Geotechnical engineering model optimization training method and system based on multi-source knowledge base

Abstract

The invention discloses a geotechnical engineering model optimization training method and system based on a multi-source knowledge base. The method comprises the following steps: constructing a geotechnical engineering multi-source knowledge base; collecting field sensor monitoring data in real time and preprocessing to obtain standardized time sequence monitoring data; performing time sequence feature extraction on the standardized data by adopting a Mamba model to generate a time sequence state feature data set; generating initial computable prior information and parameter constraint conditions of the geotechnical engineering model based on the knowledge base, and establishing an initial particle state set; iteratively updating the particle state by adopting a particle flow filtering algorithm to obtain geotechnical engineering model parameter posterior distribution; carrying out model numerical simulation and calculating a simulation error evaluation index; and judging whether the posterior distribution reaches a predetermined optimization convergence condition according to the error evaluation index, and outputting the geotechnical engineering model after optimization training. According to the method, the accuracy and convergence stability of geotechnical engineering model parameter optimization are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of model optimization training, and particularly relates to a geotechnical engineering model optimization training method and system based on a multi-source knowledge base. BACKGROUND

[0002] Geotechnical engineering numerical modeling and parameter optimization technology is one of the important research contents in the field of modern geotechnical engineering. The geotechnical engineering model is usually based on field monitoring data, geological survey data and historical engineering cases, and the mechanical properties and evolution process of complex rock-soil mass are analyzed and predicted by numerical simulation method to support the design and decision of actual engineering. With the development of monitoring technology and data analysis method, in recent years, the geotechnical engineering model optimization training based on multi-source knowledge fusion and data driving has been widely applied.

[0003] At present, the mainstream technical route of geotechnical engineering model parameter optimization includes inversion analysis method and data assimilation method. The inversion analysis method depends on the explicit mathematical model, and the engineering parameters are inversely calculated by minimizing the error between the monitoring data and the numerical model. The implementation process needs to be highly dependent on the model and data accuracy, and the parameter convergence performance is difficult to guarantee, which is easy to fall into local optimum. The particle filter algorithm, which is a representative of the data assimilation method, can estimate the probability distribution of parameters through the particle set in the state space, and iteratively optimize in real time by fusing observation data. However, in the application process of the existing particle filter algorithm in the field of geotechnical engineering, the multi-source geological data, geophysical data, historical engineering experience and specification requirements and other knowledge information have not been fully and effectively utilized as prior constraints of model parameters, resulting in insufficient data utilization and poor convergence stability in the parameter optimization process. At the same time, in the complex geotechnical environment, there are missing and abnormal situations in the monitoring data, and the current data processing method cannot effectively balance the data quality and model parameter updating accuracy requirements, so that the accuracy and reliability of the parameter optimization result are difficult to guarantee.

[0004] Therefore, how to provide a geotechnical engineering model optimization training method and system based on a multi-source knowledge base is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] One object of the present application is to provide a geotechnical engineering model optimization training method and system based on a multi-source knowledge base. In view of the problems of difficult multi-source knowledge unified modeling, insufficient parameter optimization accuracy and convergence stability in the existing geotechnical engineering model optimization process, the technical scheme of constructing a multi-source knowledge base of geotechnical engineering, real-time acquisition and preprocessing of sensor monitoring data, extracting time series features by Mamba model, and iterative optimization by particle flow filtering algorithm is proposed. The present application has the advantages of high model parameter optimization accuracy and strong convergence stability.

[0006] According to an embodiment of the present application, a geotechnical engineering model optimization training method based on a multi-source knowledge base comprises the following steps: A geotechnical engineering multi-source knowledge base is constructed. Real-time sensor monitoring data of a geotechnical engineering site are collected and preprocessed to obtain standardized time-series monitoring data. Mamba model is used to extract time-series features from the standardized time-series monitoring data to generate a time-series state feature dataset. Initial computable prior information and parameter constraints of a geotechnical engineering model are generated according to the geotechnical engineering multi-source knowledge base, and an initial particle state set is established in combination with the time-series state feature dataset. Particle flow filtering algorithm is used to iteratively update the initial particle state set to obtain an updated geotechnical engineering model parameter posterior distribution. Numerical simulation of the geotechnical engineering model is performed based on the updated geotechnical engineering model parameter posterior distribution, and an error evaluation index is calculated. Whether the geotechnical engineering model parameter posterior distribution meets a predetermined optimization convergence condition is determined based on the error evaluation index, and the final optimized and trained geotechnical engineering model is obtained when the predetermined optimization convergence condition is met.

