability of the second intelligent decoupling model to fit nonlinear coupling relationships.

[0028] Preferably, the third intelligent prediction model is constructed based on fractional derivative theory, and the settlement prediction formula is as follows:

[0029]

[0030] Among them, for Predict settlement in real time. The order of the fractional derivative. For gamma function, To adapt the parameter set nonlinear mapping function, For integration variables, Indicates the integral variable Differential; nonlinear mapping function The system is trained through a deep learning network. The input is a set of parameters to adapt to the settlement trend, and the output is the instantaneous settlement rate. The preset settlement warning threshold range is divided into multiple levels, and the warning threshold for each level is determined according to the bearing capacity requirements of soft soil foundation and engineering safety standards.

[0031] Preferably, the risk level coefficient is calculated based on the matching degree between the early warning threshold range and the prediction result, using the following formula:

[0032]

[0033] in, This is the risk level coefficient. , These are the weighting coefficients. To predict the cumulative settlement, The highest level of warning cumulative settlement threshold, To predict the settlement rate, The highest level of warning is the settlement rate threshold; the risk level coefficient corresponds to different warning levels, and the corresponding level of warning is triggered according to the interval in which the risk level coefficient is located.

[0034] Preferably, the AI ​​edge computing node adopts a distributed architecture of edge gateway and edge server. The edge gateway is responsible for data preprocessing and command issuance, and the edge server deploys a first intelligent parsing model, a second intelligent decoupling model, and a third intelligent prediction model. Data preprocessing includes outlier removal and data standardization. Outlier removal adopts statistical criteria, and standardization adopts normalization processing. The edge computing node supports real-time computing and offline computing modes. When the network is interrupted, it can independently complete the generation and issuance of early warning commands. After the network is restored, the data is automatically synchronized to the remote monitoring platform.

[0035] The AI ​​edge computing-based intelligent early warning system for dynamic settlement of soft soil foundation includes a data acquisition module, an intelligent analysis module, an intelligent decoupling module, an intelligent prediction module, an edge computing module, an early warning and control module, and a remote monitoring module. These modules work together to achieve intelligent early warning for dynamic settlement of soft soil foundation.

[0036] The data acquisition module is used to collect time-series data of stress and strain in deep soil of the foundation, data of pore water pressure distribution and data of dynamic changes in groundwater level, and output multi-dimensional dynamic monitoring data.

[0037] The intelligent analysis module is used to receive multi-dimensional dynamic monitoring data output by the data acquisition module, and output a set of key influencing parameters of foundation settlement by fusing analysis and feature dimensionality reduction.

[0038] The intelligent decoupling module is used to receive the set of key influencing parameters of foundation settlement output by the intelligent analysis module, remove the coupling interference components between parameters, and output the settlement trend adaptation parameter set.

[0039] The intelligent prediction module is used to receive the settlement trend adaptation parameter set output by the intelligent decoupling module, and generate a dynamic settlement prediction curve and risk level coefficient by combining it with the preset settlement early warning threshold range.

[0040] The edge computing module is used to receive the dynamic settlement prediction curve and risk level coefficient output by the intelligent prediction module, judge the risk status through real-time calculation, and generate graded early warning instructions and foundation reinforcement control suggestions.

[0041] The early warning and control module is used to receive hierarchical early warning instructions issued by the edge computing module, output early warning information through sound and light alarms, visual displays, etc., and respond to foundation reinforcement control suggestions and link up to execute reinforcement operations.

[0042] The remote monitoring module is used to receive dynamic monitoring data, prediction results and early warning information synchronously uploaded by the edge computing module, and provides data storage, historical query, system configuration and data traceability functions to realize remote visual control of the settlement status of soft soil foundation.

[0043] Compared with the prior art, the technical solution of this application has the following technical effects:

[0044] This invention integrates multi-dimensional core monitoring data, utilizes a nonlinear rheological constitutive model for soft soil and spatiotemporal sequence feature extraction technology to accurately identify key influencing factors of foundation settlement, decouples the model through an attention mechanism to remove coupling interference between parameters, and combines a fractional derivative prediction model to fully match the rheological characteristics of soft soil, significantly improving the reliability of settlement trend prediction and providing accurate decision-making basis for engineering safety management.

[0045] This invention relies on the distributed architecture of AI edge computing nodes to realize localized processing of model calculation, risk assessment and instruction generation, which significantly shortens the early warning response cycle and can quickly respond to subsidence change scenarios. It also supports real-time and offline dual-mode computing, and can still independently complete early warning and data caching when the network is interrupted. It automatically synchronizes after the network is restored, ensuring that the early warning work is continuous and uninterrupted.

[0046] This invention constructs a fully automated system encompassing data acquisition, analysis, decoupling, prediction, early warning, and control. Through a hierarchical early warning mechanism linked with reinforcement and control recommendations, it achieves closed-loop management. The remote monitoring platform provides visualization, data storage, and traceability functions, facilitating real-time monitoring of settlement status by management personnel while ensuring data authenticity and comprehensively improving the intelligent level of soft soil foundation engineering management.

[0047] This invention system is adaptable to different types of soft soil geological conditions. The sensor deployment and early warning thresholds can be flexibly adjusted according to the actual project. The hardware and software are selected from industrial-grade equipment, which can adapt to the complex environment of the engineering site. It is convenient to deploy and maintain, and can be widely used in various soft soil foundation projects such as roads and buildings, providing reliable technical support for engineering safety.

