Numerical simulation software parameter inversion optimization method based on artificial intelligence

Abstract

The invention belongs to the technical field of data processing, and particularly relates to a numerical simulation software parameter inversion optimization method based on artificial intelligence. The method comprises the following steps: collecting geological, site, support and deformation monitoring data of the deep foundation pit; preprocessing the data; an FLAC3D initial numerical model is constructed; building an AI optimization model by taking simulation and monitoring deformation errors as a target function, and outputting optimal parameters through double closed-loop optimization, outer-loop improved particle swarm global optimization and inner-loop CNN-LSTM dynamic correction parameters; and a parameter value range is determined through multiple times of inversion statistics. According to the method, the parameter optimization precision and efficiency are improved, the simulation reliability is enhanced, and technical support can be provided for deformation prediction and safety evaluation of the deep foundation pit.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to a method for parameter inversion and optimization in numerical simulation software based on artificial intelligence. Background Technology

[0002] With the increasing development of urban underground space, the scale and complexity of deep foundation pit projects are constantly rising. Deformation control during construction directly impacts the structural safety of surrounding buildings, underground pipelines, and roads. FLAC3D finite element simulation software is currently the mainstream tool for predicting deep foundation pit deformation, but its simulation accuracy depends entirely on the accuracy of its core parameters. Since these parameters are difficult to obtain directly through on-site measurement in actual projects, they need to be derived in reverse from on-site deformation monitoring data (i.e., parameter inversion). Therefore, developing efficient and accurate parameter inversion optimization methods has become a key requirement for improving the accuracy of deep foundation pit deformation prediction and ensuring project safety. Existing parameter optimization processes are prone to getting trapped in local optima and have slow iterative convergence speeds, making it difficult to efficiently find globally optimal parameters. This results in insufficient scientific validity and reliability of the parameter range, ultimately causing numerical simulation software to fail to meet the engineering safety requirements for predicting deep foundation pit deformation. Summary of the Invention

[0003] To address the technical problems existing in the background art described above, this invention proposes a parameter inversion optimization method for numerical simulation software based on artificial intelligence.

[0004] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0005] S1. Collect geological conditions, site conditions, support methods, and deformation monitoring data of historical deep foundation pit projects;

[0006] S2. Perform data preprocessing: Convert the format of the collected unstructured data and handle outliers in the deformation monitoring data.

[0007] S3. Based on FLAC3D finite element simulation software, and combined with preprocessed data, construct the initial numerical model of deep foundation pit engineering.

[0008] S4. An artificial intelligence optimization model is constructed using the error value between the simulated deformation data and the deformation monitoring data output from the initial numerical model of FLAC3D finite element method as the objective function; the objective function is: Where M is the total number of tests, These are actual monitoring data values. These are the simulated deformation data values ​​output by the finite element model;

[0009] S5. The AI-based optimization model employs a dual-loop parameter optimization approach for parameter optimization. The outer loop uses an improved particle swarm optimization algorithm for global parameter optimization, while the inner loop uses a deep learning model to learn deformation rules and constrain parameter search methods to achieve dynamic parameter correction and output the optimal parameters.

[0010] S6. Based on the optimal parameters obtained from the artificial intelligence optimization model, and combined with the results of multiple inversion verifications, statistical analysis is conducted to determine the value range of various parameters to be inverted in the FLAC3D finite element numerical simulation software.

[0011] Preferably, the geological conditions include soil layer thickness, physical parameters, and soil permeability coefficient; the site conditions include site topographic slope and ground elevation; and the support methods include support type and support material parameters.

[0012] Preferably, the implementation of format conversion of the collected unstructured data and outlier processing of the deformation monitoring data in step S2 is as follows:

[0013] S21. First, for unstructured data of support type, number it as structured data and input it;

[0014] S22. For deformation monitoring data, obtain the original deformation monitoring data of each monitoring point in the deep foundation pit project. The original deformation monitoring data includes the displacement value, settlement value and corresponding sampling timestamp of each monitoring point at different sampling times.

