Method for monitoring overload safety risk of construction load in real time

Abstract

The invention provides a construction load overload safety risk real-time monitoring method, which belongs to the technical field of construction safety, and comprises the following steps: arranging multiple types of sensors to form a three-dimensional monitoring network to collect real-time data, and extracting load response time-frequency characteristics through principal component analysis dimensionality reduction and variational mode decomposition; using particle swarm optimization to invert a support stiffness parameter and a pre-stress parameter, constructing a double-layer game model to calculate a load safety margin coefficient, determining a load state sudden change and triggering matrix reconstruction, dynamically adjusting a sensor sampling frequency according to the load safety margin coefficient, and determining a load state sudden change; the inversion convergence is judged through the Jacobian matrix spectral radius, the search step size is adaptively adjusted, the Kriging interpolation and the radial basis function neural network are fused to reconstruct the three-dimensional deformation field, and the technical problem that cooperative monitoring of parameter inversion and load state change of the foundation pit support structure cannot be achieved in the construction load overload state is solved.

Description

A method for real-time monitoring of safety risks from construction load overload Technical Field

[0001] This invention belongs to the field of construction safety technology, and more specifically, relates to a method for real-time monitoring of safety risks associated with construction load overload. Background Technology

[0002] In the field of foundation pit construction monitoring, traditional methods for monitoring the safety of retaining structures mainly rely on threshold alarm mechanisms based on displacement and stress. These methods collect monitoring data by deploying a single type of sensor and determine the structural safety status based on preset warning values. However, these traditional methods have significant drawbacks. Firstly, the use of a fixed sampling frequency leads to data redundancy or omission of crucial information. Secondly, alarm mechanisms based on static thresholds cannot accurately identify dynamic changes in load conditions and lack the ability to perform real-time inversion of support system parameters. In current foundation pit monitoring, traditional methods struggle to combine retaining structure parameter inversion with load state identification, making it impossible to dynamically adjust monitoring strategies based on load changes. This results in the inability to accurately obtain true structural parameter values ​​or promptly capture sudden load changes under overload conditions, impacting the accuracy and timeliness of safety risk assessments. In other words, existing technologies suffer from the technical problem of failing to achieve coordinated monitoring of parameter inversion and load state changes in foundation pit retaining structures under overload conditions. Summary of the Invention

[0003] In view of this, the present invention provides a method for real-time monitoring of safety risks caused by construction load overload, which can solve the technical problem in the prior art that the coordinating monitoring of parameter inversion and load state changes of the foundation pit retaining structure under construction load overload conditions cannot be achieved.

[0004] This invention is implemented as follows: It provides a method for real-time monitoring of construction load overload safety risks. Displacement sensors, tilt sensors, and strain sensors are uniformly deployed on the surface of the foundation pit retaining structure to form a three-dimensional spatial monitoring network, collecting real-time displacement, tilt, and strain data. Principal component analysis is performed on the collected data to reduce dimensionality and extract low-dimensional feature vectors. Variational mode decomposition is then used to decompose the low-dimensional feature vectors and extract intrinsic modal components. Hilbert transform is combined to calculate the instantaneous frequency and amplitude of each intrinsic modal component, constructing a load response time-frequency feature matrix. A particle swarm optimization model is then established. The support stiffness parameters and prestress parameters are obtained by inverting the parameter inversion model of the enclosure structure. A two-layer game optimization model is constructed to solve the load risk assessment and obtain the load safety margin coefficient. The Frobenius norm and its time change rate of the load response time-frequency characteristic matrix are calculated to determine the load state change and perform matrix reconstruction operation. The sensor sampling frequency is dynamically adjusted according to the load safety margin coefficient. An iterative convergence determination mechanism based on the Jacobian matrix spectral radius is constructed to adjust the search step size of the particle swarm optimization algorithm. The three-dimensional deformation field of the enclosure structure is reconstructed based on the Kriging interpolation method to generate a continuous deformation cloud map.

[0005] Principal component analysis dimensionality reduction refers to projecting the original high-dimensional monitoring data onto several orthogonal directions with the largest variance through linear transformation to form uncorrelated principal components, and selecting the top principal components with a cumulative variance contribution rate of 95% as the feature vectors after dimensionality reduction.

[0006] Variational mode decomposition is an adaptive signal decomposition technique that decomposes the original signal into several intrinsic mode components with finite bandwidth by constructing a variational problem.

[0007] The Hilbert transform is used to calculate the analytic signals of each intrinsic mode component to obtain the instantaneous frequency and instantaneous amplitude. The load response time-frequency characteristic matrix is ​​a two-dimensional matrix formed by arranging the instantaneous frequency and instantaneous amplitude of each intrinsic mode component according to the time sequence and frequency sequence.

[0008] Among them, the retaining structure parameter inversion model based on particle swarm optimization takes minimizing the root mean square error between measured displacement data and theoretical displacement data as the objective function, and adopts an adaptive adjustment strategy of inertial weight and chaotic mapping to initialize particle positions.

[0009] Among them, the adaptive adjustment strategy of inertia weight refers to the dynamic change of the inertia weight coefficient in the particle swarm algorithm with the number of iterations. In the early stage, a larger inertia weight is used to enhance the global search capability, and in the later stage, a smaller inertia weight is used to enhance the local fine search capability.

[0010] Among them, the chaotic mapping initialization utilizes the ergodicity and randomness of the chaotic sequence to initialize the particle positions, making the initial particle distribution more uniform, and uses the Logistic mapping to generate the chaotic sequence.

[0011] In the two-layer game optimization model, the upper-layer model aims to minimize the maximum displacement amplitude of the retaining structure, while the lower-layer model aims to maximize the uniformity of stress distribution in the support system. The two-layer model achieves Nash equilibrium by coupling the solution of the support stiffness parameters.

[0012] The load safety margin coefficient is defined as the ratio of the ultimate bearing capacity of the enclosure structure to the current load effect, which is calculated by combining real-time displacement data, tilt data, and strain data.

[0013] The Frobenius norm is the square root of the sum of the squares of all elements of the matrix. The formula for calculating the rate of change of the Frobenius norm is the difference between the Frobenius norm at the current time and the Frobenius norm at the previous time, divided by the Frobenius norm at the previous time.

[0014] Among them, when the Frobenius norm change rate exceeds 0.15, it is determined that the load state has changed significantly. The load response time-frequency characteristic matrix reconstruction operation refers to re-executing the variational mode decomposition and Hilbert transform process to update the intrinsic mode components and instantaneous frequency and instantaneous amplitude.

[0015] Specifically, when the load safety margin factor is within the safe range of 1.5 to 2.0, the sensor sampling frequency is kept constant at 10Hz. If the load safety margin factor is within the controllable range of 1.2 to 1.5, the sensor sampling frequency is adjusted to 20Hz.

[0016] The Jacobian matrix spectral radius is defined as the maximum absolute value of the matrix eigenvalues. When the spectral radius is less than 0.8, the prediction inversion algorithm will converge quickly. When the spectral radius is greater than 0.95, the search step size of the particle swarm optimization algorithm is adjusted to 70% of the original search step size.

[0017] Among them, the Kriging interpolation method establishes a covariance function by analyzing the spatial correlation between known points to perform interpolation estimation of unknown points, and incorporates the mechanical constraints of the finite element shape function into the construction of the covariance function of Kriging interpolation.

[0018] Specifically, the interpolation error is corrected by learning the interpolation residual between Kriging interpolation and the actual deformation field using a radial basis function neural network. The input of the radial basis function neural network is the three-dimensional coordinates of the monitoring point, and the output is the interpolation residual of the monitoring point.

[0019] Among them, the continuous deformation cloud map is a graphic that visualizes the three-dimensional deformation field of the reconstructed retaining structure in a color gradient manner. The horizontal axis represents the position coordinates of the retaining structure along the perimeter of the foundation pit, the vertical axis represents the position coordinates of the retaining structure along the depth direction, and the color value corresponds to the displacement magnitude.

[0020] This invention collects displacement, tilt, and strain data by constructing a multi-sensor 3D monitoring network. It extracts the time-frequency characteristics of the load response using principal component analysis for dimensionality reduction and variational mode decomposition, and combines this with a particle swarm optimization-based parameter inversion model to invert support stiffness and preload parameters in real time. This invention uses the Frobenius norm change rate of the load response time-frequency characteristic matrix as the criterion for determining load state abrupt changes. When the change rate exceeds a threshold, a matrix reconstruction operation is triggered. Simultaneously, the sensor sampling frequency is dynamically adjusted based on the inverted load safety margin coefficient, achieving real-time linkage between the parameter inversion process and load state monitoring. This invention uses the Jacobian matrix spectral radius to determine the convergence of the inversion algorithm and adaptively adjusts the search step size to ensure parameter inversion accuracy. Furthermore, it reconstructs the 3D deformation field based on the Kriging interpolation method and finite element mechanical constraints, comprehensively solving the technical problem of the inability to achieve coordinated monitoring of parameter inversion and load state changes in foundation pit retaining structures under construction load overload conditions. Attached Figure Description

[0021] Figure 1 is a flowchart of the method of the present invention.