[0048] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0049] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0051] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0052] Figure 1 A schematic diagram of the overall process of an intelligent early warning method for dynamic settlement of soft soil foundation based on AI edge computing;

[0053] Figure 2 First Intelligent Parsing Model Architecture Diagram;

[0054] Figure 3 Second intelligent decoupling model architecture diagram;

[0055] Figure 4 Third Intelligent Prediction Model Architecture Diagram;

[0056] Figure 5 Architecture diagram of an intelligent early warning system for dynamic settlement of soft soil foundation based on AI edge computing;

[0057] Figure 6 Comparison chart of dynamic settlement prediction curve and actual monitoring curve at monitoring point S1;

[0058] Figure 7 : Risk level coefficient change curve of monitoring point S2;

[0059] Figure 8 : Comparison of prediction errors before and after decoupling;

[0060] Figure 9 Histogram of computational delay distribution of edge computing nodes;

[0061] Figure 10 Screenshot of the settlement data visualization interface of the remote monitoring platform. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0063] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0064] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0065] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0066] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0067] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0068] Example 1

[0069] This embodiment mainly describes an intelligent early warning method for dynamic settlement of soft soil foundations based on AI edge computing, such as... Figure 1 As shown, it specifically includes:

[0070] Acquire multi-dimensional dynamic monitoring data of soft soil foundation, wherein the multi-dimensional dynamic monitoring data includes first core data and second core data;

[0071] A first intelligent analytical model is constructed based on the nonlinear rheological constitutive model of soft soil and the spatiotemporal sequence feature extraction algorithm. The first core data and the second core data are input into the first intelligent analytical model to obtain the set of key influencing parameters of foundation settlement.

[0072] The key influencing parameter set of foundation settlement is input into the second intelligent decoupling model constructed based on the attention mechanism, and the coupling interference components between parameters are stripped to obtain the settlement trend adaptation parameter set.

[0073] The set of settlement trend adaptation parameters is input into the third intelligent prediction model constructed based on fractional derivative theory, and combined with the preset settlement early warning threshold range, to generate a dynamic settlement prediction curve and risk level coefficient.

[0074] The dynamic settlement prediction curve and risk level coefficient are calculated in real time by AI edge computing nodes. When the risk level coefficient exceeds the preset safety threshold or the predicted settlement approaches the warning threshold, a graded warning instruction and foundation reinforcement control suggestion are generated.

[0075] The tiered early warning instructions are sent to the on-site early warning terminals in real time through the edge gateway. At the same time, dynamic monitoring data, prediction results and early warning information are uploaded to the remote monitoring platform to realize intelligent early warning and collaborative management of dynamic settlement of soft soil foundation.

[0076] Furthermore, the data acquisition module synchronously acquires the first core data (time-series data of stress and strain in deep foundation soil) and the second core data (data on the distribution of pore water pressure and dynamic changes in groundwater level). The first core data uses a distributed fiber Bragg grating sensor array, with strain-type fiber Bragg grating sensors, a wavelength range of 1525nm-1565nm, and a strain measurement range of ±2000 nm. Measurement accuracy ±1 To ensure the sensitivity and accuracy of data acquisition, sensors are deployed vertically along the depth direction of the soft soil foundation, covering the bearing layer and underlying soft soil layer. Sensor nodes are arranged at uniform intervals, with the spacing between adjacent nodes adjusted according to the uniformity of the foundation, ensuring that the spatial sampling density meets the analytical requirements. The second core data uses a vibrating wire pore water pressure sensor with a measurement range of 0-2MPa and a measurement accuracy of ±0.5%FS. The deployment depth corresponds one-to-one with the fiber optic grating sensor nodes. The groundwater level monitor uses an immersion-type liquid level sensor with a measurement range of 0-50m and a measurement accuracy of ±0.1%FS. It is deployed in a dedicated monitoring well around the foundation to collect groundwater level elevation data in real time and calculate the rate of water level change. The time synchronization of the pore water pressure sensor, groundwater level monitor, and fiber optic grating sensor is achieved through a GPS timing module, with a time error ≤1ms, ensuring the spatiotemporal consistency of the two types of core data.

[0077] Furthermore, such as Figure 2 As shown, the construction of the first intelligent analytical model is as follows: Based on the nonlinear rheological constitutive equation of soft soil: an improved three-element rheological constitutive equation is adopted: ,in, Instantaneous strain rate reflects the instantaneous deformation rate of soft soil under load. The effective stress is calculated by subtracting the pore water pressure from the total stress. , For the total stress, Pore ​​water pressure, The instantaneous elastic modulus was determined through an indoor triaxial test on soft soil. , The viscosity coefficients are all obtained through rheological experiments. For elastic modulus, For the duration of load application, As the integral variable, the collected stress and strain data are substituted into the equation, and the equation parameters are solved by numerical integration to quantify the rheological properties of soft soil.