[0007] Optionally, the geotechnical engineering multi-source knowledge base is constructed in particular as follows: Geological survey data, geophysical survey data, historical engineering case data, construction log data, test data, and specification experience data are dynamically evaluated for confidence, and low-confidence data is automatically eliminated according to a confidence threshold to obtain a high-confidence filtered data set. The high-confidence filtered data set is semantically analyzed to obtain a semantic constraint mapping table. The high-confidence filtered data set is dynamically standardized according to the semantic constraint mapping table to generate a standardized data set. A spatio-temporal topology index library is constructed using the spatial position, time stamp, and semantic label of each data record in the standardized data set. The historical engineering case data and the construction log data are processed for multi-layer spatio-temporal association according to the spatio-temporal topology index library to obtain a multi-source dynamic association data set. The standardized data set and the multi-source dynamic association data set are integrated for unified structure to obtain the geotechnical engineering multi-source knowledge base.

[0008] Optionally, the real-time sensor monitoring data of the geotechnical engineering site are collected and preprocessed to obtain standardized time-series monitoring data in particular as follows: The real-time collected sensor monitoring data are checked for data integrity, and missing data are interpolated and completed using the statistical distribution features of the data before and after the time series to obtain sensor monitoring data after missing data completion. The sensor monitoring data after the missing data is completed is subjected to abnormal data identification and elimination, and sensor monitoring data after abnormal data elimination is obtained; The data records of the sensor monitoring data after abnormal data elimination are mapped into the predetermined standard data field by using a semantic constraint mapping table, and the sensor monitoring data after semantic standardization is obtained; The sensor monitoring data after semantic standardization is subjected to data smoothing processing, and the sensor monitoring data after denoising is obtained; The sensor monitoring data after denoising is subjected to time series data resampling, and the sensor monitoring data with uniform time series sampling frequency is obtained; The sensor monitoring data with uniform time series sampling frequency is subjected to numerical range standardization operation, and each numerical value in the data is converted in a uniform standard numerical interval to generate standardized time series monitoring data.

[0009] Optionally, the standardized time series monitoring data is subjected to time series feature extraction by using the Mamba model to generate a time series state feature data set, specifically: A multi-scale time series group is constructed based on the standardized time series monitoring data, and an adaptive scale learning mechanism built-in the Mamba model is used to determine the optimal convolution scale of each time series, to generate adaptive scale feature data containing optimal scale information; The adaptive scale feature data is subjected to dynamic state space mapping by using a state space conversion mechanism in the Mamba model, to convert the time series feature representation into dynamic state space feature data; The dynamic state space feature data is subjected to time series feature pattern similarity matching according to a historical case time series feature pattern library in the geotechnical engineering multi-source knowledge base, to calculate the matching degree between the current monitoring data features and the historical feature patterns, and to obtain state feature data; The state feature data is subjected to a time series dynamic sparse attention mechanism, to highlight key state features with discriminability in the geotechnical engineering evolution process and to suppress redundant and low-correlation features, to generate sparse optimized key state feature data; The sparse optimized key state feature data is subjected to importance scoring by using an interpretable time series state feature screening mechanism, to select a few key feature dimensions with high importance in both historical time series pattern matching and dynamic state space mapping for reservation, to obtain time series state principal component feature data; The geotechnical engineering time series state feature data set is constructed based on the time series state principal component feature data.