[0022] Figure 2 shows the three-dimensional deformation field cloud map of the retaining structure under surcharge conditions.

[0023] Figure 3 is a time history curve of the load safety margin coefficient during the monitoring period. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0025] Figure 1 shows a flowchart of a method for real-time monitoring of construction load overload safety risks provided by the present invention. This method includes the following steps:

[0026] S01. Displacement sensors, tilt sensors and strain sensors are evenly distributed on the surface of the foundation pit retaining structure to form a three-dimensional spatial monitoring network, and real-time displacement data, tilt data and strain data of the retaining structure are collected. The sensor sampling frequency is set to 10Hz.

[0027] S02. Perform principal component analysis to reduce the dimensionality of the collected real-time displacement data, tilt data and strain data, extract the top principal components that retain 95% of the variance contribution rate, and convert the high-dimensional monitoring data into low-dimensional feature vectors.

[0028] S03. The low-dimensional eigenvector is decomposed using the variational mode decomposition method to extract the intrinsic mode components. The instantaneous frequency and instantaneous amplitude of each intrinsic mode component are calculated by combining the Hilbert transform to construct the load response time-frequency characteristic matrix.

[0029] S04. Establish a parameter inversion model for the retaining structure based on particle swarm optimization. The objective function is to minimize the root mean square error between the measured displacement data and the theoretical displacement data. The particle positions are initialized using an adaptive adjustment strategy of inertial weight and chaotic mapping. The support stiffness parameters and preload parameters are obtained through inversion.

[0030] S05. Construct a two-layer game optimization model for load risk assessment. The upper-layer model aims to minimize the maximum displacement amplitude of the retaining structure, while the lower-layer model aims to maximize the uniformity of stress distribution in the support system. The two-layer model are coupled through the support stiffness parameter to solve for the load safety margin coefficient under the current working condition.

[0031] S06. Calculate the Frobenius norm and its time rate of change of the load response time-frequency characteristic matrix. When the Frobenius norm rate of change exceeds 0.15, it is determined that the load state has changed significantly, and the load response time-frequency characteristic matrix reconstruction operation is performed.

[0032] S07. Extract the real-time value of the load safety margin coefficient. When the load safety margin coefficient is within the safe range of 1.5 to 2.0, maintain the current sensor sampling frequency of 10Hz and continuously monitor the changes in real-time displacement data, tilt data and strain data. If the load safety margin coefficient is within the controllable range of 1.2 to 1.5, adjust the sensor sampling frequency to 20Hz.

[0033] S08. Construct an iterative convergence determination mechanism based on the spectral radius of the Jacobian matrix. Calculate the spectral radius value of the Jacobian matrix during the iteration of the inversion model of the retaining structure parameters. When the spectral radius value is less than 0.8, the prediction inversion algorithm will converge quickly. When the spectral radius value is greater than 0.95, adjust the search step size of the particle swarm optimization algorithm to 70% of the original search step size.

[0034] S09. Reconstruct the three-dimensional deformation field of the enclosure structure based on the Kriging interpolation method, integrate the mechanical constraints of the finite element shape function into the covariance function of the Kriging interpolation, learn the interpolation residual using the radial basis function neural network, and generate a continuous deformation cloud map that satisfies the displacement compatibility condition using the real-time displacement data.

[0035] Principal component analysis (PCA) dimensionality reduction involves projecting the original high-dimensional monitoring data onto several orthogonal directions with the largest variance through linear transformation, forming uncorrelated principal components. Each principal component is a linear combination of the original variables. The top principal components with a cumulative variance contribution rate of 95% are selected as the feature vectors after dimensionality reduction, thereby compressing the data dimensionality while retaining the main information. The 95% variance contribution rate threshold was derived through statistical analysis of historical monitoring data from 12 actual foundation pit projects. Experiments show that this threshold can retain sufficient data features while effectively reducing the computational load. The 12 actual foundation pit projects include subway station foundation pits with excavation depths of 8 to 15 meters and commercial complex foundation pits, with a monitoring period of 3 to 6 months and a cumulative data sample collection of over 5 million sets.

[0036] Variational mode decomposition (VMD) is an adaptive signal decomposition technique that decomposes the original signal into several intrinsic mode components with finite bandwidth by constructing a variational problem. Each intrinsic mode component represents the oscillation mode of the signal within a certain frequency range. This method overcomes the mode aliasing problem existing in traditional empirical mode decomposition and can accurately extract each frequency component in non-stationary signals.

[0037] The Hilbert transform is used to calculate the analytic signals of each intrinsic mode component, thereby obtaining the instantaneous frequency and instantaneous amplitude. The instantaneous frequency reflects the frequency characteristics of the load excitation over time, and the instantaneous amplitude reflects the time-varying characteristics of the load magnitude.

[0038] The load response time-frequency characteristic matrix is ​​a two-dimensional matrix formed by arranging the instantaneous frequency and instantaneous amplitude of each intrinsic mode component according to the time sequence and frequency sequence. The rows of the matrix correspond to different times, the columns correspond to different frequency components, and the values ​​of the matrix elements represent the energy intensity of the frequency component at the given time. The load response time-frequency characteristic matrix comprehensively describes the time-domain and frequency-domain characteristics of the load response signal.

[0039] The inversion model of retaining structure parameters in particle swarm optimization treats the support stiffness parameters and prestress parameters to be inverted as decision variables in the optimization problem. Each particle represents a candidate solution in the parameter space. The particle updates its velocity and position based on its own historical best position and the global best position of the swarm, and gradually converges to the true value of the parameters.

[0040] The root mean square error function is used to measure the degree of deviation between the measured displacement and the theoretical displacement. The input includes the support stiffness parameter, the prestress parameter, and the measured displacement data. The output is the normalized root mean square error value. The function form is the square root of the sum of the squares of the differences between the measured displacement and the theoretical displacement at each measuring point, divided by the average value of the measured displacement. The theoretical displacement data is obtained through finite element numerical simulation. The finite element model adopts the Mohr-Coulomb constitutive model to describe the mechanical behavior of the soil. The mesh size is controlled between 0.5 and 1.0 m, and the boundary conditions are set to lateral constraint and vertical freedom.

[0041] The adaptive adjustment strategy of inertia weight refers to the dynamic change of the inertia weight coefficient in the particle swarm optimization algorithm with the number of iterations. In the early stage, a larger inertia weight is used to enhance the global search capability, and in the later stage, a smaller inertia weight is used to enhance the local fine search capability. The adjustment formula is that the inertia weight is equal to the maximum inertia weight minus the difference between the maximum and minimum inertia weight multiplied by the ratio of the current iteration number to the maximum iteration number. The maximum inertia weight is 0.9, the minimum inertia weight is 0.4, and the maximum number of iterations is set to 200.

[0042] Chaotic mapping initialization utilizes the ergodicity and randomness of chaotic sequences to initialize particle positions, making the initial particle distribution more uniform and avoiding getting trapped in local optima. Logistic mapping is used to generate chaotic sequences. The recursive formula for Logistic mapping is that the next generation chaotic variable is equal to the chaotic control parameter multiplied by the current chaotic variable multiplied by 1 minus the current chaotic variable. The chaotic control parameter is set to 4.0.

[0043] Through small-scale 1:20 scale model experiments on eight typical foundation pit projects, displacement response data under known parameter conditions were recorded. The parameter inversion error using this method was controlled within 5%, verifying the effectiveness of the algorithm. In the eight typical foundation pit project model experiments, the support stiffness parameter was set from 50 to 200 kN / mm, the prestressing parameter was set from 100 to 500 kN, the excavation depth was scaled to 0.4 to 0.75 m, and quartz sand was used to simulate the soil. The similarity ratio met the conditions of geometric similarity, stress similarity, and displacement similarity.

[0044] The objective function of the upper-level model in the two-layer game optimization model is expressed as follows: The function for minimizing the maximum displacement amplitude of the retaining structure is used to find the support arrangement scheme that minimizes the structural deformation. The input includes support stiffness parameters, support spacing and load distribution. The output is the normalized maximum displacement amplitude. The function form is the maximum displacement value of all monitoring points divided by the design allowable displacement value of the retaining structure. The design allowable displacement value of the retaining structure is determined according to 0.3% to 0.5% of the excavation depth of the foundation pit.

[0045] The objective function of the lower-level model is expressed as follows: The stress distribution uniformity maximization function of the support system is used to optimize the force balance of each support member. The input includes support stiffness parameters, preload parameters and load distribution. The output is the normalized stress distribution uniformity index. The function form is 1 minus the stress standard deviation of each support member divided by the stress average value. The stress standard deviation and stress average value are obtained by finite element calculation.

[0046] The two-layer model is coupled through the support stiffness parameter. The support stiffness parameter output by the upper-layer model serves as the input constraint of the lower-layer model. The optimization result of the lower-layer model is fed back to adjust the decision of the upper-layer model. The solution is iterated until the Nash equilibrium state is reached. The Nash equilibrium state is determined when the change of the objective function of the upper-layer model is less than 0.01 and the change of the objective function of the lower-layer model is less than 0.01 in three consecutive iterations.