[0015] S23. Based on the spatial coordinates of the monitoring points and the sampling timestamps, construct a spatiotemporal matrix D of deformation monitoring data, with matrix elements... This represents the deformation monitoring value of the i-th monitoring point during the j-th sampling.

[0016] S24. A spatiotemporal attention mechanism is introduced to extract features from the spatiotemporal matrix D, and spatial attention weights and temporal attention weights are calculated respectively. The spatial attention weights are based on the spatial distance and geological correlation between each monitoring point to construct a spatial correlation matrix S, wherein... ,in Let be the straight-line distance between the i-th and k-th monitoring points. Let be the geological correlation coefficient between the i-th monitoring point and the k-th monitoring point. These are the weighting coefficients. The spatial attention weight matrix is ​​obtained by row-normalizing the spatial correlation matrix S, using the attenuation coefficient. The attention weights are constructed based on a time-related vector T constructed from the sampling time interval, where... ,in The time interval between the j-th and (j-1)-th samples is... The time interval weighting coefficient, The temporal attention weight matrix is ​​obtained by normalizing the temporal correlation vector T, which serves as the time decay coefficient. ,pass Obtain the deformation monitoring matrix after attention weighting ,in, Indicates will Transform into a diagonal matrix;

[0017] S25, Based on the weighted matrix LSTM prediction is used to initially identify outliers. The LSTM prediction model selects a matrix... The first 70% of the sampled data was used as the training set. The trained LSTM model was then used to predict the remaining 30% of the data to obtain the predicted values ​​for each monitoring point at each sampling time. Calculate the predicted residual And based on the statistical threshold of the training set residuals ,in To obtain the mean of the residuals in the training set, The standard deviation of the training set residuals; if The detected value will then be marked as an outlier and removed.

[0018] Preferably, step S3, based on FLAC3D finite element simulation software and combined with preprocessed data, constructs the initial numerical model of the deep foundation pit project as follows: First, a geometric model is constructed according to the actual shape of the object and mesh is generated; the mesh quality assessment tool of FLAC3D is used to check the element angle and twist index, and the mesh shape is optimized through smoothing algorithm or node repositioning; values ​​are assigned to different regions according to the data collected after preprocessing; and boundary constraints are set to output the simulated deformation data.

[0019] Preferably, an artificial intelligence optimization model is constructed using the error value between the simulated deformation data output by the finite element model and the deformation monitoring data as the objective function. This artificial intelligence optimization model employs a dual-loop parameter optimization approach for parameter optimization. Specifically, the outer loop uses an improved particle swarm optimization algorithm for global parameter optimization.

[0020] S511. First, perform population initialization, dividing ecological niches according to the functional subdomains of the parameters to be optimized: The vector of parameters to be inverted... ,in Represented by the total number of parameters, categorized by function as follows: There are 1 ecological niche, of which the 1st The parameter subset corresponding to each ecological niche is ,satisfy ;

[0021] S512, Improved Logistic Chaotic Mapping for Generating Initial Particles for Each Niche generate For each initial particle, the mapping formula is: ,in, The value of the chaotic sequence at the t-th iteration is 0 to 1. Let be the chaotic sequence value of the (t+1)th iteration. For chaos control parameters, The coefficient is a sinusoidal disturbance.