[0010] Optionally, the initial computable prior information and parameter constraint conditions of the geotechnical engineering model are generated according to the geotechnical engineering multi-source knowledge base, and an initial particle state set is established in combination with the time series state feature data set, specifically: Determine the initial structure and parameter experience value range of the geotechnical engineering model according to the historical engineering case data and test data in the geotechnical engineering multi-source knowledge base, and form a parameter experience constraint set; According to the spatial distribution characteristics and geological attribute classification information in the geophysical prospecting data and geological survey data, establish the spatial parameter partition of the geotechnical engineering model, and obtain the initial parameter spatial partition information; According to the specification experience data and construction log data, extract the initial physical and mechanical constraint conditions of the geotechnical engineering model, and obtain the physical constraint set containing the initial stress boundary conditions and mechanical parameter constraints; Constrain and fuse the parameter experience constraint set, the initial parameter spatial partition information and the physical constraint set to form the initial computable prior information and parameter constraint conditions; Classify and map the principal component feature dimensions in the geotechnical engineering time sequence state feature data set according to the spatial position and parameter type to obtain a time sequence feature mapping set; According to the initial computable prior information and parameter constraint conditions of the unified structure geotechnical engineering model, and combined with the time sequence feature mapping set, generate an initial particle state set by using random sampling.

[0011] Optionally, the particle flow filtering algorithm is used to iteratively update the initial particle state set to obtain the updated geotechnical engineering model parameter posterior distribution, specifically: Determine the initial particle state distribution of the particle flow filtering algorithm according to the initial particle state set, and calculate the statistical characteristic parameters of the initial particle state distribution; Collect the latest standardized time sequence monitoring data, and determine the mapping relationship between the real-time observation data and the particle state parameters according to the time sequence feature mapping set; Calculate the particle flow path between the particle state parameters from the initial state distribution to the real-time observation data condition using the particle flow filtering algorithm, and determine the particle weight change of each particle on the particle flow path; Update the numerical value of each particle state parameter through the particle flow path and the particle weight change, and perform boundary check and constraint adjustment on the updated particle state parameter according to the parameter constraint condition to generate a corrected particle state parameter that satisfies the constraint condition; Calculate the statistical characteristic parameters of the updated particle state set using the corrected particle state parameter, and estimate the geotechnical engineering model parameter posterior distribution according to the statistical characteristic parameters; Determine whether the geotechnical engineering model parameter posterior distribution is stable and convergent, if not, recalculate the particle flow path and continue to iteratively update until the stable and convergent geotechnical engineering model parameter posterior distribution is obtained.

[0012] Optionally, the geotechnical engineering model numerical simulation is performed based on the updated geotechnical engineering model parameter posterior distribution, and an error evaluation index is calculated, specifically: According to the updated geotechnical engineering model parameter posterior distribution, the parameters of the geotechnical engineering model are valued, the parameters of the geotechnical engineering model are corresponded one by one with the geotechnical engineering model spatial grid, and the geotechnical engineering model spatial grid data after parameter valuation is obtained; According to the initial physical and mechanical constraint conditions of the geotechnical engineering model, the initial boundary conditions of the geotechnical engineering model are determined, and the initial boundary conditions are applied to the geotechnical engineering model spatial grid data after parameter valuation; The geotechnical engineering model spatial grid data after parameter valuation and initial boundary condition application is subjected to numerical simulation calculation, and the model response result is obtained; According to the model response result, the simulation calculation values at corresponding positions and times are extracted to form a simulation data sequence that can be directly compared with the actual monitoring data; The numerical difference between the simulation data sequence and the actual monitoring data at each monitoring point is calculated to obtain the difference sequence at the corresponding time of each monitoring point; The arithmetic mean of the absolute value of the difference sequence of all monitoring points, the variance of the difference sequence, and the root mean square error are calculated, and are recorded as the error evaluation index of the model simulation result.

[0013] Optionally, the error evaluation index is used to determine whether the geotechnical engineering model parameter posterior distribution meets the predetermined optimization convergence condition, and when the predetermined optimization convergence condition is met, the final optimized and trained geotechnical engineering model is obtained, specifically: According to the error evaluation index of the geotechnical engineering model numerical simulation result, the error absolute value, variance and root mean square error of each monitoring point and time sequence are comprehensively scored to generate an error comprehensive score sequence corresponding to each iteration; In the error comprehensive score sequence, the change rate of the error comprehensive scores of adjacent two iterations is calculated according to the time sequence index to obtain an error convergence trend curve; According to the preset optimization convergence threshold and the error convergence trend curve, the difference between the current iteration error comprehensive score and the threshold is calculated, and it is determined whether the error convergence trend curve remains below the threshold fluctuation range at consecutive multiple time indexes; When the error convergence trend curve remains below the threshold fluctuation range for a continuous preset number of iterations, the current geotechnical engineering model parameter posterior distribution is determined to meet the predetermined optimization convergence condition, and the geotechnical engineering model parameter posterior distribution of the current iteration is output as the final optimized and trained geotechnical engineering model parameter posterior distribution; When the error convergence trend curve does not reach the predetermined optimization convergence condition, the error comprehensive score of the current iteration is recorded, and the particle flow filtering algorithm is executed to continue the iterative updating of the parameter posterior distribution until the predetermined optimization convergence condition is met.