[0047] The load safety margin coefficient is defined as the ratio of the ultimate bearing capacity of the enclosure structure to the current load effect, reflecting the safety reserve of the structure against overload failure. The larger the coefficient, the more sufficient the safety reserve. The current load effect is calculated by comprehensively considering the real-time displacement data, tilt data, and strain data. The ultimate bearing capacity of the enclosure structure is determined by calculating the material strength design value and the cross-sectional geometric parameters.

[0048] The safety range of 1.5 to 2.0 was determined through statistical analysis of construction loads and finite element numerical simulation of 35 completed foundation pit projects. Within this range, the structure is in an elastic working state and its safety meets the specifications. The 35 completed foundation pit projects cover sandy soil, silty soil and clay soil strata, with excavation depths ranging from 6 to 18 meters. The types of retaining structures include diaphragm walls, bored piles and SMW method piles.

[0049] The controllable range of 1.2 to 1.5 indicates that the structure has entered the elastoplastic stage but has not yet reached the failure state. It is necessary to increase the monitoring frequency and closely track it. The threshold is determined by analyzing the stress and strain monitoring data of 15 foundation pit projects under overload conditions. When the load safety margin coefficient drops below 1.2, plastic strain occurs in some areas of the retaining structure, but the overall structure remains stable.

[0050] The Frobenius norm is the square root of the sum of the squares of all elements of a matrix, used to measure the overall energy of the matrix. The rate of change of the Frobenius norm of the load response time-frequency characteristic matrix reflects the severity of the load state change. The formula for calculating the rate of change of the Frobenius norm is the difference between the Frobenius norm at the current time and the Frobenius norm at the previous time, divided by the Frobenius norm at the previous time.

[0051] The 0.15 rate of change threshold was obtained through statistical analysis of monitoring data from six foundation pit projects under load abrupt change conditions. When the rate of change exceeds the threshold, it indicates that the load distribution or magnitude has changed significantly, and the load response time-frequency characteristic matrix needs to be reconstructed to adapt to the new load state. The six foundation pit projects include three cases of surcharge abrupt change conditions, two cases of vehicle traffic conditions, and one case of pile driving conditions. The monitoring data acquisition time window is 30 minutes before and after the load change.

[0052] The load response time-frequency characteristic matrix reconstruction operation refers to re-executing the variational mode decomposition and Hilbert transform process, updating the intrinsic mode components and instantaneous frequencies and instantaneous amplitudes, and generating a new load response time-frequency characteristic matrix. The reconstruction operation uses real-time displacement data, tilt data and strain data within the last 60 seconds as input.

[0053] The Jacobian matrix is ​​a matrix composed of the partial derivatives of multivariate functions. In the iterative process of the inversion model of the building envelope parameters, the Jacobian matrix describes the sensitivity of the objective function to the decision variables. The spectral radius of the matrix is ​​defined as the maximum value of the absolute value of the matrix eigenvalues. When the spectral radius is less than 1, the iterative algorithm converges. The smaller the spectral radius, the faster the convergence speed.

[0054] The thresholds of 0.8 and 0.95 were determined through numerical experiments. When the spectral radius value is less than 0.8, the algorithm exhibits good convergence. When the spectral radius value is greater than 0.95, the convergence slows down and the search step size needs to be adjusted. The numerical experiments selected 20 sets of different initial parameters and load conditions for inversion calculations, and the relationship between the spectral radius value and the convergence algebra was statistically analyzed. The experimental results show that when the spectral radius value is less than 0.8, the average convergence algebra is 85 generations, and when the spectral radius value is greater than 0.95, the average convergence algebra exceeds 180 generations.

[0055] The search step size is the maximum distance a particle can move in the parameter space in the particle swarm optimization algorithm. The original search step size is initially set to 10% of the parameter value range. When the spectral radius value is greater than 0.95, it is adjusted to 70% of the original search step size to accelerate convergence.

[0056] Kriging interpolation is an optimal unbiased estimation method based on spatial statistics. It establishes a covariance function by analyzing the spatial correlation between known points and then performs interpolation estimation for unknown points.

[0057] The finite element shape function embodies the principles of displacement compatibility and mechanical equilibrium in structural mechanics. By incorporating these mechanical constraints into the construction of the covariance function of Kriging interpolation, the interpolation results are not only mathematically smooth but also physically reasonable. The mechanical constraints include the continuity constraint of the displacement gradient between adjacent nodes and the internal force equilibrium constraint.

[0058] Radial basis function neural networks are a type of feedforward neural network that uses radial basis functions as hidden layer activation functions. By training, they learn the interpolation residuals between Kriging interpolation and the actual deformation field, correcting interpolation errors and improving reconstruction accuracy. The radial basis function neural network includes an input layer, a hidden layer, and an output layer. The number of hidden layer nodes is set to 1.5 times the number of monitoring points. Gaussian radial basis functions are used as activation functions, and the training algorithm uses the least squares method.

[0059] The interpolation residual is defined as the difference between the displacement value calculated by Kriging interpolation and the measured displacement data. The radial basis function neural network takes the three-dimensional coordinates of the monitoring point as input and outputs the interpolation residual of the monitoring point.

[0060] The displacement compatibility condition requires that the displacements of adjacent nodes in a continuous deformation field satisfy geometric continuity. This is achieved by introducing displacement gradient constraints during the interpolation process. The displacement gradient constraint is expressed as the result of dividing the displacement difference between adjacent nodes by the node spacing being less than the maximum strain limit of the material.

[0061] The continuous deformation cloud map is a graphic representation of the reconstructed three-dimensional deformation field of the retaining structure using color gradients. The horizontal axis represents the position coordinates of the retaining structure along the perimeter of the foundation pit, and the vertical axis represents the position coordinates of the retaining structure along the depth direction. The color values ​​correspond to the displacement magnitude. The continuous deformation cloud map is used to intuitively determine the deformation distribution pattern of the retaining structure and identify dangerous areas.

[0062] The sensor sampling frequency adjustment mechanism dynamically adjusts the sensor sampling frequency based on the real-time value of the load safety margin coefficient. When the load safety margin coefficient is high, a lower frequency of 10Hz is used to save computing resources. When the load safety margin coefficient decreases to a controllable range, the frequency is increased to 20Hz to enhance monitoring sensitivity. The adjustment strategy was verified through comparative experiments on four foundation pit projects. The four foundation pit projects adopted three schemes: a fixed sampling frequency of 10Hz, a fixed sampling frequency of 20Hz, and a dynamically adjusted sampling frequency. The comparative experiment period was 30 days. The results showed that the dynamic adjustment scheme reduced data storage by 42% and shortened the calculation time by 35% while ensuring the monitoring effect.

[0063] The specific implementation methods of the above steps are described in detail below.

[0064] The specific implementation of step S01 involves arranging displacement sensors on the surface of the foundation pit retaining structure in a grid with a horizontal spacing of 3 to 5 meters and a vertical spacing of 2 to 3 meters. The displacement sensors use laser rangefinders or total stations to measure the three-dimensional displacement of the retaining structure surface relative to a reference point. The tilt sensors use high-precision inclinometers to measure the tilt angle changes of the retaining structure. The strain sensors use vibrating wire strain gauges or fiber optic strain sensors to measure the internal strain of the retaining structure. The sensor sampling frequency is set to 10Hz to ensure that it can capture the rapid change response caused by construction loads. All sensors are connected to the central monitoring system through a data acquisition unit to achieve real-time data transmission. The sensor placement locations are selected in areas sensitive to the deformation of the retaining structure, including the maximum excavation depth, the maximum support spacing, and corner locations. The three-dimensional spatial monitoring network achieves comprehensive monitoring of the displacement field, rotation angle, and stress state of the retaining structure through the combination of different types of sensors.

[0065] The specific implementation of step S02 involves arranging the collected real-time displacement data, tilt data, and strain data according to a time series to form an original monitoring data matrix. The rows of the matrix correspond to different times, and the columns correspond to different sensor measurement points. The original monitoring data matrix is ​​standardized to eliminate the differences in the dimensions of different physical quantities. The standardization method is to subtract the mean of the column containing each data point and then divide it by the standard deviation of the column. The covariance matrix of the standardized matrix is ​​calculated to describe the correlation between different measurement points. Eigenvalue decomposition is performed on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors. The magnitude of the eigenvalues ​​reflects the degree of contribution of the principal components to the data variance. The top few eigenvectors with a cumulative variance contribution rate of 95% are selected according to the eigenvalues ​​from largest to smallest as the principal component directions. The original monitoring data matrix is ​​projected onto the principal component directions to obtain low-dimensional eigenvectors. The purpose of principal component analysis dimensionality reduction is to reduce the data dimension and improve the efficiency of subsequent calculations while retaining the main features of the monitoring data. The reference value is that the first 5 to 8 principal components can usually meet the 95% variance contribution rate requirement.