[0022] S513. Perform fitness variance verification and supplementation, calculate the particle fitness variance within the niche, where fitness is the reciprocal of the objective function, and the fitness variance is calculated as follows: ,like If the value is less than the set threshold, then supplement. A chaotic particle;

[0023] S514, Next, we will perform particle updates, following... ,in For inertial weights, As a learning factor, For the individual's optimal, It is the optimal ecological niche. For global optimality, The values ​​are random numbers, ranging from 0 to 1; calculate the population convergence. ,in For the standard deviation of the parameter, For the first The upper and lower limits of the parameters corresponding to each niche are set. If the convergence C is less than the set threshold, the random particles are perturbed. ,in The range of values ​​for the parameter vector. Let be the parameter vector of the particle after chaotic perturbation. This represents the intensity coefficient of chaotic perturbation; every ten iterations, niches are re-clustered based on particle fitness, with the cluster centers being the top-ranked particles. Preferred particle position;

[0024] S515. Then, adjust the parameters and calculate the error change rate. The error ,in The target error is used as the objective function; the parameters of the improved particle swarm are adjusted using the error and the rate of change of error as the second objective function, and the optimal improved particle swarm parameters are output through a genetic algorithm.

[0025] S516, Finally, perform a loop iteration. Or, when the number of iterations reaches its maximum, output... .

[0026] As a preferred approach, based on the optimization of the outer ring, the inner ring uses a deep learning model to learn deformation law constraint parameter search method to achieve dynamic parameter correction. The specific implementation of this method is as follows:

[0027] S521. Extract the trend consistency coefficient between measured deformation data and FLAC3D simulated deformation data of deep foundation pit engineering. If the trend consistency coefficient is greater than or equal to the set threshold, the inner loop will not be optimized and the product will be retained directly; otherwise, inner loop optimization will be performed.

[0028] S522. Inner loop optimization is used to construct a CNN-LSTM deep learning model. The parameters obtained in the outer loop optimization are input, and the output is a vector of parameter correction coefficients. ,in For the first The correction ratio for each parameter;

[0029] S523. Output correction coefficient vector based on characteristic deviation. Calculate the corrected parameters ,in The parameters are obtained from the outer loop optimization. The corrected parameters are then set with physical limits to prevent them from exceeding reasonable ranges.

[0030] As a preferred approach, the optimal parameters obtained based on the artificial intelligence optimization model are combined with the results of multiple inversions. The statistical analysis and determination of the value range of various parameters to be inverted in the FLAC3D finite element numerical simulation software are achieved by first performing multiple inversions to obtain multiple parameters, calculating the average and standard deviation of the multiple parameters, and finally using the average ± 1.96 × standard deviation to determine the value range of each parameter.

[0031] Compared with existing technologies, the advantages and positive effects of this invention are as follows: By introducing artificial intelligence algorithms to optimize the inversion process of deep foundation pit parameters, this invention can effectively overcome the problem of easily getting trapped in local optima in traditional optimization methods, thus improving the accuracy and efficiency of parameter optimization. Secondly, by adopting a dual closed-loop optimization mechanism, combining global and local optimization strategies, the parameter inversion can more accurately approximate the true value, avoiding the computational inefficiency and local optimization problems of traditional methods. Thirdly, through multiple inversion verifications and statistical analyses, the reasonable value range of each parameter to be inverted is accurately determined, further improving the reliability of numerical simulation. Finally, the adaptability and flexibility of this invention enable it to adapt to deep foundation pit projects of different scales and complexities, and it can be widely applied in practical engineering, providing strong technical support for deep foundation pit deformation prediction and safety assessment. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart illustrating a parameter inversion and optimization method for numerical simulation software based on artificial intelligence. Detailed Implementation

[0034] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0035] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0036] In this example, deep foundation pit engineering is widely used in urban underground space development. With the increasing scale and complexity of buildings, accurately predicting deformation during construction, especially accurately obtaining and retrieving the parameters required for numerical simulation software, becomes crucial to ensuring structural safety. However, existing parameter optimization methods are prone to getting trapped in local optima and have slow iterative convergence speeds, resulting in low efficiency in the optimization process and ultimately affecting the accuracy of numerical simulation and the safety of the project. Therefore, this invention proposes an artificial intelligence-based parameter retrieving and optimization method for numerical simulation software. The specific implementation process is as follows... Figure 1 As shown.