[0078] Implementation of spatiotemporal sequence feature extraction: Sliding time window division: Adaptive sliding time window is adopted, with window length... The formula is: In the formula, To adjust the coefficient, The settling rate is used; the higher the settling rate, the shorter the window length, ensuring the capture of instantaneous changes. The window step size is a fraction of the window length. To avoid data omissions; Feature extraction: Extracting statistical and trend features of the data within each window, including the mean. , The number of data points within the window. For a single data point, variance Peak factor , Maximum value within the window, kurtosis Trend characteristics include trend term coefficients Fitting via linear regression get, The coefficient for the trend term. This is a constant term, representing the overall trend of data change;

[0079] Kernel Principal Component Analysis (KPCA) Dimensionality Reduction Implementation: Feature Vector Construction: The extracted statistical features and trend features are combined sequentially to construct a multidimensional feature vector. The kernel function used is the Gaussian kernel function. In the formula, The kernel parameters are determined using cross-validation. , The Euclidean distance between two eigenvectors; Dimensionality reduction process: Calculate the kernel matrix. The kernel matrix is ​​centered. , for Given an all-1 matrix, find the eigenvalues ​​of the centered kernel matrix. With feature vectors Sort by feature value from largest to smallest, and select the cumulative contribution rate. The former The principal components form a set of key influencing parameters for foundation settlement. It includes core parameters such as effective stress, rheological coefficient, pore water pressure gradient, and water level change rate.

[0080] Furthermore, statistical characteristics include: mean reflects the central tendency of the data; variance reflects the dispersion of the data; kurtosis reflects the influence of extreme values ​​in the data; and kurtosis reflects the steepness of the data distribution (kurtosis > 3 indicates a leptokurtic distribution, suggesting abrupt changes in data; kurtosis < 3 indicates a flat-peaked distribution, suggesting gradual data change). Trend characteristics include the trend term coefficient. A positive value indicates that the data is trending upward (e.g., increased strain, intensified settlement). A negative value indicates that the data is trending downwards. A value close to 0 indicates stable data. Feature vector dimension: determined by the number of feature types extracted. If 5 types of features are extracted (mean, variance, peak factor, kurtosis, trend coefficient), the feature vector dimension is 5-dimensional. If other features are added (such as skewness), the dimension increases accordingly.

[0081] Kernel parameter optimization in Gaussian kernel function: Optimization of kernel parameters using 5-fold cross-validation. The dataset was divided into 5 parts: 4 parts as the training set and 1 part as the test set. This process was repeated 5 times, and the test set with the smallest reconstruction error after dimensionality reduction was selected. Value; High-dimensional mapping: Transforming low-dimensional feature vectors using a Gaussian kernel function Mapping to a high-dimensional feature space It eliminates the need to explicitly compute high-dimensional space vectors; the inner product of high-dimensional space can be directly calculated through the kernel matrix. To avoid the curse of dimensionality; Principal component extraction: The principal components in the high-dimensional space correspond to the eigenvectors of the kernel matrix. Each principal component is a nonlinear combination of the original features, which can capture the nonlinear correlation information in the original data and retain the feature components that have the greatest impact on settlement.

[0082] Furthermore, such as Figure 3 As shown, the architecture and operational logic of the second intelligent decoupling model are implemented as follows: Model structure deployment: Embedded layer: This layer incorporates the key influencing parameters of foundation settlement. (dimension is) Mapped to a fixed-dimensional embedding vector The formula is: In the formula, For embedding vectors, the dimension is set to 64. for The weight matrix is ​​initialized using the Xavier initialization method. For the key influencing parameter set vector, The 64-dimensional bias vector is initialized to zero. The embedding layer maps the original parameters to a unified feature space, facilitating subsequent attention calculation. The multi-head attention layer uses eight attention heads, each independently calculating its attention weights to achieve multi-scale feature capture. The weight calculation formula for a single attention head is as follows: ,in, These are the query, key, and value matrices, each composed of an embedding vector. Obtained through linear transformation, , , , , , The weight matrix is ​​a linear transformation matrix with dimensions of 1. 64, Key matrix The transpose ensures dimension matching for matrix multiplication. The product is ), The dimension of the key vector. , divided by Normalization is used to prevent excessively large weight values ​​from saturating the softmax function. The softmax function maps the weight matrix to the [0,1] interval, with the sum of the weights being 1, thus focusing on important parameters. The feedforward neural network layer uses a two-layer fully connected network. The first layer has an output dimension of 256, and the second layer has an output dimension of 64. A non-linear activation function is used to enhance the mapping of decoupled features, improving the model's ability to fit complex non-linear relationships. The output layer maps the output vector of the feedforward neural network layer to a set of settling trend adaptation parameters. The dimensions and key impact parameter sets are consistent to ensure a one-to-one correspondence between the parameters;

[0083] Coupling interference stripping implementation: Coupling identification: The coupling strength between parameters is analyzed through the attention weight matrix. The larger the element value in the weight matrix, the stronger the correlation between the two corresponding parameters (the more serious the coupling); Stripping process: The parameters are weighted and reorganized based on the attention weight to highlight independent features and suppress coupling features. For example, for the coupling of stress and pore water pressure, the cross influence between the two is reduced by adjusting the weights to obtain independent parameters that only reflect their own effect on settlement.

[0084] Furthermore, the activation function types for the feedforward neural network layers are implemented as follows: The ReLU activation function formula is: The formula for the GELU activation function is:

[0085] In the formula, The input vector for the ReLU activation function is... The input vector for the GELU activation function is... , This is the weight matrix. , For bias vectors, The cumulative distribution function of the standard normal distribution. It can be approximated by the following formula: Fast computation is achieved by enhancing the second intelligent decoupling model's ability to fit nonlinear coupling relationships through activation functions. The activation function is selected based on engineering requirements: if computational speed is prioritized, the ReLU activation function is chosen; if decoupling accuracy is prioritized, the GELU activation function is chosen. During model training, the optimal activation function can be adaptively selected based on the validation set error.