[0066] The specific implementation of step S03 involves using a low-dimensional feature vector as the input signal, setting the number of modes in the variational mode decomposition to 4 to 6, and iteratively solving the variational problem using the alternating direction multiplier method to decompose the input signal into several intrinsic mode components. Each intrinsic mode component has narrowband characteristics and a different center frequency. A Hilbert transform is performed on each intrinsic mode component to obtain an analytic signal. The magnitude of the analytic signal is the instantaneous amplitude, and the derivative of the phase of the analytic signal with respect to time is the instantaneous frequency. The instantaneous frequencies and instantaneous amplitudes of each intrinsic mode component are organized into a two-dimensional matrix according to the time axis and the frequency axis to form the load response time-frequency feature matrix. Compared with the traditional Fourier transform, variational mode decomposition can adaptively extract the time-varying frequency components of non-stationary signals, which is suitable for the analysis of monitoring signals caused by dynamic changes in construction loads. The joint characterization of instantaneous frequency and instantaneous amplitude can simultaneously depict the frequency and energy characteristics of load changes, providing a time-frequency domain feature basis for subsequent load identification and risk assessment.

[0067] The specific implementation of step S04 involves setting the search range of the support stiffness parameter to 50 to 200 kN / mm, the search range of the preload parameter to 100 to 500 kN, initializing the particle swarm size to 30 to 50 particles, using Logistic chaotic mapping to generate a chaotic sequence between 0 and 1 to uniformly distribute the initial particle positions within the parameter space, and implementing the recursive calculation of the chaotic mapping through iterative formulas. A finite element model is established for each particle position representing the parameter combination to calculate theoretical displacement data. The root mean square error between the measured displacement data and the theoretical displacement data is calculated as the fitness value; the smaller the root mean square error, the closer the parameter combination is to the true value. Based on each... The historical best fitness and global best fitness of each particle are used to update the particle velocity and position. The velocity update considers three parts: inertia, cognition, and society. The inertia weight coefficient decreases linearly from 0.9 to 0.4 with the number of iterations to achieve global search in the early stage and local optimization in the later stage. The maximum number of iterations is set to 200 generations or the iteration is terminated when the global best fitness changes by less than 0.001 for 20 consecutive generations. The output global best particle position is the support stiffness parameter and preload parameter obtained by inversion. The particle swarm optimization algorithm achieves efficient search of the parameter space through swarm intelligence and avoids getting trapped in local optima by the traditional gradient method. Chaotic mapping initialization enhances the diversity of the initial population and improves the global search capability.

[0068] The specific implementation of step S05 is as follows: In the upper-level model, the objective is to minimize the maximum displacement amplitude of the enclosure structure. Input parameters include support stiffness parameters, support spacing, and load distribution. The displacement values ​​of all monitoring points are calculated using finite element analysis, and the maximum value is extracted and divided by the design allowable displacement value to obtain the normalized maximum displacement amplitude. In the lower-level model, the objective is to maximize the stress distribution uniformity of the support system. Input parameters include support stiffness parameters, prestressing parameters, and load distribution. The axial stress of each support member is calculated, and the stress standard deviation and average stress are statistically analyzed. The result of subtracting the stress standard deviation from 1 and dividing by the stress average is taken as the normalized stress distribution uniformity. The upper-level model uses the support stiffness parameters obtained from the upper-level model as constraints on the lower-level model. The prestress parameters obtained from the optimization of the lower-level model influence the displacement calculation of the upper-level model. A sequential solution strategy is adopted to alternately optimize the two-level models until the changes in the objective functions of the two levels are less than 0.01 in three consecutive iterations, reaching Nash equilibrium. Under the Nash equilibrium state, the ratio of the ultimate bearing capacity of the retaining structure to the current load effect is calculated to obtain the load safety margin coefficient. The two-level game optimization model achieves a balance between structural safety and the economy of the support system through the collaborative optimization of the upper and lower level objectives. The support stiffness parameters are used as coupling variables to connect the two-level models to form a closed-loop optimization process.

[0069] The specific implementation of step S06 involves extracting all elements of the load response time-frequency feature matrix, calculating the sum of squares, and then taking the square root to obtain the Frobenius norm. The Frobenius norm at the current moment and the Frobenius norm at the previous moment are recorded. The difference between the two is calculated and divided by the Frobenius norm at the previous moment to obtain the Frobenius norm change rate. It is then determined whether the Frobenius norm change rate exceeds the 0.15 threshold. If the change rate exceeds 0.15, a load response time-frequency feature matrix reconstruction operation is initiated. The reconstruction operation extracts real-time displacement data, tilt data, and strain data from the last 60 seconds and re-executes principal component analysis for dimensionality reduction, variational mode decomposition, and Hilbert transform to generate an updated load response time-frequency feature matrix. The Frobenius norm, as a measure of the overall energy of the matrix, can sensitively reflect the overall change in load state. The 0.15 threshold is determined statistically through monitoring data of typical load abrupt changes such as sudden load increases, vehicle traffic, and pile driving. When the change rate exceeds the threshold, it indicates a significant change in load distribution or load magnitude, requiring an update of the time-frequency features to ensure the accuracy of subsequent analysis.

[0070] The specific implementation of step S07 is to read the real-time value of the load safety margin coefficient output in step S05, and determine whether the real-time value is within the safe range of 1.5 to 2.0. If it is within the safe range, the sensor sampling frequency is kept at 10Hz and real-time displacement data, tilt data, and strain data are collected. If the load safety margin coefficient drops to the controllable range of 1.2 to 1.5, the sensor sampling frequency is adjusted to 20Hz through control commands. The dynamic adjustment of the sensor sampling frequency realizes the optimized allocation of monitoring resources according to the risk level reflected by the load safety margin coefficient. A lower sampling frequency is used within the safe range to save data storage space and computing resources. Increasing the sampling frequency within the controllable range enhances the ability to capture rapid changes in response. The thresholds of 1.5 and 1.2 correspond to the critical point of the structure transitioning from an elastic state to an elastic-plastic state and the critical point of entering plastic deformation, respectively.

[0071] The specific implementation of step S08 involves calculating the partial derivatives of the objective function with respect to the support stiffness parameters and prestress parameters in each iteration of the inversion model of the retaining structure parameters to construct a Jacobian matrix. Eigenvalue decomposition is performed on the Jacobian matrix to obtain all eigenvalues. The maximum absolute value of each eigenvalue is taken as the spectral radius value. The relationship between the spectral radius value and two thresholds, 0.8 and 0.95, is then determined. When the spectral radius value is less than 0.8, the algorithm is deemed to have fast convergence, and iteration continues at the current search step size. When the spectral radius value is greater than 0.95, the search step size of the particle swarm optimization algorithm is reduced to 70% of the original search step size to accelerate the convergence process. The spectral radius of the Jacobian matrix reflects the stability and convergence speed of the linearized system. A spectral radius less than 1 ensures iterative convergence, while a spectral radius closer to 1 results in slower convergence. By monitoring the spectral radius value and dynamically adjusting the search step size, the algorithm achieves adaptive optimization, improving inversion efficiency.

[0072] The specific implementation of step S09 involves extracting the three-dimensional coordinates and displacement values ​​of each monitoring point from the real-time displacement data, calculating the spatial distance between monitoring points to construct a distance matrix, selecting a Gaussian covariance function to describe spatial correlation and incorporating displacement compatibility conditions expressed by finite element shape functions, determining the covariance function parameters using the maximum likelihood estimation method, and using the Kriging interpolation method to perform weighted estimation of any unknown point on the surface of the retaining structure based on the displacement values ​​of surrounding monitoring points. The weighting coefficients are calculated from the covariance function and the distance matrix. A radial basis function neural network is constructed, with the three-dimensional coordinates of the unknown points as input and the interpolation residuals as output. The hidden layers of the neural network employ... The Gaussian radial basis function is used as the activation function, with the number of nodes set to 1.5 times the number of monitoring points. The difference between the Kriging interpolation value and the measured displacement data of the known monitoring points is used to train the neural network to learn the nonlinear residual law. The Kriging interpolation result is superimposed with the residual predicted by the neural network to obtain the corrected displacement value. All grid nodes on the surface of the retaining structure are traversed to complete the reconstruction of the three-dimensional deformation field and generate a continuous deformation cloud map. Kriging interpolation uses spatial statistics theory to achieve optimal unbiased estimation. Finite element shape functions are incorporated to ensure that the interpolation result meets mechanical constraints. The radial basis function neural network compensates for the systematic error of Kriging interpolation and improves the reconstruction accuracy.