[0037] First, geological conditions, site conditions, support methods, and deformation monitoring data of historical deep foundation pit projects are collected. Geological conditions include soil layer thickness, physical parameters, and soil permeability coefficient; site conditions include site topography slope and ground elevation; support methods include support type and support material parameters. Then, these data are preprocessed. The data preprocessing steps include converting unstructured data into structured data and handling outliers in the deformation monitoring data to ensure accuracy and validity. Specifically, firstly, unstructured data on support types are numbered and input as structured data; for deformation monitoring data, the original deformation monitoring data of each monitoring point in the deep foundation pit project is obtained. This original deformation monitoring data includes the displacement value, settlement value, and corresponding sampling timestamp of each monitoring point at different sampling times; based on the spatial coordinates of the monitoring points and the sampling timestamps, a spatiotemporal matrix D of deformation monitoring data is constructed, with matrix elements... The deformation monitoring value of the i-th monitoring point at the j-th sampling time is represented by the following: A spatiotemporal attention mechanism is introduced to extract features from the spatiotemporal matrix D, and spatial attention weights and temporal attention weights are calculated respectively. The spatial attention weights are based on the spatial distance and geological correlation between each monitoring point to construct a spatial correlation matrix S, where... ,in Let be the straight-line distance between the i-th and k-th monitoring points. Let be the geological correlation coefficient between the i-th monitoring point and the k-th monitoring point. These are the weighting coefficients. The spatial attention weight matrix is ​​obtained by row normalizing the spatial correlation matrix S, which serves as the attenuation coefficient. The attention weights are constructed based on a time-related vector T constructed from the sampling time interval, where... ,in The time interval between the j-th and (j-1)-th samples is... The time interval weighting coefficient, The temporal attention weight matrix is ​​obtained by normalizing the temporal correlation vector T, which serves as the time decay coefficient. ,pass Obtain the deformation monitoring matrix after attention weighting ,in, Indicates will Transform into a diagonal matrix; based on the weighted matrix LSTM prediction is used to initially identify outliers. The LSTM prediction model selects a matrix... The first 70% of the sampled data was used as the training set. The trained LSTM model was then used to predict the remaining 30% of the data to obtain the predicted values ​​for each monitoring point at each sampling time. Calculate the predicted residual And based on the statistical threshold of the training set residuals ,in To obtain the mean of the residuals in the training set, The standard deviation of the training set residuals; if The detected value is then marked as an outlier and removed. After preprocessing, the collected data is integrated and matched together as input to the finite element simulation software, ensuring that each input is complete data.

[0038] Secondly, based on the FLAC3D finite element simulation software and combined with preprocessed data, an initial numerical model of the deep foundation pit project is constructed. First, a geometric model is built according to the actual shape of the object, and meshing is performed. FLAC3D's mesh quality evaluation tool is used to check element angles and torsion indices, and the mesh shape is optimized through smoothing algorithms or node repositioning. Values ​​are assigned to different regions based on the preprocessed data, and boundary constraints are set to output simulated deformation data. Specifically, when constructing the initial numerical model of the deep foundation pit project based on the FLAC3D finite element simulation software, the first step is to build a geometric model based on the actual shape of the object. This process involves accurately modeling the spatial geometry of the deep foundation pit project, including defining the shape, size, and relative position of the foundation, support structure, and other related facilities. After the geometric model is constructed, the meshing stage begins. Mesh generation is a crucial step in finite element analysis, requiring the appropriate selection of mesh size and shape based on the complexity and accuracy requirements of the geometric model. FLAC3D provides various meshing strategies, allowing different mesh densities to be set according to the needs of different regions to ensure accuracy and computational efficiency. To ensure the accuracy and stability of the numerical model, FLAC3D provides a mesh quality assessment tool. This tool evaluates mesh quality by checking indicators such as element angles and distortion. If the mesh quality is found to be unsatisfactory, smoothing algorithms or node relocation can be used to optimize the mesh shape, eliminating irregular elements or element mismatches, thereby improving computational accuracy and convergence. Next, based on the preprocessed data, different regions in the model are assigned values. Specifically, based on field monitoring data or geological survey data, different regions of the model are assigned material properties, such as the mechanical properties of the soil and the stiffness of the support structure. Simultaneously, reasonable boundary constraints are set to simulate the deformation behavior of the deep foundation pit. Finally, the simulated deformation data output by FLAC3D provides the foundational data for subsequent parameter inversion and optimization. Through a precisely constructed finite element model, combined with actual geological and engineering data, the deformation that may occur in deep foundation pits during construction can be effectively predicted, providing a reliable basis for engineering design and safety assessment.