[0086] Furthermore, such as Figure 4 As shown, the construction logic of the third intelligent prediction model is as follows: Settlement prediction formula: , among which, among which for Predict settlement in real time. The order of the fractional derivative, ranging from 0.5 to 1.0, was determined through soft soil rheological tests. The closer the value is to 1, the closer the soft soil is to an elastic body. The closer the value is to 0.5, the more pronounced the cohesive properties of the soft soil. For gamma function, For integers , !, for non-integers It was obtained through numerical calculation. To adapt the parameter set The nonlinear mapping function outputs the instantaneous settlement rate. For integration variables, Indicates the integral variable Differentiate the derivative and construct a continuous-time integral relationship; A deep learning network was constructed, consisting of an input layer (with dimensions consistent with the adaptation parameter set), three hidden layers (each with 128, 64, and 32 neurons respectively), and an output layer (one neuron outputting the instantaneous settlement rate). Training data: A training set was constructed using historical settlement monitoring data and corresponding parameter sets. The input was the settlement trend adaptation parameter set, and the output was the measured instantaneous settlement rate. Training process: The network parameters were optimized using the stochastic gradient descent (SGD) algorithm, with the mean squared error as the loss function. , These are measured values. The training iterations are 1000, and the initial learning rate is 0.001, adjusted exponentially.

[0087] Warning threshold range setting: Setting basis: determined according to the bearing capacity requirements of soft soil foundation, the allowable settlement of the superstructure (e.g., allowable settlement ≤30mm for road engineering, allowable settlement ≤20mm for building engineering), and engineering safety standards; Grade setting: warning threshold ranges are divided into multiple levels, with the first-level warning threshold (high risk): cumulative settlement ≥ or settling rate ≥ Level 2 warning threshold (medium risk): ≤Cumulative Settlement< or ≤Settlement rate< Level 3 warning threshold (low risk): ≤Cumulative Settlement< or ≤Settlement rate< ;in > > , > > The specific value is determined based on the type of project.

[0088] The calculation logic for the risk level coefficient is as follows: Core formula: , This is a risk level coefficient, ranging from 0 to 1, with values ​​closer to 1 indicating higher risk. , Let be the weighting coefficient, satisfying Adjustments can be made based on the level of importance placed on cumulative settlement and settlement rate in engineering projects, such as in road engineering. =0.6 (focusing on cumulative settlement) =0.4 (focusing on settlement rate); To predict the cumulative settlement (unit: m); The highest level of cumulative settlement threshold (i.e., the first-level warning cumulative settlement amount) ); To predict settlement rate (unit: m / s); The highest-level warning settlement rate threshold (i.e., the first-level warning settlement rate) );

[0089] Alert level matching:

[0090] Level 1 Warning (High Risk): If the value is ≥0.8, emergency reinforcement measures must be initiated immediately, and on-site work must be suspended for investigation.

[0091] Level 2 Warning (Medium Risk): 0.6≤ If the value is less than 0.8, routine reinforcement measures should be initiated and the monitoring frequency increased.

[0092] Level 3 Warning (Low Risk): 0.4≤ If the value is less than 0.6, the existing monitoring frequency should be maintained and the settlement trend should be closely monitored.

[0093] Safety status: <0.4, no reinforcement required, operate at normal monitoring frequency.

[0094] Furthermore, the architecture and functions of the AI ​​edge computing node are as follows: Edge Gateway: An industrial-grade edge gateway is selected, with a quad-core ARM Cortex-A53 processor, a main frequency ≥1.2GHz, memory ≥2GB, and storage ≥16GB. It supports Ethernet, Wi-Fi, and 4G / 5G multi-network access and has data preprocessing, protocol conversion, and command issuance functions; Edge Server: A rack-mounted edge server is selected, with an Intel Xeon E3 processor, a main frequency ≥3.5GHz, memory ≥16GB, and storage ≥512GB SSD. It deploys the first to third intelligent models, supports parallel computing, and has a computing latency ≤500ms; Connection Method: The edge gateway and edge server are connected via wired Ethernet with a transmission rate ≥1Gbps; the edge gateway is connected to field sensors and early warning terminals via a hybrid wired and wireless method to ensure connection stability.

[0095] Data preprocessing: Outlier removal: using Criteria for filtering collected data; if the data... satisfy: , If the value is less than the standard deviation, it is considered an outlier and linear interpolation is used. Supplement; Data standardization: A normalization process is used to map the data to the [0,1] interval. The formula is: , , These are the minimum and maximum values ​​of the data, determined through historical data statistics.

[0096] Operation mode switching: Real-time operation mode: When the network is normal, the edge server receives real-time data from the sensors, calls the model to perform calculations, generates early warning commands and prediction results, and uploads them synchronously to the remote monitoring platform; Offline operation mode: When the network is interrupted, the edge server automatically switches to offline mode, performs independent calculations based on locally stored historical data and real-time collected data, generates early warning commands and sends them to the on-site early warning terminals, and caches the data locally; After the network is restored, the cached data is automatically synchronized to the remote monitoring platform to ensure that the data is not lost.