[0073] It should be noted that the key technical ideas of this invention include a load response time-frequency feature extraction method based on variational mode decomposition and Hilbert transform, an adaptive inversion algorithm for retaining structure parameters based on particle swarm optimization and Jacobian matrix spectral radius monitoring, and a dynamic evaluation mechanism for load safety margin based on a two-level game model. Variational mode decomposition overcomes the limitations of traditional Fourier transform's assumption of stationary signals by adaptively decomposing non-stationary monitoring signals through a variational framework. Combined with Hilbert transform, it obtains instantaneous frequency and amplitude to achieve a joint time-frequency domain characterization of the load response. Compared with traditional single time-domain or frequency-domain analysis, it can accurately capture the frequency characteristics and energy evolution laws of dynamic changes in construction loads, providing richer feature information for risk identification under complex load conditions. The particle swarm optimization algorithm enhances global search capability and convergence speed through chaotic mapping initialization and adaptive adjustment of inertia weights. It introduces the Jacobian matrix spectral radius to monitor the convergence state of the inversion iteration process in real time and dynamically adjust the search step size, achieving adaptive optimization of parameter inversion and avoiding the inefficiency or local optima problems caused by traditional methods relying on empirical parameter settings. The two-level game model achieves force equilibrium of the support system by synergistically optimizing the safety objective of the upper structure and the uniformity objective of the lower support system, and by using the support stiffness parameter as a coupling variable to establish a closed-loop feedback mechanism. Under the premise of ensuring the deformation control of the enclosure structure, it can comprehensively consider multiple constraints to obtain more reasonable risk assessment results.

[0074] The synergistic effect of the above key technical approaches is reflected in the fact that time-frequency feature extraction provides high-quality input data for parameter inversion, the support stiffness and preload parameters obtained from parameter inversion provide an accurate mechanical parameter basis for the game optimization model, and the load safety margin coefficient output by the game model provides feedback control for monitoring frequency and trigger feature matrix reconstruction, forming a complete closed-loop monitoring system from signal processing to parameter identification to risk assessment. Compared with traditional discretized monitoring methods that rely only on a single threshold judgment, this invention achieves continuous tracking and dynamic early warning of construction load overload risk through multi-level technology integration, significantly improving the real-time performance, accuracy and adaptability of the monitoring system.

[0075] It should be noted that this invention also solves the following technical problems: In traditional foundation pit monitoring, fixed sampling frequencies lead to low data acquisition efficiency, high-frequency sampling causes data redundancy and computational burden, while low-frequency sampling may miss key information about load abrupt changes. This invention establishes a dynamic adjustment mechanism for sensor sampling frequency based on the load safety margin coefficient. When the load safety margin coefficient is within a safe range, a lower frequency is used to save computational resources; when the coefficient drops to a controllable range, the frequency is increased to enhance monitoring sensitivity, thus achieving optimal allocation of monitoring resources. Meanwhile, traditional parameter inversion methods lack a convergence prediction mechanism, and are prone to getting stuck in local optima or slow convergence when initial parameters are improperly selected or load conditions are complex. This invention introduces the Jacobian matrix spectral radius as a convergence criterion. When the spectral radius is small, the prediction algorithm converges quickly; when the spectral radius is large, the search step size is adaptively reduced to accelerate convergence. By dynamically adjusting the optimization strategy through real-time evaluation of the convergence characteristics of the inversion algorithm, the stability and computational efficiency of parameter inversion are ensured.

[0076] Specifically, the principle of this invention is as follows: The core principle of this invention in solving the above-mentioned technical problems lies in establishing a two-way coupling mechanism between parameter inversion and load state identification. First, the time-frequency characteristics of the load response signal are extracted through variational mode decomposition and Hilbert transform, transforming the load state change into a quantitative index of the norm of the time-frequency characteristic matrix. When the Frobenius norm change rate exceeds a threshold, it indicates a significant change in the load state. At this time, the recalculation of the parameter inversion model and the reconstruction of the time-frequency characteristic matrix are triggered, achieving a rapid response to load abrupt changes. Second, the particle swarm optimization parameter inversion model inverts the support stiffness and preload parameters by minimizing the root mean square error between the measured displacement and the theoretical displacement. The inversion result is input into a two-layer game optimization model to calculate the load safety margin coefficient. This coefficient serves as a control signal to dynamically adjust the sensor sampling frequency, forming a closed-loop control from parameter identification to monitoring strategy adjustment. The Jacobian matrix spectral radius is used to evaluate the convergence of the inversion algorithm in real time. When the spectral radius value is abnormal, the search step size is adaptively adjusted to ensure the stability and accuracy of parameter inversion. Thus, under overload conditions, the accurate acquisition of structural parameters and sensitive monitoring of load conditions can be achieved simultaneously.

[0077] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0078] The specific implementation of step S01 is to uniformly deploy displacement sensors, tilt sensors, and strain sensors on the surface of the foundation pit retaining structure according to the grid arrangement principle of horizontal spacing of 3 to 5m and vertical spacing of 2 to 4m. The displacement sensors use laser rangefinders to measure the three-dimensional displacement of the retaining structure surface relative to a fixed reference point. The tilt sensors use biaxial tiltmeters to measure the tilt angle of the retaining structure. The strain sensors use vibrating wire strain gauges to measure the strain value inside the retaining structure. The sensor sampling frequency is uniformly set to 10Hz. The data acquisition system transmits real-time monitoring data to the monitoring center through a wireless transmission module.

[0079] The specific implementation of step S02 involves performing principal component analysis to reduce the dimensionality of the collected real-time displacement data, tilt data, and strain data. First, the monitoring data is arranged according to the time series to form the original data matrix. The covariance matrix of principal component analysis The calculation formula is as follows:

[0080] ;

[0081] In the formula, The original data matrix has dimensions of . , The number of time sampling points, To monitor the number of variables; This is the mean matrix of the original data matrix, with dimensions [missing information]. ; Let be the covariance matrix with dimension . ; This represents the matrix transpose. For the covariance matrix... Eigenvalues ​​are obtained by performing eigenvalue decomposition. and the corresponding feature vector Sort the eigenvalues ​​from largest to smallest and calculate the cumulative variance contribution rate. The formula is as follows:

[0082] ;

[0083] In the formula, For the front The cumulative variance contribution rate of each principal component, dimensionless; For the first 1 eigenvalue, dimensionless; The eigenvalue index; The number of main components; To monitor the total number of variables, select those that meet the following criteria. The smallest The value corresponding to the previous The eigenvectors constitute the projection matrix. Projecting the original data matrix onto the principal component space yields low-dimensional feature vectors. The calculation formula is as follows:

[0084] ;

[0085] In the formula, It is a low-dimensional eigenvector matrix with dimension . ; Let be the projection matrix, with dimension . .

[0086] The specific implementation of step S03 involves using variational mode decomposition to decompose the low-dimensional eigenvectors, followed by Hilbert transform. The definition is as follows:

[0087] ;

[0088] In the formula, For function Hilbert transform; The variable is for integration, and the unit is seconds. The variable is time, and the unit is seconds; Let π be the mathematical constant pi. The constrained variational problem of variational mode decomposition is formulated as follows:

[0089] ;

[0090] In the formula, For the first One intrinsic mode component; The intrinsic modal component number; For the first The center frequency of each intrinsic mode component, in Hz; This represents the total number of intrinsic modal components, with an empirical value of 5 to 8. It is the Dirac function; This represents the convolution operation; Indicates time Find the partial derivative; The imaginary unit satisfies ; Represents the 2-norm; Let be the base of the natural logarithm. By introducing a quadratic penalty term and Lagrange multipliers, the constrained optimization problem is transformed into an unconstrained optimization problem. The alternating direction multiplier method is then used to iteratively solve for each eigenmode component. and center frequency For each intrinsic mode component Perform Hilbert transform to obtain analytic signal The calculation formula is as follows:

[0091] ;

[0092] In the formula, For the first The analytic signal of each intrinsic mode component. Instantaneous amplitude of the analytic signal. and instantaneous phase The calculation formula is as follows:

[0093] ;

[0094] ;

[0095] In the formula, For the first The instantaneous amplitude of each intrinsic mode component, in the same units as the original signal; For the first The instantaneous phase of each intrinsic mode component, in radians; The modulus of a complex number; It is the arctangent function. Instantaneous frequency. The calculation formula is obtained by taking the time derivative with respect to the instantaneous phase, as follows:

[0096] ;

[0097] In the formula, For the first The instantaneous frequency of each intrinsic mode component, in Hz; Indicates time Find the derivative. Construct the time-frequency characteristic matrix of the load response. Matrix elements Indicates the first At the [time]th moment The energy intensity of each frequency component is calculated using the following formula:

[0098] ;

[0099] In the formula, The load response time-frequency characteristic matrix is ​​the first... Line number The elements of the column have the same units as the original signal units, except that the units are in mm when the input is displacement data; For time sequence number; The frequency component number; For the first Each moment is measured in seconds. For the first A discrete frequency point, in Hz; Let Kronecker function be used when equal It takes the value 1 when it is true, and 0 otherwise. It is dimensionless.