[0039] Then, an artificial intelligence optimization model is constructed using the error value between the simulated deformation data and the deformation monitoring data output from the initial numerical model of FLAC3D finite element method as the objective function; the objective function is: Where M is the total number of tests, These are actual monitoring data values. These are the simulated deformation data values ​​output by the finite element model. An AI-based optimization model employs a dual-loop parameter optimization approach. The outer loop uses an improved particle swarm optimization algorithm for global parameter optimization, while the inner loop uses a deep learning model to learn deformation rules and constrain parameter search methods, achieving dynamic parameter correction and outputting the optimal parameters.

[0040] Specifically, the outer ring employs an improved particle swarm optimization algorithm for global parameter optimization. This involves first initializing the population and then dividing the ecological niche according to the functional subdomains of the parameters to be optimized. The vector of parameters to be inverted is then... ,in Represented by the total number of parameters, categorized by function as follows: There are 1 ecological niche, of which the 1st The parameter subset corresponding to each ecological niche is ,satisfy Improved Logistic chaotic mapping for generating initial particles, for each ecological niche. generate For each initial particle, the mapping formula is: ,in, The value of the chaotic sequence at the t-th iteration is 0 to 1. Let be the chaotic sequence value of the (t+1)th iteration. For chaos control parameters, The coefficients are sinusoidal perturbations; supplementary fitness variance verification is performed, and the particle fitness variance within the niche is calculated. The fitness is the reciprocal of the objective function, and the fitness variance is calculated as follows: ,like If the value is less than the set threshold, then supplement. One chaotic particle; next, particle updates will be performed, according to... ,in For inertial weights, As a learning factor, For the individual's optimal, This is for the optimal ecological niche. For global optimality, The values ​​are random numbers, ranging from 0 to 1; calculate the population convergence. ,in For the standard deviation of the parameter, For the first The upper and lower limits of the parameters corresponding to each niche are set. If the convergence C is less than the set threshold, the random particles are perturbed. ,in The range of values ​​for the parameter vector. Let be the parameter vector of the particle after chaotic perturbation. This represents the intensity coefficient of chaotic perturbation; every ten iterations, niches are re-clustered based on particle fitness, with the cluster centers being the top-ranked particles. Optimize particle position; then adjust parameters and calculate the error change rate. The error ,in The objective is to set the error as the target; the parameters of the improved particle swarm are adjusted using the error and the rate of change of error as a second objective function, and the optimal improved particle swarm parameters are output through a genetic algorithm; finally, iterative iterations are performed. Or, when the number of iterations reaches its maximum, output... The outer loop employs an improved Particle Swarm Optimization (PSO) algorithm. By introducing chaotic mapping, adaptive variance verification, particle update mechanisms, and random particle perturbation, it significantly enhances the global search capability and convergence speed of the optimization process. First, during population initialization, niches are divided according to functional subdomains, and initial particles are generated through an improved Logistic chaotic mapping, avoiding the initialization bias problem in traditional PSO algorithms and improving the diversity and coverage of initial solutions. The introduction of chaotic sequences enhances the exploration capability of the search space, enabling the PSO to effectively avoid getting trapped in local optima. The fitness variance verification and replenishment mechanism ensures that the population does not lose diversity during optimization. If the fitness variance of particles in a certain niche is less than a threshold, the system automatically replenishes chaotic particles, further enhancing the exploratory nature of the PSO and improving the global optimization capability. During particle update, the search efficiency of the PSO is optimized by adaptively adjusting the inertia weight and learning factor, ensuring rapid approximation of the global optimum. The random particle perturbation mechanism and the strategy of clustering to re-divide niches effectively avoid premature convergence of the algorithm and improve the flexibility of the search process. Furthermore, by introducing a combined optimization strategy of error change rate and genetic algorithm, the parameter adjustment of particle swarm was further optimized, enabling the algorithm to dynamically adapt to the optimization needs at different stages and improving the accuracy and stability of optimization.