[0097] Example 2

[0098] This embodiment mainly describes an intelligent early warning system for dynamic settlement of soft soil foundations based on AI edge computing, such as... Figure 5 As shown, it specifically includes:

[0099] The data acquisition module consists of a distributed fiber optic grating sensor array, a pore water pressure sensor, a groundwater level monitor, a data acquisition unit, and a GPS timing module. It acquires time-series data on stress and strain of deep foundation soil, pore water pressure distribution data, and dynamic changes in groundwater level at preset frequencies. It performs preliminary data filtering and time synchronization, outputting multi-dimensional dynamic monitoring data to the intelligent analysis module. The intelligent analysis module consists of a software module deploying the first intelligent analysis model, including a data receiving interface, a feature extraction unit, a KPCA dimensionality reduction unit, and a parameter output interface. It receives data from the data acquisition module through the data receiving interface, extracts spatiotemporal sequence features through the feature extraction unit, reduces the dimensionality of the feature vectors through KPCA, and outputs a set of key parameters affecting foundation settlement to the intelligent decoupling module. The decoupling module consists of a software module that deploys the second intelligent decoupling model, including a parameter receiving interface, an embedded layer operation unit, a multi-head attention operation unit, a feedforward neural network operation unit, and an adaptation parameter output interface. It receives the set of key influencing parameters, removes the coupling interference components between parameters through each operation unit, and outputs the settlement trend adaptation parameter set to the intelligent prediction module. The intelligent prediction module consists of a software module that deploys the third intelligent prediction model, including an adaptation parameter receiving interface, a nonlinear mapping operation unit, an integral operation unit, and a prediction result output interface. It receives the settlement trend adaptation parameter set, obtains the instantaneous settlement rate through the nonlinear mapping operation unit, solves the settlement prediction value through the integral operation unit, and generates a dynamic settlement prediction curve and risk level coefficient by combining the early warning threshold range, and outputs it to the edge computing module.

[0100] The edge computing module consists of: an AI edge computing node, an edge gateway, a computational logic unit, an instruction generation unit, and a data synchronization unit. It receives dynamic settlement prediction curves and risk level coefficients; the computational logic unit determines the risk status; the instruction generation unit generates tiered early warning instructions and foundation reinforcement control suggestions, which are then distributed to the early warning and control module via the edge gateway; and the data synchronization unit synchronizes and uploads relevant data to the remote monitoring module. The early warning and control module consists of: on-site early warning terminals (audio-visual alarm, LED display, mobile APP), and a reinforcement equipment controller. Upon receiving early warning instructions, the audio-visual alarm provides continuous alarm (red light and high-frequency sound) for Level 1 warnings, intermittent alarm (yellow light and medium-frequency sound) for Level 2 warnings, and only a light indicator (blue light) for Level 3 warnings. The LED display shows the warning level, risk level coefficient, predicted settlement data, and reinforcement suggestions in real time. The mobile terminal APP pushes early warning notifications to relevant personnel, and the reinforcement equipment controller responds to the reinforcement control suggestions, linking high-pressure jet grouting piles, cement mixing piles, and other reinforcement equipment to execute reinforcement operations according to the preset plan. The remote monitoring module consists of: a cloud monitoring platform including a data storage server, a visualization interface, a configuration management unit, and a data traceability unit. The data storage server adopts a distributed storage architecture to store dynamic monitoring data, prediction results, and early warning information for a storage time of ≥5 years. The visualization interface displays settlement trends, risk levels, and equipment status in the form of line charts, heat maps, and tables, and supports multi-dimensional data queries. The configuration management unit supports remote configuration of parameters such as early warning thresholds, model parameters, and sampling frequency. The data traceability unit adopts blockchain technology, with each data block containing a timestamp, data content, and signature information to ensure that the data is tamper-proof and supports historical data traceability queries.

[0101] Based on Examples 1 and 2, this example verifies the AI ​​edge computing-based intelligent early warning system and method for dynamic settlement of soft soil foundation proposed in this application. A highway subgrade project in a coastal soft soil area was selected as the test site. The soft soil thickness at this site is 12-18m, the natural water content is 42%-58%, and the void ratio is 1.2-1.6, which is a typical highly compressible soft soil. The test section is 200m long, 34m wide, and the subgrade filling height is 4.5m. The test period is 6 months. By deploying the intelligent early warning system described in this application, monitoring data is collected synchronously and compared with traditional monitoring methods (settlement plate combined with manual reading of pore water pressure gauge) to verify the monitoring accuracy, early warning response speed, and prediction accuracy of the system.

[0102] Three sets of distributed fiber Bragg grating sensor arrays (numbered S1, S2, and S3) were deployed at the test site. Each sensor array contained eight sensing nodes, with deployment depths of 2m, 4m, 6m, 8m, 10m, 12m, 15m, and 18m, respectively, and the spacing between adjacent nodes was 2-4m. Correspondingly, three sets of pore water pressure sensors (numbered P1, P2, and P3) and three groundwater level monitoring instruments (numbered W1, W2, and W3) were deployed, with each node depth corresponding to a fiber Bragg grating sensor node. The AI ​​edge computing nodes utilize industrial-grade edge gateways (model: Advantech EIS-D720) and edge servers (model: Dell PowerEdge T440). On-site early warning terminals include audible and visual alarms (model: LTE-1101J) and LED displays (size: 55 inches). The remote monitoring platform is deployed in the project management center's computer room, transmitting data via a 4G network. During the test, the system collected data at an adaptive sampling frequency: 5 minutes / time during the roadbed filling stage and 30 minutes / time after filling was completed. A total of 128,640 sets of valid data were collected, including 69,120 sets of the first core data (stress-strain time series data) and 59,520 sets of the second core data (pore water pressure + groundwater level data). Statistical and trend features were extracted using the first intelligent analytical model. After dimensionality reduction using KPCA, a set of key influencing parameters containing six core parameters, including effective stress, rheological coefficient, pore water pressure gradient, and water level change rate, was obtained. Then, the parameter coupling interference was removed using the second intelligent decoupling model. Finally, the dynamic settlement prediction curve and risk level coefficient were generated by the third intelligent prediction model.