[0014] A geotechnical engineering model optimization training system based on a multi-source knowledge base comprises: A multi-source knowledge base construction module is configured to construct a geotechnical engineering multi-source knowledge base. A sensor data acquisition and preprocessing module is configured to acquire sensor monitoring data of a geotechnical engineering site and perform preprocessing to obtain standardized time series monitoring data. A time series feature extraction module is configured to extract time series features from the standardized time series monitoring data using a Mamba model to generate a time series state feature dataset. An initial particle state set generation module is configured to generate initial computable prior information and parameter constraints of a geotechnical engineering model based on the geotechnical engineering multi-source knowledge base and establish an initial particle state set. A particle flow filtering iterative updating module is configured to iteratively update the initial particle state set using a particle flow filtering algorithm to obtain an updated geotechnical engineering model parameter posterior distribution. A geotechnical engineering model numerical simulation module is configured to perform numerical simulation of the geotechnical engineering model and calculate error evaluation indicators. An optimization convergence judgment module is configured to determine whether the geotechnical engineering model parameter posterior distribution meets a predetermined optimization convergence condition, and output a final optimized geotechnical engineering model parameter posterior distribution when the predetermined optimization convergence condition is met.

[0015] The present application has the following advantages: (1) The present application effectively realizes unified modeling and full utilization of multi-source knowledge in the field of geotechnical engineering by constructing a geotechnical engineering multi-source knowledge base and extracting features from time series monitoring data using a Mamba model, effectively improving the accuracy and stability of parameter estimation in the optimization process of the geotechnical engineering model, and enhancing the reliability and precision of model optimization training.

[0016] (2) The present application realizes fine optimization of geotechnical engineering model parameters by iteratively updating particle state parameters and performing constraint adjustment using a particle flow filtering algorithm, significantly improving the parameter convergence and prediction accuracy of the geotechnical engineering model, and showing better adaptability in complex geotechnical engineering application scenarios.

[0017] (3) The application effectively solves the problems of low data quality and insufficient convergence in the prior art, breaks through the stability and accuracy bottleneck of particle filtering algorithm in the application of geotechnical engineering, realizes specific and significant improvement of model optimization reliability and stability, and effectively improves the engineering applicability of the geotechnical engineering model optimization training method. BRIEF DESCRIPTION OF DRAWINGS

[0018] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of the specification, illustrate embodiments of the application and are used to explain the application, and do not constitute a limitation on the application. In the drawings: Figure 1 A flowchart of a geotechnical engineering model optimization training method based on a multi-source knowledge base according to the application. DETAILED DESCRIPTION

[0019] The application will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams and only schematically show the basic structure of the application, and therefore only show the components related to the application.

[0020] REFERENCE Figure 1 A geotechnical engineering model optimization training method based on a multi-source knowledge base, comprising: constructing a geotechnical engineering multi-source knowledge base; real-time acquisition and preprocessing of sensor monitoring data of a geotechnical engineering site to obtain standardized time series monitoring data; extracting time series features from the standardized time series monitoring data using a Mamba model to generate a time series state feature dataset; generating initial computable prior information and parameter constraints of the geotechnical engineering model according to the geotechnical engineering multi-source knowledge base, and combining the time series state feature dataset to establish an initial particle state set for optimization training of the geotechnical engineering model; iterative updating of the initial particle state set using a particle flow filtering algorithm to obtain an updated geotechnical engineering model parameter posterior distribution; numerical simulation of the geotechnical engineering model based on the updated geotechnical engineering model parameter posterior distribution to obtain a numerical simulation result of the geotechnical engineering model, and calculating an error evaluation index between the numerical simulation result of the geotechnical engineering model and the actual monitoring data; judging whether the geotechnical engineering model parameter posterior distribution meets a predetermined optimization convergence condition based on the error evaluation index, and obtaining a final optimized and trained geotechnical engineering model when the predetermined optimization convergence condition is met.