[0041] After the outer loop optimization is completed, a decision is made on whether to perform the inner loop optimization. A deep learning model is used to learn the deformation pattern constraint parameter search method, enabling dynamic parameter correction. Specifically, the trend consistency coefficient between measured deformation data and FLAC3D simulated deformation data of deep foundation pit engineering is extracted. If the trend consistency coefficient is greater than or equal to the set threshold, the model is retained directly without inner loop optimization; otherwise, inner loop optimization is performed. Inner loop optimization involves constructing a CNN-LSTM deep learning model, inputting the parameters obtained from outer loop optimization, and outputting a vector of parameter correction coefficients. ,in For the first The correction ratio for each parameter; output the correction coefficient vector based on the characteristic deviation. Calculate the corrected parameters ,in The parameters are obtained from the outer loop optimization. The corrected parameters are then used, and physical limits are set to prevent them from exceeding reasonable ranges. The parameters obtained from outer-loop optimization are used as input to the CNN-LSTM model. The main function of the CNN (Convolutional Neural Network) part in the CNN-LSTM model is to extract local features from the input data. Convolution operations are performed on the parameters obtained from outer-loop optimization to extract local features, such as the changing trend of each parameter and its impact on deformation. Through multiple convolutional and pooling layers, the CNN can extract the spatial features of the deep pit parameters, helping the LSTM part better understand the parameter change patterns. The main function of the LSTM (Long Short-Term Memory) part is to capture the dynamic change patterns of the input parameters over time, especially the temporal dependence of pit deformation. The features extracted by the CNN and historical data are input into the LSTM model. The LSTM model processes the temporal dependence in the historical data and outputs a correction coefficient vector, representing the correction ratio of each parameter. Finally, the corrected parameters are calculated using the correction coefficient vector, and physical limits are set to prevent the corrected parameters from exceeding reasonable ranges.

[0042] Finally, based on the optimal parameters obtained from the artificial intelligence optimization model and combined with the results of multiple inversions, the value ranges of various parameters to be inverted in the FLAC3D finite element numerical simulation software were statistically analyzed and determined. Specifically, multiple inversions were first performed to obtain multiple parameters, and the average and standard deviation of the parameters were calculated. Finally, the value range of each parameter was determined using the average ± 1.96 × standard deviation. Through statistical analysis and confidence interval calculation of multiple inversion results, this method can effectively determine the reasonable value ranges of various parameters to be inverted, reduce uncertainty, improve the accuracy and stability of numerical simulation results, and provide a scientific parameter basis for practical engineering applications.