[0103] The monitoring accuracy verification results show that the average relative error between the stress and strain data collected by this system and the traditional resistance strain gauge monitoring data is 2.3%, the average relative error between the pore water pressure data and the groundwater level data is 1.8%, and the average relative error between the groundwater level data and the traditional manual monitoring method (average relative error 5.7%-8.2%). Table 1 compares the settlement monitoring data at different depths of the three monitoring points in the third month of the experiment (one month after the completion of the filling). S1-S3 are the system monitoring values, and TS1-TS3 are the traditional settlement plate monitoring values. The unit is mm. As can be seen from Table 1, the maximum absolute error between the system monitoring values ​​and the traditional monitoring values ​​is 3.2 mm, and the minimum is 0.5 mm. The relative errors are all controlled within 3%, indicating that the system has high monitoring accuracy.

[0104] Table 1 Comparison of Settlement Monitoring Data at Different Depths

[0105] In verifying the accuracy of settlement prediction, monitoring data from months 1-4 of the experiment were selected as the training set, and monitoring data from months 5-6 were selected as the validation set. The difference between the system's predicted settlement and the actual monitored settlement was compared. At the end of month 6, the cumulative settlement predicted by the system was 89.3 mm, 87.1 mm, and 88.5 mm (corresponding to monitoring points S1-S3), while the actual monitored values ​​were 87.6 mm, 85.8 mm, and 87.2 mm, with relative errors of 1.9%, 1.5%, and 1.5%, respectively. In contrast, the relative error of the prediction value using the traditional empirical formula (stratified summation method) was 7.3%–9.5%. The dynamic settlement prediction curve of monitoring point S1 was compared with the actual monitoring curve, as shown below. Figure 6 As shown in the figure, the horizontal axis represents the monitoring time (days), and the vertical axis represents the cumulative settlement (mm). The three curves are the prediction curve of this system, the actual monitoring curve, and the prediction curve of the traditional stratified summation method, respectively. It can be clearly seen from the figure that the system prediction curve and the actual monitoring curve have a very high degree of fit and no significant deviation throughout the entire monitoring period. In contrast, the prediction curve of the traditional stratified summation method shows a significant deviation in the later stage, indicating that the settlement prediction accuracy of this system is far superior to that of the traditional method.

[0106] The early warning response speed was verified by simulating a sudden settlement scenario. On the 10th day of the 4th month of the experiment, the settlement rate of monitoring point S2 was suddenly increased from 1.2 mm / d to 3.5 mm / d (exceeding the secondary early warning threshold of 2.0 mm / d) through local loading. After detecting the sudden settlement rate change, the system completed data analysis, model calculation and risk level determination in only 0.3 seconds, and generated a secondary early warning command. 0.8 seconds later, the on-site audible and visual alarm was activated intermittently, and the LED display screen simultaneously displayed the early warning information and reinforcement suggestions. The remote monitoring platform also received the synchronous data within 1.2 seconds. In contrast, traditional manual monitoring methods require readings every 24 hours and cannot respond to sudden settlement changes in real time. The early warning response time exceeds 24 hours. The rapid response capability of this system can buy valuable time for emergency response in engineering projects.

[0107] In verifying the accuracy of risk level determination, based on the soft soil foundation bearing capacity requirements and engineering safety standards of the test site, three warning thresholds were set: Level 1 (cumulative settlement ≥ 100 mm or settlement rate ≥ 4.0 mm / d), Level 2 (80 mm ≤ cumulative settlement < 100 mm or 2.0 mm / d ≤ settlement rate < 4.0 mm / d), and Level 3 (60 mm ≤ cumulative settlement < 80 mm or 1.0 mm / d ≤ settlement rate < 2.0 mm / d). During the test, Level 3 warnings were triggered 3 times, Level 2 warnings were triggered 1 time, and Level 1 warnings did not occur. Each warning was consistent with the actual settlement state. The risk level coefficient change curve of monitoring point S2 is shown in the figure below. Figure 7As shown, the horizontal axis represents the monitoring time (days), and the vertical axis represents the risk level coefficient. The horizontal line in the figure represents the level 3 warning threshold (0.4) and the level 2 warning threshold (0.6). The curve touched the 0.4 threshold three times on the 45th, 68th and 92nd days, triggering the level 3 warning. On the 130th day, due to a sudden change in the settlement rate, it touched the 0.6 threshold, triggering the level 2 warning. All warning events were confirmed by on-site verification, and the settlement status matched the warning level 100%.