[0021] In the embodiment, the construction of the geotechnical engineering multi-source knowledge base specifically comprises: The geological survey data, geophysical survey data, historical engineering case data, construction log data, test data and specification experience data are dynamically evaluated in confidence, and low-confidence data is automatically eliminated according to a confidence threshold to obtain a high-confidence filtered data set; The dynamic confidence evaluation is specifically: After obtaining the geological survey data, geophysical survey data, historical engineering case data, construction log data, test data and specification experience data, multi-dimensional quality analysis is performed on the collection time, collection frequency, collection location, collection instrument type, historical reliability record, spatial correlation with similar data and time continuity of each data source, and a comprehensive confidence score of each data record is calculated according to a preset confidence evaluation rule, including a time freshness factor, a source reliability factor, a spatial consistency factor and a cross-validation factor. The time freshness factor is used to reflect the closeness of the data to the current time, the source reliability factor is used to reflect the reliability of the collection instrument, the collection unit and the data transmission process, the spatial consistency factor is used to measure the matching degree of the data in space with adjacent data sets, and the cross-validation factor is used to evaluate the mutual consistency between the data and other data sources. The factors are combined according to the predetermined weights to form a comprehensive confidence score matrix, the data records below the confidence threshold are marked and removed, and a high-confidence filtered data set containing only data records with a confidence score higher than the threshold is obtained. The high-confidence filtered data set is semantically parsed to obtain a semantic constraint mapping table; The high-confidence filtered data set is dynamically standardized according to the semantic constraint mapping table to generate a standardized data set; The dynamic standardization processing is specifically: According to the field name, dimension definition and value range in the semantic constraint mapping table, each data record in the high-confidence filtered data set is converted by numerical standardization one by one, the mapping relationship between the original value of the data record and the predetermined standard value range of the corresponding field in the semantic constraint mapping table is calculated, the standardization mapping coefficient is dynamically adjusted based on the confidence score attached to the data record, and the original value of each data record is dynamically standardized and mapped into a unified standard value interval to obtain a standardized data set. The spatial position, timestamp and semantic label of each data record in the standardized data set are used to construct a spatio-temporal topology index library; The historical engineering case data and the construction log data are processed by multi-layer spatio-temporal association according to the spatio-temporal topology index library to obtain a multi-source dynamic association data set; The multi-layer spatio-temporal association processing is specifically: According to the spatial position, the time stamp and the semantic label information recorded in the historical engineering case data and the construction log data in the space-time topology index library, the historical engineering case data and the construction log data are associated and matched in the spatial dimension according to the spatial position, and data records in the same or adjacent spatial units are determined; The data records matched in space are associated in the time dimension, time sequence association links are established according to the collection time of each data record, and a double-layer association structure with the spatial position as the first layer and the time sequence as the second layer is formed; The data in the association structure is further subdivided in the semantic dimension according to the semantic label of the data record, a third layer of association in the semantic dimension is established, and a multi-source dynamic association data set is formed; The standardized data set and the multi-source dynamic association data set are integrated in a unified structure, and a geotechnical engineering multi-source knowledge base is obtained.

[0022] In the embodiment, the sensor monitoring data of the geotechnical engineering site is collected in real time and preprocessed to obtain standardized time sequence monitoring data, specifically: The real-time collected sensor monitoring data is subjected to data integrity check, and the missing data is interpolated and completed by using the statistical distribution characteristics of the data before and after the time sequence, to obtain sensor monitoring data after completion of missing data; The sensor monitoring data after completion of missing data is subjected to abnormal data identification and elimination, to obtain sensor monitoring data after elimination of abnormal data; Each data record of the sensor monitoring data after elimination of abnormal data is corresponded to a predetermined standard data field by using a semantic constraint mapping table, to obtain semantically standardized sensor monitoring data; The semantically standardized sensor monitoring data is subjected to data smoothing processing, to obtain denoised sensor monitoring data; The denoised sensor monitoring data is subjected to time sequence data resampling, to obtain sensor monitoring data with uniform time sequence sampling frequency; The sensor monitoring data with uniform time sequence sampling frequency is subjected to numerical range standardization operation, each value in the data is subjected to numerical range conversion according to a uniform standard numerical interval, and standardized time sequence monitoring data is generated.