[0043] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A parameter inversion and optimization method for numerical simulation software based on artificial intelligence, characterized in that, Includes the following steps: S1. Collect geological conditions, site conditions, support methods, and deformation monitoring data of historical deep foundation pit projects; S2. Perform data preprocessing: Convert the format of the collected unstructured data and handle outliers in the deformation monitoring data. S3. Based on FLAC3D finite element simulation software, and combined with preprocessed data, construct the initial numerical model of deep foundation pit engineering. S4. An artificial intelligence optimization model is constructed using the error value between the simulated deformation data and the deformation monitoring data output from the initial numerical model of FLAC3D finite element method as the objective function; the objective function is: Where M is the total number of tests, These are actual monitoring data values. These are the simulated deformation data values ​​output by the finite element model; S5. The AI-based optimization model employs a dual-loop parameter optimization approach for parameter optimization. The outer loop uses an improved particle swarm optimization algorithm for global parameter optimization, while the inner loop uses a deep learning model to learn deformation rules and constrain parameter search methods to achieve dynamic parameter correction and output the optimal parameters. S6. Based on the optimal parameters obtained from the artificial intelligence optimization model, and combined with the results of multiple inversion verifications, statistical analysis is conducted to determine the value range of various parameters to be inverted in the FLAC3D finite element numerical simulation software.

2. The method for parameter inversion and optimization in numerical simulation software based on artificial intelligence according to claim 1, characterized in that, The geological conditions include soil layer thickness, physical parameters, and soil permeability coefficient; the site conditions include site topographic slope and ground elevation; the support methods include support type and support material parameters.

3. The method for parameter inversion and optimization in numerical simulation software based on artificial intelligence according to claim 1, characterized in that, The implementation of step S2, which involves format conversion of the collected unstructured data and outlier handling of the deformation monitoring data, is as follows: S21. First, for unstructured data of support type, number it as structured data and input it; S22. For deformation monitoring data, obtain the original deformation monitoring data of each monitoring point in the deep foundation pit project. The original deformation monitoring data includes the displacement value, settlement value and corresponding sampling timestamp of each monitoring point at different sampling times. S23. Based on the spatial coordinates of the monitoring points and the sampling timestamps, construct a spatiotemporal matrix D for deformation monitoring data, with matrix elements... This represents the deformation monitoring value of the i-th monitoring point during the j-th sampling. S24. A spatiotemporal attention mechanism is introduced to extract features from the spatiotemporal matrix D, and spatial attention weights and temporal attention weights are calculated respectively. The spatial attention weights are based on the spatial distance and geological correlation between each monitoring point to construct a spatial correlation matrix S, wherein... ,in Let be the straight-line distance between the i-th and k-th monitoring points. Let be the geological correlation coefficient between the i-th monitoring point and the k-th monitoring point. These are the weighting coefficients. The spatial attention weight matrix is ​​obtained by row normalizing the spatial correlation matrix S, which serves as the attenuation coefficient. The attention weights are constructed based on a time-related vector T constructed from the sampling time interval, where... ,in The time interval between the j-th and (j-1)-th samples is... The time interval weighting coefficient, The temporal attention weight matrix is ​​obtained by normalizing the temporal correlation vector T, which serves as the time decay coefficient. ,pass Obtain the deformation monitoring matrix after attention weighting ,in, Indicates will Transform into a diagonal matrix; S25, Based on the weighted matrix LSTM prediction is used to initially identify outliers. The LSTM prediction model selects a matrix... The first 70% of the sampled data was used as the training set. The trained LSTM model was then used to predict the remaining 30% of the data to obtain the predicted values ​​for each monitoring point at each sampling time. Calculate the predicted residual And based on the statistical threshold of the training set residuals ,in To obtain the mean of the residuals in the training set, The standard deviation of the training set residuals; if The detected value is then marked as an outlier and removed.

4. The method for parameter inversion and optimization in numerical simulation software based on artificial intelligence according to claim 1, characterized in that, The implementation method of step S3, which is based on FLAC3D finite element simulation software and combined with preprocessed data to construct the initial numerical model of deep foundation pit engineering, is as follows: First, construct a geometric model according to the actual shape of the object and perform mesh generation; use FLAC3D's mesh quality evaluation tool to check the unit angle and twist index, and optimize the mesh shape through smoothing algorithm or node repositioning. Values ​​are assigned to different regions based on the preprocessed data; boundary constraints are set to output simulated deformation data.