[0108] The effectiveness of parameter coupling decoupling was verified by comparing the prediction errors before and after decoupling. Without the second intelligent decoupling model, the average relative error of settlement prediction was 6.8%. After decoupling, the average relative error decreased to 1.7%, demonstrating a significant decoupling effect. A comparison of prediction errors before and after decoupling is shown in the figure below. Figure 8 As shown, the horizontal axis represents the monitoring point number (1-24 monitoring nodes), and the vertical axis represents the relative error (%). The bar charts show the error before and after decoupling. It can be seen from the figure that the prediction error of all monitoring nodes is significantly reduced after decoupling, and the error distribution is more concentrated. This indicates that the second intelligent decoupling model can effectively remove the coupling interference between parameters and improve the prediction accuracy.

[0109] The real-time and offline computing capabilities of the edge computing nodes were fully verified. Under normal network conditions, the processing time for a single set of data on the edge server was 0.21 seconds, meeting the requirements for real-time monitoring. Simulating a 24-hour network outage, the edge computing nodes independently completed data acquisition, model computation, and early warning command generation, storing a total of 2880 sets of offline data. After network recovery, all offline data was synchronized within 15 seconds, with no data loss or corruption. The histogram of the edge computing node's computational latency distribution is shown below. Figure 9 As shown in the figure, the horizontal axis represents the computation delay (seconds), and the vertical axis represents the number of data sets. It can be seen from the figure that the computation delay of more than 98% of the data is between 0.2 and 0.3 seconds, indicating good computational stability.

[0110] The data storage and traceability functions of the remote monitoring platform have been verified, demonstrating that the platform can completely store six months of monitoring data with a data query response time of ≤0.5 seconds. The early warning data traceability function, implemented through blockchain technology, can clearly trace the trigger time, data source, model calculation parameters, and other key information for each early warning event, ensuring data immutability. A screenshot of the settlement data visualization interface of the remote monitoring platform is shown below. Figure 10 As shown, the interface displays the distribution of monitoring points and real-time risk levels on the left, the dynamic settlement prediction curve in the middle, and the values ​​of key monitoring parameters and historical warning records on the right. The intuitive and clear visualization makes it easy for managers to keep track of the roadbed settlement status in real time.

[0111] In summary, the AI ​​edge computing-based intelligent early warning system and method for dynamic settlement of soft soil foundation described in this application demonstrates excellent performance in terms of monitoring accuracy, prediction accuracy, and early warning response speed. It can effectively solve the problems of low accuracy, slow response, and inaccurate prediction in traditional soft soil foundation settlement monitoring, and provide reliable technical support for the safety management and control of engineering construction in soft soil areas.

[0112] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A method for intelligent early warning of dynamic settlement of soft soil foundation based on AI edge computing, characterized in that, include: Acquire multi-dimensional dynamic monitoring data of soft soil foundation, wherein the multi-dimensional dynamic monitoring data includes first core data and second core data; A first intelligent analytical model is constructed based on the nonlinear rheological constitutive model of soft soil and the spatiotemporal sequence feature extraction algorithm. The first core data and the second core data are input into the first intelligent analytical model to obtain the set of key influencing parameters of foundation settlement. The key influencing parameter set of foundation settlement is input into the second intelligent decoupling model constructed based on the attention mechanism, and the coupling interference components between parameters are stripped to obtain the settlement trend adaptation parameter set. The set of settlement trend adaptation parameters is input into the third intelligent prediction model constructed based on fractional derivative theory, and combined with the preset settlement early warning threshold range, to generate a dynamic settlement prediction curve and risk level coefficient. The dynamic settlement prediction curve and risk level coefficient are calculated in real time by AI edge computing nodes. When the risk level coefficient exceeds the preset safety threshold or the predicted settlement approaches the warning threshold, a graded warning instruction and foundation reinforcement control suggestion are generated. The tiered early warning instructions are sent to the on-site early warning terminals in real time through the edge gateway. At the same time, dynamic monitoring data, prediction results and early warning information are uploaded to the remote monitoring platform to realize intelligent early warning and collaborative management of dynamic settlement of soft soil foundation.

2. The intelligent early warning system and method for dynamic settlement of soft soil foundation based on AI edge computing according to claim 1, characterized in that, The first core data is the time series data of stress and strain in the deep soil of the foundation, which is collected by a distributed fiber optic grating sensor array; the second core data is the data on the distribution of pore water pressure and the dynamic changes of groundwater level in the foundation, which is collected synchronously by a pore water pressure sensor and a groundwater level monitor.

3. The intelligent early warning method for dynamic settlement of soft soil foundation based on AI edge computing according to claim 1, characterized in that, The construction of the first intelligent analytical model includes: based on the nonlinear rheological constitutive equation of soft soil: in, Instantaneous strain rate For effective stress, For instantaneous elastic modulus, , The viscosity coefficient, For elastic modulus, For time, As an integral variable, a spatiotemporal sequence feature extraction algorithm is used; data segments are divided by a sliding time window, and statistical and trend features of the data in each window are extracted to construct a multidimensional feature vector; kernel principal component analysis (KPCA) is used to reduce the dimensionality of the feature vector, and principal components whose cumulative contribution rate meets the preset requirements are retained to form a set of key influencing parameters of foundation settlement.

4. The intelligent early warning method for dynamic settlement of soft soil foundation based on AI edge computing according to claim 3, characterized in that, The statistical features include mean, variance, peak factor, and kurtosis; the trend features include the trend term coefficient. The dimension of the multidimensional feature vector is determined according to the number of feature types extracted. The kernel function of the kernel principal component analysis is the Gaussian kernel function. After the feature vector is mapped to a high-dimensional feature space through the kernel function, principal component extraction is performed.