[0023] In the embodiment, the Mamba model is used to extract time sequence characteristics from the standardized time sequence monitoring data, to generate a time sequence state feature data set, specifically: A multi-scale time sequence group is constructed based on the standardized time sequence monitoring data, and an adaptive scale feature data containing optimal scale information is generated by using an adaptive scale learning mechanism built in the Mamba model to determine the optimal convolution scale of each time sequence; After constructing the multi-scale time series group from the standardized time series monitoring data, input each time series into the pre-defined multi-scale convolution kernel group in the Mamba model, different convolution kernel groups correspond to different receptive fields and time windows, and establish trainable scale weight parameters for each convolution kernel group in the adaptive scale learning unit of the Mamba model; Calculate the convolution response of each convolution kernel group to the input time series using the forward propagation process, and dynamically adjust the scale weight parameters based on the response through the back propagation algorithm and gradient optimization method, so that each time series automatically selects the convolution kernel group with the highest feature extraction score in the multi-scale convolution kernel group; Mark the convolution kernel group parameters corresponding to the highest feature extraction score obtained by each time series in the training iteration process as the optimal convolution scale of the time series, and finally output the optimal convolution scale index and the corresponding convolution feature response of each time series to form adaptive scale feature data containing optimal scale information; Map the dynamic state space of the adaptive scale feature data using the state space conversion mechanism in the Mamba model to convert the time series feature representation into dynamic state space feature data that can reflect the dynamic change characteristics of the geotechnical engineering physical process; After obtaining the adaptive scale feature data containing the optimal convolution scale information, input the feature data into the state space conversion unit built-in the Mamba model, and jointly encode each time series feature vector with its time index, spatial index and historical feature pattern parameters according to the pre-set dynamic state variable set to form an initial state representation containing input state, output state and state transition factor; Use the Mamba model to dynamically learn the transition rules between state variables, input response rules and output observation rules in the training stage through trainable state transition matrix, input mapping matrix and output mapping matrix to obtain state space parameters that describe the evolution of time series features over time; Apply the learned state transition matrix and input-output mapping matrix to the adaptive scale feature data to perform state update and state output calculation on each time step feature vector to generate dynamic state space feature data that can reflect the dynamic change characteristics of the geotechnical engineering physical process, and record the corresponding state variables, state transition parameters and time index in the output to form a complete and traceable dynamic state space feature sequence; According to the historical case time series feature pattern library in the multi-source knowledge base of geotechnical engineering, perform time series feature pattern similarity matching on the dynamic state space feature data, calculate the matching degree between the current monitoring data features and the historical feature patterns, and obtain state feature data with historical time series feature pattern matching weight; After obtaining the dynamic state space feature data, the state space feature vectors at each time step are combined in chronological order into fixed-length state sequence fragments, and the key state variables, state transition parameters and time indexes of each state sequence fragment are extracted to form the sequence to be matched; The historical time sequence feature mode stored in the geotechnical engineering multi-source knowledge base is constructed into a standard state sequence template according to the same state variable and state transition parameter; By calculating the Euclidean distance, dynamic time warping distance and state transition matrix difference of the sequence to be matched and the standard state sequence template in the state variable value, state transition parameter and time index, the multi-index similarity score between each sequence to be matched and all standard state sequence templates is obtained; The multi-index similarity score is integrated according to the preset weight to form a comprehensive matching degree score matrix, and the optimal matching historical time sequence feature mode template number and the corresponding comprehensive matching degree score are labeled for each time step in the score matrix, and the state feature data with historical time sequence feature mode matching weight is output; The state feature data with historical time sequence feature mode matching weight is processed by the time sequence dynamic sparse attention mechanism to highlight the key state features that are discriminative in the geotechnical engineering evolution process and suppress redundant and low correlation features, and sparse optimized key state feature data is generated; After obtaining the state feature data with historical time sequence feature mode matching weight, the feature tensor is constructed according to the time step and state variable dimension, and the initial attention weight matrix is generated based on the historical time sequence feature mode matching weight; The initial attention weight matrix is dynamically updated by using the time sequence dynamic sparse attention mechanism, and the importance of each state variable in the feature tensor in the time sequence is weighted by introducing sparse regularization constraints in the attention calculation process, and the sparse attention weight distribution is obtained; The sparse attention weight distribution and the feature tensor are multiplied element by element to obtain the weighted feature tensor after attention sparsification; All state variable features in the weighted feature tensor are accumulated and summed along the time axis direction, and the key state variable features that contribute more to the state change in the geotechnical engineering evolution process are selected by using the preset discriminative threshold, and the redundant or low correlation state variable features with weight lower than the discriminative threshold are suppressed, and sparse optimized key state feature data is generated; The sparse optimized key state feature data is scored for importance by using the interpretable time sequence state feature screening mechanism, and a few key feature dimensions with high importance in historical time sequence mode matching and dynamic state space mapping are selected to be retained, and the time sequence state principal component feature data is obtained; After obtaining the key state feature data of sparse optimization, a feature importance evaluation function is established according to the historical time sequence feature mode matching weight and the state transition matrix parameters of the state variable in the dynamic state space mapping process; The feature importance evaluation function is used to calculate the contribution score of each key state feature variable in the historical time sequence mode matching and the sensitivity score in the dynamic state space feature mapping respectively; The contribution score and the sensitivity score are weighted and fused according to the predetermined weight to form a comprehensive importance score index; According to the comprehensive importance score index, all key state feature variables are sorted, and a few feature variables with higher comprehensive importance score are selected as the key feature dimensions that can be explained according to the importance score preset threshold, and time sequence state principal component feature data containing the key feature dimensions and corresponding importance scores are output; Based on the time sequence state principal component feature data, a geotechnical engineering time sequence state feature data set is constructed; After obtaining the time sequence state principal component feature data, the feature values of each key feature dimension in the data are restructured in a unified data storage format, and the corresponding space-time index is generated based on the time stamp and spatial location information of the feature data; The restructured feature data is mapped into the spatial coordinate system and time coordinate system of the geotechnical engineering model one by one using the space-time index, and the corresponding relationship between the feature data and the actual monitoring points and monitoring time of the geotechnical engineering site is established; According to the established space-time mapping relationship, the data of all key feature dimensions in each space-time unit are integrated to generate time sequence state feature records with monitoring points as basic units and time sequence as sorting; All time sequence state feature records are uniformly numbered and labeled to construct a geotechnical engineering time sequence state feature data set with unified data structure and clear space-time index.