[0100] The specific implementation of step S04 is to establish a parameter inversion model of the enclosure structure based on particle swarm optimization, and to incorporate the support stiffness parameters... and prestress parameters The parameters to be inverted are used as the objective function, which is the root mean square error between the measured displacement data and the theoretical displacement data. Minimize, the calculation formula is as follows:

[0101] ;

[0102] In the formula, The normalized root mean square error is dimensionless. Number of monitoring points; For monitoring point serial numbers; For the first The measured displacement values ​​of each monitoring point are in mm. For the first The theoretical displacement value of each monitoring point, in mm; This is the average measured displacement of all monitoring points, in mm. Support stiffness parameters, in kN / mm; These are prestressing parameters, in kN. Theoretical displacement value. Finite element numerical simulations show that the soil mechanical behavior is described using the Mohr-Coulomb constitutive model, and the expression for the Mohr-Coulomb yield criterion is as follows:

[0103] ;

[0104] In the formula, This is the yield function value, in kPa. The maximum principal stress is expressed in kPa. The minimum principal stress is expressed in kPa. These are parameters related to the internal friction angle, and are dimensionless. The internal friction angle of the soil is expressed in degrees. Soil cohesion, expressed in kPa; It is a sine function; It is the square root function. In the particle swarm optimization algorithm, the... The generation The velocity update formula for each particle is as follows:

[0105] ;

[0106] In the formula, For the first The generation The velocity vector of each particle; For iterative algebra; For particle serial numbers; For the first The generation The velocity vector of each particle; For the first The inertial weighting coefficient of the generation is dimensionless; and The learning factor is dimensionless and typically takes a value of 2.0. and A random number between 0 and 1, dimensionless; For the first The historical optimal position of each particle; The optimal position for the entire group; For the first The generation The position vector of each particle. The generation The formula for updating the position of each particle is as follows:

[0107] ;

[0108] In the formula, For the first The generation The position vector of each particle, containing the support stiffness parameter and prestress parameters Inertia weighting coefficient The adaptive adjustment formula is as follows:

[0109] ;

[0110] In the formula, The maximum inertia weight is 0.9, which is dimensionless. The minimum inertia weight is 0.4, which is dimensionless. The maximum number of iterations is set to 200. The chaotic mapping initialization uses a Logistic mapping to generate a chaotic sequence. The recursive formula is as follows:

[0111] ;

[0112] In the formula, For the first The chaotic variable in the next iteration has a value range of 0 to 1 and is dimensionless. The number of iterations for the chaotic sequence; The chaotic control parameter is 4.0, which is dimensionless. The chaotic sequence... Mapping to parameter space yields the initial particle positions. The mapping formula is as follows:

[0113] ;

[0114] In the formula, and These are the minimum and maximum values ​​of the parameter, respectively. For the first The chaotic sequence value corresponding to each particle is dimensionless.

[0115] The specific implementation of step S05 is to construct a two-layer game optimization model for load risk assessment. The objective function of the upper-layer model is to minimize the maximum displacement amplitude of the retaining structure, as expressed below:

[0116] ;

[0117] In the formula, The objective function value of the upper-level model is dimensionless. The support spacing is in meters (m). The allowable displacement value for the retaining structure is given in mm, based on the excavation depth of the foundation pit. The percentage is determined to be between 0.3% and 0.5%, of which This represents the excavation depth of the foundation pit, in meters (m). This indicates taking the maximum value over all monitoring points. The objective function of the lower-level model is to maximize the uniformity of stress distribution in the support system, expressed as follows:

[0118] ;

[0119] In the formula, This represents the objective function value of the lower-level model, which is dimensionless. This represents the standard deviation of the stress in each support member, in MPa. This represents the average stress of each support member, in MPa. The standard deviation of stress in the support members is also shown. The calculation formula is as follows:

[0120] ;

[0121] In the formula, The number of supporting members; For the support member serial number; For the first The stress values ​​of the support members, in MPa. Average stress of the support members. The calculation formula is as follows:

[0122] ;

[0123] In the formula, This represents the average stress of each support member, in MPa. The support stiffness parameters output from the upper-level model. As input constraints to the lower-level model, the upper-level model's objective function changes through iterative solutions until a Nash equilibrium is reached. and the change in the objective function of the lower-level model The calculation formula is as follows:

[0124] ;

[0125] ;

[0126] In the formula, For the first The change in the objective function of the upper-level model in the next iteration is dimensionless. For the first The change in the objective function of the lower-level model in the next iteration is dimensionless. and The first The objective function values ​​of the upper and lower model in the next iteration are dimensionless. and The first The objective function values ​​of the upper and lower model in the next iteration are dimensionless. This represents the number of iterations. This represents the absolute value. The Nash equilibrium criterion is the change in the objective function of the upper-level model over three consecutive iterations. and the change in the objective function of the lower-level model All are less than 0.01. Load safety margin factor The definition is as follows:

[0127] ;

[0128] In the formula, The load safety margin factor is dimensionless. The ultimate bearing capacity of the enclosure structure is expressed in kN. This represents the current load effect, expressed in kN. Current load effect The expression is calculated by combining real-time displacement data, tilt angle data, and strain data, as follows:

[0129] ;

[0130] In the formula, The reference load effect value, in kN, was determined through finite element numerical simulation, with an empirical value ranging from 100 to 500 kN. This is the average measured displacement, in mm; The value is a reference displacement in mm and is taken as the allowable displacement value for the enclosure structure design. This is the average measured tilt angle, in degrees; This is a reference tilt angle value, in degrees, with an empirical value of 0.5 degrees. This represents the average measured strain, expressed in microstrain. The strain value is for reference only, and the unit is microstrain. The empirical value is 1000 microstrain. , , The weighting coefficients are dimensionless and satisfy the following conditions: The empirical values ​​are 0.5, 0.3, and 0.2 respectively.

[0131] The specific implementation of step S06 is to calculate the time-frequency characteristic matrix of the load response. Frobenius norm The calculation formula is as follows:

[0132] ;

[0133] In the formula, is the Frobenius norm of the time-frequency characteristic matrix of the load response, with the same units as the matrix elements. When the input is displacement data, the unit is mm. This represents the number of discrete points in time. The number of discrete points in the frequency domain. The Frobenius norm rate of change over time. The calculation formula is as follows:

[0134] ;

[0135] In the formula, The rate of change of the Frobenius norm is dimensionless. For the current moment The Frobenius norm; The current time is expressed in seconds. For the previous moment The Frobenius norm; The previous moment, in seconds; time interval The default value is 10 seconds. When If a significant change in load state is detected, a load response time-frequency characteristic matrix reconstruction operation is performed. The variational mode decomposition and Hilbert transform processes are re-executed using real-time displacement data, tilt data, and strain data from the last 60 seconds as input.

[0136] The specific implementation method of step S07 is to extract the load safety margin factor. The real-time value, when Maintain sensor sampling frequency To maintain a constant 10Hz frequency and continuously monitor changes in real-time displacement, tilt, and strain data, if The sensor sampling frequency will be used at this time. The dynamic adjustment strategy expression for the sensor sampling frequency, adjusted to 20Hz, is as follows:

[0137] ;

[0138] In the formula, This is the sensor sampling frequency, measured in Hz.

[0139] The specific implementation of step S08 involves constructing an iterative convergence determination mechanism based on the spectral radius of the Jacobian matrix. During the iteration process of the retaining structure parameter inversion model, the Jacobian matrix... The formula for calculating the elements is as follows:

[0140] ;

[0141] In the formula, For the Jacobian matrix, the first... Line number The elements of the column are dimensionless; For decision variable index; For column indexes; For the first The decision variables include the support stiffness parameters. and prestress parameters ; Indicates the decision variable Find the partial derivatives. Find the spectral radius of the Jacobian matrix. Defined as the maximum absolute value of the matrix's eigenvalues, the expression is as follows:

[0142] ;

[0143] In the formula, Let be the spectral radius of the Jacobian matrix, which is dimensionless; Jacobian matrix The One eigenvalue, dimensionless; For feature value index; This indicates taking the maximum value among all eigenvalues. When... The time-prediction inversion algorithm will converge quickly, when Adjusting the search step size of the particle swarm optimization algorithm Adjust the search step size to 70% of the original step size using the following formula:

[0144] ;

[0145] In the formula, This is the adjusted search step size; This is the original search step size, initially set to 10% of the parameter value range.