5. The method for parameter inversion and optimization in numerical simulation software based on artificial intelligence according to claim 1, characterized in that, An artificial intelligence optimization model is constructed using the error value between the simulated deformation data output by the finite element model and the deformation monitoring data as the objective function. This model employs a dual-loop parameter optimization approach. Specifically, the outer loop utilizes an improved particle swarm optimization algorithm for global parameter optimization. S511. First, perform population initialization, dividing ecological niches according to the functional subdomains of the parameters to be optimized: The vector of parameters to be inverted... ,in Represented by the total number of parameters, categorized by function as follows: There are 1 ecological niche, of which the 1st The parameter subset corresponding to each ecological niche is ,satisfy ; S512, Improved Logistic Chaotic Mapping for Generating Initial Particles for Each Niche generate For each initial particle, the mapping formula is: ,in, The value of the chaotic sequence at the t-th iteration is 0 to 1. Let be the chaotic sequence value of the (t+1)th iteration. For chaos control parameters, The coefficient is a sinusoidal disturbance. S513. Perform fitness variance verification and supplementation, calculate the particle fitness variance within the niche, where fitness is the reciprocal of the objective function, and the fitness variance is calculated as follows: ,like If the value is less than the set threshold, then supplement. A chaotic particle; S514, Next, we will perform particle updates, following... ,in For inertial weights, As a learning factor, For the individual's optimal, It is the optimal ecological niche. For global optimality, The values ​​are random numbers, ranging from 0 to 1; calculate the population convergence. ,in For the standard deviation of the parameter, For the first The upper and lower limits of the parameters corresponding to each niche are set. If the convergence C is less than the set threshold, the random particles are perturbed. ,in The range of values ​​for the parameter vector. Let be the parameter vector of the particle after chaotic perturbation. This represents the intensity coefficient of chaotic perturbation; every ten iterations, niches are re-clustered based on particle fitness, with the cluster centers being the top-ranked particles. Preferred particle position; S515. Then, adjust the parameters and calculate the error change rate. The error ,in The target error is used as the objective function; the parameters of the improved particle swarm are adjusted using the error and the rate of change of error as the second objective function, and the optimal improved particle swarm parameters are output through a genetic algorithm. S516, Finally, perform a loop iteration. Or, when the number of iterations reaches its maximum, output... .

6. The method for parameter inversion and optimization in numerical simulation software based on artificial intelligence according to claim 1, characterized in that, Based on the optimization of the outer ring, the inner ring uses a deep learning model to learn deformation law constraint parameter search method to achieve dynamic parameter correction. The specific implementation is as follows: S521. Extract the trend consistency coefficient between measured deformation data and FLAC3D simulated deformation data of deep foundation pit engineering. If the trend consistency coefficient is greater than or equal to the set threshold, the inner loop will not be optimized and the product will be retained directly; otherwise, inner loop optimization will be performed. S522. Inner loop optimization is used to construct a CNN-LSTM deep learning model. The parameters obtained in the outer loop optimization are input, and the output is a vector of parameter correction coefficients. ,in For the first The correction ratio for each parameter; S523. Output correction coefficient vector based on characteristic deviation. Calculate the corrected parameters ,in The parameters are obtained from the outer loop optimization. The corrected parameters are then set with physical limits to prevent them from exceeding reasonable ranges.

7. The method for parameter inversion and optimization in numerical simulation software based on artificial intelligence according to claim 1, characterized in that, Based on the optimal parameters obtained from the artificial intelligence optimization model, combined with the results of multiple inversion verifications, the range of values ​​for various parameters to be inverted in the FLAC3D finite element numerical simulation software is statistically analyzed and determined. The method is to first perform multiple inversions to obtain multiple parameters, calculate the average and standard deviation of the multiple parameters, and finally use the average ± 1.96 × standard deviation to determine the range of values ​​for each parameter.

Images