5. The intelligent early warning method for dynamic settlement of soft soil foundation based on AI edge computing according to claim 1, characterized in that, The second intelligent decoupling model is built based on an attention mechanism, specifically employing a multi-head self-attention architecture. The model structure includes an embedding layer, a multi-head attention layer, a feedforward neural network layer, and an output layer. The embedding layer maps the key influencing parameter set of foundation settlement into a fixed-dimensional embedding vector, as shown in the formula: in, For embedding vectors, This is the weight matrix. For the key influencing parameter set vector, The bias vector is used; the multi-head attention layer sets up multiple attention heads, and the attention weight of each attention head is calculated using the following formula: in, These are query, key, and value matrices, respectively. Key matrix transpose, Using the key vector dimension, the coupling interference components between parameters are accurately removed; the feedforward neural network layer uses a non-linear activation function to enhance the mapping of the decoupled features and output a set of parameters that adapt to the settling trend.

6. The intelligent early warning method for dynamic settlement of soft soil foundation based on AI edge computing according to claim 5, characterized in that, The nonlinear activation function of the feedforward neural network layer adopts either the ReLU activation function or the GELU activation function. The formula for the ReLU activation function is as follows: The formula for the GELU activation function is: in, The input vector for the ReLU activation function is... The input vector for the GELU activation function is... , This is the weight matrix. , For bias vectors, The cumulative distribution function is the standard normal distribution. The activation function enhances the ability of the second intelligent decoupling model to fit nonlinear coupling relationships.

7. The intelligent early warning method for dynamic settlement of soft soil foundation based on AI edge computing according to claim 1, characterized in that, The third intelligent prediction model is constructed based on fractional derivative theory, and the settlement prediction formula is as follows: Among them, for Predict settlement in real time. The order of the fractional derivative. For gamma function, To adapt the parameter set nonlinear mapping function, For integration variables, Indicates the integral variable Differential; nonlinear mapping function The data is obtained through deep learning network training. The input is a set of parameters for settlement trend adaptation, and the output is the instantaneous settlement rate. The preset settlement early warning threshold range is divided into multiple levels, and the early warning threshold of each level is determined according to the bearing requirements of soft soil foundation and engineering safety standards.

8. The intelligent early warning method for dynamic settlement of soft soil foundation based on AI edge computing according to claim 7, characterized in that, The risk level coefficient is calculated based on the matching degree between the early warning threshold range and the prediction result, using the following formula: in, Risk level coefficient , These are the weighting coefficients. To predict the cumulative settlement, The highest level of warning cumulative settlement threshold, To predict the settlement rate, The highest level of warning is the settlement rate threshold; the risk level coefficient corresponds to different warning levels, and the corresponding level of warning is triggered according to the interval in which the risk level coefficient is located.

9. The intelligent early warning method for dynamic settlement of soft soil foundation based on AI edge computing according to claim 1, characterized in that, The AI ​​edge computing node adopts a distributed architecture of edge gateway and edge server. The edge gateway is responsible for data preprocessing and command issuance, and the edge server deploys the first intelligent parsing model, the second intelligent decoupling model and the third intelligent prediction model. Data preprocessing includes outlier removal and data standardization. Outlier removal uses statistical criteria, and standardization uses normalization. Edge computing nodes support both real-time and offline computing modes. When the network is interrupted, they can independently generate and issue early warning commands, and automatically synchronize data to the remote monitoring platform after the network is restored.

10. A smart early warning system for dynamic settlement of soft soil foundation based on AI edge computing, used to implement the method of any one of claims 1-9, characterized in that, It includes a data acquisition module, an intelligent analysis module, an intelligent decoupling module, an intelligent prediction module, an edge computing module, an early warning and control module, and a remote monitoring module. These modules work together to achieve intelligent early warning of dynamic settlement of soft soil foundation. The data acquisition module is used to collect time-series data of stress and strain in deep soil of the foundation, data of pore water pressure distribution and data of dynamic changes in groundwater level, and output multi-dimensional dynamic monitoring data. The intelligent analysis module is used to receive multi-dimensional dynamic monitoring data output by the data acquisition module, and output a set of key influencing parameters of foundation settlement by fusing analysis and feature dimensionality reduction. The intelligent decoupling module is used to receive the set of key influencing parameters of foundation settlement output by the intelligent analysis module, remove the coupling interference components between parameters, and output the settlement trend adaptation parameter set. The intelligent prediction module is used to receive the settlement trend adaptation parameter set output by the intelligent decoupling module, and generate a dynamic settlement prediction curve and risk level coefficient by combining it with the preset settlement early warning threshold range. The edge computing module is used to receive the dynamic settlement prediction curve and risk level coefficient output by the intelligent prediction module, judge the risk status through real-time calculation, and generate graded early warning instructions and foundation reinforcement control suggestions. The early warning and control module is used to receive hierarchical early warning instructions issued by the edge computing module, output early warning information through sound and light alarms, visual displays, etc., and respond to foundation reinforcement control suggestions and execute reinforcement operations in conjunction with them. The remote monitoring module is used to receive dynamic monitoring data, prediction results and early warning information synchronously uploaded by the edge computing module, and provides data storage, historical query, system configuration and data traceability functions to realize remote visual management and control of the settlement status of soft soil foundation.

Images