[0146] The specific implementation of step S09 is based on reconstructing the three-dimensional deformation field of the enclosure structure using the Kriging interpolation method, with the initial covariance function... An exponential covariance model is used, and the expression is as follows:

[0147] ;

[0148] In the formula, The initial covariance function value, in units of ; and Spatial location coordinates; This is the variance parameter, in units of... The variance of the measured displacement data is estimated. The relevant distance parameter, in meters, represents the characteristic length of spatial correlation, with an empirical value of 2 to 5 meters. For position and The Euclidean distance between them, in meters; It is an exponential function. Kriging interpolation estimate. The expression is as follows:

[0149] ;

[0150] In the formula, For position The Kriging interpolation estimates the displacement at the location, in mm; The spatial coordinates of the point to be interpolated; For the first The weighting coefficients for each monitoring point are dimensionless. By solving the Kriging equations, we determine that the Kriging equations are expressed as follows:

[0151] ;

[0152] In the formula, For monitoring points and The covariance function values ​​between them, in units of ; These are Lagrange multipliers, with units of 1. ; For monitoring points and interpolation points The covariance function values ​​between them, in units of Displacement gradient vector The expression, calculated using the finite difference method, is as follows:

[0153] ;

[0154] In the formula, For position The displacement gradient vector at the location; , , Displacement at , , Partial derivative in direction, in mm / m; , , They are respectively three-dimensional spatial coordinate systems , , The coordinate axis directions are in meters. The mechanical constraints of the finite element shape function are incorporated into the construction of the covariance function, resulting in the modified covariance function. The expression is as follows:

[0155] ;

[0156] In the formula, The corrected covariance function value, in units of ; This is the mechanical constraint coefficient, with an empirical value of 0.1 to 0.3, and is dimensionless. This represents the inner product of the displacement gradient vectors, in units of 1. ; This is a reference length, in meters, and is taken as the depth of the foundation pit excavation. Interpolation residuals are learned using radial basis function neural networks. The interpolation residual is defined as the difference between the displacement value calculated by Kriging interpolation and the measured displacement data, and the expression is as follows:

[0157] ;

[0158] In the formula, For the first The interpolation residuals for each monitoring point are expressed in mm. The output of the radial basis function neural network. The expression is as follows:

[0159] ;

[0160] In the formula, For position The interpolation residual prediction value at the location, in mm; The number of hidden layer nodes is set to 1.5 times the number of monitoring points; For the first The output weights of each hidden layer node; These are Gaussian radial basis functions; For the first The center vector of each hidden layer node; This represents the Euclidean norm. Gaussian radial basis functions. The expression is as follows:

[0161] ;

[0162] In the formula, The Euclidean distance between the input vector and the center vector; The width parameter of the radial basis function is typically taken as 0.5 to 2.0. The final reconstructed 3D deformation field displacement values... The sum of the Kriging interpolation value and the neural network prediction residual is expressed as follows:

[0163] ;

[0164] In the formula, For position The reconstructed displacement value is given in mm. A continuous deformation contour map satisfying the displacement compatibility condition is generated using real-time displacement data. The displacement compatibility condition constrains the continuity of the displacement gradient between adjacent nodes, and the constraint expression is as follows:

[0165] ;

[0166] In the formula, and The coordinates of the adjacent nodes; This is the maximum strain limit for the material, typically ranging from 0.002 to 0.003 for concrete.

[0167] To better understand and implement this invention, the following is a specific application scenario example 2: A technical team undertook a construction monitoring task for the retaining structure of a subway station foundation pit. The excavation depth of the pit was 12.5m, and the retaining structure adopted an underground continuous wall support system with a wall thickness of 0.8m. Three concrete supports were installed, with support spacing of 4.5m, 4.0m, and 4.0m respectively. The surrounding strata of the foundation pit were mainly silty clay and fine sand layers, with a groundwater level of 3.2m. The construction site was only 8m away from the existing building, and one side of the foundation pit was adjacent to a main urban road. The construction load was complex and variable, and traditional manual inspection methods could not capture the dynamic response of the retaining structure in real time. Therefore, the technical team decided to adopt the real-time monitoring method for construction load overload safety risks of this invention.

[0168] The technical team deployed a row of sensors every 2.5 meters along the depth direction outside the diaphragm wall, and a monitoring section every 6 meters around the perimeter of the excavation pit, for a total of 24 monitoring sections. Each section was equipped with 5 displacement sensors, 5 tilt sensors, and 8 strain sensors, forming a three-dimensional spatial monitoring network covering the entire retaining structure. The sampling frequency of all sensors was initially set to 10Hz, and displacement, tilt, and strain data of the retaining structure were collected in real time. When the excavation pit reached 2 meters below the second support, the system continuously monitored for 72 hours, accumulating approximately 3.1 million sets of data from various sensors.

[0169] The technical team performed principal component analysis (PCA) to reduce the dimensionality of the massive monitoring data collected. The original monitoring data had a dimension of 432. By calculating the eigenvalues ​​of the covariance matrix, it was found that the cumulative variance contribution rate of the first 18 principal components reached 96.2%, meeting the set threshold requirement of 95%. The system automatically extracted these 18 principal components as eigenvectors, compressing the data dimension from 432 to 18, significantly reducing the subsequent computational load. Subsequently, variational mode decomposition (VMD) was used to decompose the low-dimensional eigenvectors. The number of decomposition modes was set to 6, resulting in 6 intrinsic mode components, each corresponding to an oscillation mode in a different frequency range. The instantaneous frequency and instantaneous amplitude of each intrinsic mode component were calculated using Hilbert transform to construct the load response time-frequency feature matrix. The matrix dimension is 7200×6, where 7200 corresponds to the number of sampling points per second over 72 hours.

[0170] The technical team established a parameter inversion model for the retaining structure based on particle swarm optimization, setting the particle swarm size to 50 particles and the maximum number of iterations to 200. Logistic chaotic mapping was used to initialize particle positions, with the chaos control parameter set to 4.0, and the generated chaotic sequence traversed the entire parameter space. The initial values ​​for the support stiffness parameters to be inverted were set to a range of 80 to 180 kN / mm, and the initial values ​​for the prestressing parameters were set to a range of 200 to 450 kN. The inversion model uses the minimization of the root mean square error between measured displacement data and finite element theoretical displacement data as its objective function. The theoretical displacement data was calculated by establishing a three-dimensional finite element model of the foundation pit excavation. The model uses the Mohr-Coulomb constitutive model to describe the soil, with a mesh size controlled at 0.6 m. An adaptive adjustment strategy was adopted for the inertia weight, with a maximum inertia weight of 0.9 and a minimum inertia weight of 0.4, decreasing linearly with the number of iterations. After 158 iterations of calculation, the algorithm converged, and the inversion yielded the stiffness parameters of the first support being 132 kN / mm, the second support being 145 kN / mm, and the third support being 128 kN / mm. The preload parameters for the three supports were 285 kN, 320 kN, and 298 kN, respectively.

[0171] The technical team constructed a two-layer game-theoretic optimization model for load risk assessment. The upper-layer model aims to minimize the maximum displacement of the retaining structure, with the allowable displacement value for the foundation pit determined to be 50mm, representing 0.4% of the excavation depth of 12.5m. The lower-layer model aims to maximize the uniformity of stress distribution in the support system, obtaining stress data for each support member through finite element analysis. The two models are coupled through support stiffness parameters; during the iterative solution process, the support stiffness output by the upper-layer model serves as the input constraint for the lower-layer model, and the optimization results of the lower-layer model are fed back to adjust the decisions of the upper-layer model. After 11 iterations, a Nash equilibrium was reached, at which point the change in the objective function of the upper-layer model was 0.008, and the change in the objective function of the lower-layer model was 0.009, both less than the convergence threshold of 0.01. The final calculated load safety margin coefficient under the current working condition is 1.78, within the safe range of 1.5 to 2.0. The system maintains a constant sensor sampling frequency of 10Hz and continues monitoring.

[0172] On the fifth day after the foundation pit was excavated to the designed depth, a sudden increase in surface load occurred on the north side of the pit due to the temporary storage of building materials. The storage area was approximately 15m × 8m, with a storage height of approximately 2.5m and an estimated weight of 380kN. The monitoring system immediately detected a significant change in the response signal of the retaining structure. The Frobenius norm of the load response time-frequency characteristic matrix jumped from 184.6 to 223.8 within 10 minutes, with a calculated Frobenius norm change rate of 0.212, exceeding the preset threshold of 0.15. The system automatically determined that the load state had changed significantly and performed a load response time-frequency characteristic matrix reconstruction operation. Using real-time monitoring data from the last 60 seconds, variational mode decomposition and Hilbert transform were performed again to update the intrinsic mode components, instantaneous frequencies, and instantaneous amplitudes, generating a new load response time-frequency characteristic matrix. As shown in Figure 2, a significant concentration zone of displacement increment appeared on the north side of the retaining structure, with a maximum displacement of 38mm, an increase of 15mm compared to before the loading.

[0173] The technical team re-executed the parameter inversion and risk assessment process, with the system calculating the spectral radius of the Jacobian matrix in real time during the inversion model iteration. In the first 80 iterations, the spectral radius remained between 0.72 and 0.85, indicating rapid convergence of the inversion algorithm. However, in the 92nd iteration, the spectral radius suddenly rose to 0.97, and the system automatically determined that convergence had slowed. It immediately adjusted the search step size of the particle swarm optimization algorithm from 10% to 7% of the original parameter range; that is, the search step size for the support stiffness parameter decreased from 10 kN / mm to 7 kN / mm, and the search step size for the prestressing parameter decreased from 25 kN to 17.5 kN. After this adjustment, the algorithm converged in the 128th iteration, revealing changes in the support stiffness and prestressing parameters under the surcharge condition. The two-layer game optimization model yielded a new load safety margin coefficient of 1.38, falling within the controllable range of 1.2 to 1.5. The system automatically increased the sensor sampling frequency from 10 Hz to 20 Hz to enhance monitoring sensitivity. As shown in Figure 3, the load safety margin factor decreased rapidly after the surcharge occurred, and gradually recovered after adjustment measures were taken. The changes in key parameters monitored at different time periods are shown in Table 1.

[0174] Table 1 Comparison of monitoring parameters of the enclosure structure under different working conditions

[0175]

[0176] The technical team reconstructed the three-dimensional deformation field of the retaining structure based on the Kriging interpolation method, integrating the displacement gradient continuity constraint and internal force equilibrium constraint of the finite element shape function into the construction of the covariance function of the Kriging interpolation. A radial basis function neural network was used to learn the interpolation residuals. The number of hidden layer nodes in the neural network was set to 1.5 times the number of monitoring points (432), i.e., 648 nodes. Gaussian radial basis functions were used as activation functions, and the network was trained using the least squares method. After training, the neural network output the interpolation residuals of each monitoring point, corrected the Kriging interpolation results, and generated a continuous deformation cloud map that satisfied the displacement compatibility conditions. The deformation cloud map clearly showed the displacement distribution pattern of the retaining structure along the depth direction and the perimeter direction. The deformation of the retaining structure corresponding to the north-side loading area showed a clear bulge towards the inside of the pit, with the maximum deformation occurring at a depth of approximately 6 meters between the second and third supports.

[0177] Based on the real-time early warning from the monitoring system, the technical team immediately organized the unloading of the building materials piled up on the north side, while simultaneously adding temporary diagonal bracing to reinforce the second support. Within 24 hours of the unloading being completed, the monitoring system showed that the displacement of the enclosure structure gradually decreased to 28mm, the load safety margin coefficient rose to 1.56, and the Frobenius norm dropped to 195.2. The system automatically restored the sampling frequency to 10Hz. Throughout the entire monitoring process, the total system data storage was 86GB, saving approximately 40% of storage space compared to a fixed 20Hz sampling frequency scheme. The calculation response time was controlled within 15 seconds, meeting the requirements for real-time monitoring.

[0178] This invention represents a significant technological advancement over traditional manual inspections and single-sensor monitoring methods. Traditional methods rely on periodic manual readings of inclinometer or strain gauge data, resulting in low monitoring frequency and discrete data, making it difficult to capture transient response characteristics caused by sudden load changes. This invention, by constructing a three-dimensional spatial monitoring network and acquiring high-frequency data, combined with principal component analysis for dimensionality reduction and variational mode decomposition techniques, can accurately extract the time-frequency characteristics of load response from massive amounts of monitoring data, achieving a fine characterization of the dynamic behavior of the retaining structure. The parameter inversion model employs a particle swarm optimization algorithm combined with chaotic mapping initialization and adaptive inertia weight adjustment, overcoming the shortcomings of traditional gradient methods, such as being prone to getting trapped in local optima and being sensitive to initial values, thus improving the accuracy and robustness of the inversion parameters. The two-layer game optimization model incorporates both retaining structure deformation control and support system stress balance into the optimization objectives. Through coupled iterative solutions of the upper and lower layer models, a comprehensive balance between structural safety and support efficiency is achieved. The Jacobian matrix spectral radius determination mechanism provides a quantitative prediction basis for algorithm convergence, and dynamically adjusting the search step size accelerates the convergence speed. The adaptive sampling frequency adjustment strategy adjusts the monitoring frequency in real time based on the load safety margin coefficient, effectively reducing data storage and computational load while ensuring monitoring sensitivity. Kriging interpolation, which integrates finite element mechanical constraints and radial basis function neural network residual learning, ensures that the reconstructed three-dimensional deformation field satisfies both mathematical smoothness and physical rationality, providing engineers with an intuitive and reliable basis for structural safety assessment.

[0179] It should be noted that the variables involved in this invention are explained in detail in Tables 2, 3, and 4.

[0180] Table 2. Variable Explanation Table (Part 1)

[0181]

[0182] Table 3. Variable Explanation Table (Part Two)

[0183]

[0184] Table 4. Variable Explanation Table (Part 3)

[0185] The Frobenius norm at the current moment Current moment The Frobenius norm of the previous moment : the moment before Time interval Sensor sampling frequency Jacobian matrix Jacobian matrix Line number Column elements :No. Decision variables Spectral radius of the Jacobian matrix Jacobian matrix eigenvalues Adjusted search step size Original search step size Initial covariance function value Spatial location coordinates Variance parameter Relevant distance parameters :Location and Euclidean distance between :Location Kriging interpolation estimates displacement values ​​at [location]. Spatial coordinates of the point to be interpolated :No. Weighting coefficient of each monitoring point Monitoring point and Covariance function values ​​between Monitoring point and interpolation points Covariance function values ​​between :Location displacement gradient vector at point : Three-dimensional spatial coordinate system coordinate axis direction Corrected covariance function value Mechanical constraint coefficient Reference length :No. Interpolation residuals at each monitoring point :Location Interpolation residual prediction value at Number of hidden layer nodes :No. The output weights of each hidden node Gaussian radial basis functions :No. The center vector of each hidden node Euclidean distance between the input vector and the center vector Width parameter of radial basis functions :Location Reconstructed displacement value at the location Maximum strain limit of materials surface

[0186] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for real-time monitoring of safety risks from construction load overload, characterized in that, Displacement sensors, tilt sensors, and strain sensors are uniformly deployed on the surface of the foundation pit retaining structure to form a three-dimensional spatial monitoring network to collect real-time displacement, tilt, and strain data. Principal component analysis is performed on the collected data to reduce dimensionality and extract low-dimensional feature vectors. Variational mode decomposition is used to decompose the low-dimensional feature vectors to extract intrinsic mode components. Hilbert transform is used to calculate the instantaneous frequency and instantaneous amplitude of each intrinsic mode component to construct the load response time-frequency feature matrix. A particle swarm optimization-based retaining structure parameter inversion model is established to obtain the support stiffness parameters and prestress parameters. A two-level game optimization model is constructed to perform load risk assessment and obtain the load safety margin coefficient. The Frobenius norm and its time change rate of the load response time-frequency feature matrix are calculated to determine the load state change and perform matrix reconstruction. The sensor sampling frequency is dynamically adjusted according to the load safety margin coefficient. An iterative convergence determination mechanism based on the Jacobian matrix spectral radius is constructed to adjust the search step size of the particle swarm optimization algorithm. The three-dimensional deformation field of the retaining structure is reconstructed based on the Kriging interpolation method to generate a continuous deformation cloud map.

2. The method according to claim 1, characterized in that, Principal component analysis (PCA) dimensionality reduction refers to projecting the original high-dimensional monitoring data onto several orthogonal directions with the largest variance through linear transformation to form uncorrelated principal components. The top principal components with a cumulative variance contribution rate of 95% are selected as the feature vectors after dimensionality reduction.

3. The method according to claim 2, characterized in that, Variational mode decomposition is an adaptive signal decomposition technique that decomposes the original signal into several intrinsic mode components with finite bandwidth by constructing a variational problem.

4. The method according to claim 3, characterized in that, The Hilbert transform is used to calculate the analytic signals of each intrinsic mode component to obtain the instantaneous frequency and instantaneous amplitude. The load response time-frequency characteristic matrix is ​​a two-dimensional matrix formed by arranging the instantaneous frequency and instantaneous amplitude of each intrinsic mode component according to the time sequence and frequency sequence.

5. The method according to claim 4, characterized in that, The inversion model of retaining structure parameters based on particle swarm optimization takes minimizing the root mean square error between measured displacement data and theoretical displacement data as the objective function, and adopts an adaptive adjustment strategy of inertial weight and chaotic mapping to initialize particle positions.

6. The method according to claim 5, characterized in that, The adaptive adjustment strategy of inertia weight refers to the dynamic change of the inertia weight coefficient in the particle swarm optimization algorithm with the number of iterations. In the early stage, a larger inertia weight is used to enhance the global search capability, and in the later stage, a smaller inertia weight is used to enhance the local fine search capability.

7. The method according to claim 6, characterized in that, Chaotic mapping initialization utilizes the ergodicity and randomness of chaotic sequences to initialize particle positions, making the initial particle distribution more uniform. Logistic mapping is used to generate chaotic sequences.

8. The method according to claim 7, characterized in that, In the two-layer game optimization model, the upper-layer model aims to minimize the maximum displacement amplitude of the retaining structure, while the lower-layer model aims to maximize the uniformity of stress distribution in the support system. The two-layer model achieves Nash equilibrium by coupling the solution of the support stiffness parameters.

9. The method according to claim 8, characterized in that, The load safety margin factor is defined as the ratio of the ultimate bearing capacity of the enclosure structure to the current load effect, which is calculated by combining real-time displacement data, tilt data, and strain data.

10. The method according to claim 9, characterized in that, The Frobenius norm is the square root of the sum of the squares of all elements of a matrix. The formula for calculating the rate of change of the Frobenius norm is the difference between the Frobenius norm at the current time and the Frobenius norm at the previous time, divided by the Frobenius norm at the previous time.

Images