[0024] In this embodiment, the initial computable prior information and parameter constraint conditions of the geotechnical engineering model are generated according to the geotechnical engineering multi-source knowledge base, and an initial particle state set is established in combination with the time sequence state feature data set, specifically: The experience value range of the initial structure and parameters of the geotechnical engineering model is determined according to the historical engineering case data and test data in the geotechnical engineering multi-source knowledge base to form a parameter experience constraint set; The spatial parameter partition of the geotechnical engineering model is established according to the spatial distribution characteristics and geological attribute classification information in the geophysical prospecting data and geological survey data to obtain initial parameter spatial partition information; According to the specification experience data and the construction log data, initial physical and mechanical constraint conditions of the geotechnical engineering model are extracted, and a physical constraint set containing initial stress boundary conditions and mechanical parameter constraints is obtained; The parameter experience constraint set, the initial parameter space partition information, and the physical constraint set are fused to form initial computable prior information and parameter constraint conditions; After obtaining the parameter experience constraint set, the initial parameter space partition information, and the physical constraint set, each parameter experience value range in the parameter experience constraint set is corresponded to a specific space partition according to the initial parameter space partition information, and parameter experience constraints corresponding to the space partition are formed; According to the initial stress boundary conditions and the mechanical parameter constraints in the physical constraint set, a physical constraint condition set of each space partition is determined; The parameter experience constraints corresponding to each space partition and the physical constraint condition set are checked for compatibility, and through the checking, the consistency between the parameter experience value range and the physical and mechanical constraints in each partition is determined, conflicting constraint conditions are excluded, and constraint conditions with high consistency are retained; The parameter experience constraints, the space partition information, and the physical constraints after the compatibility checking are uniformly expressed and structurally integrated to form unified initial computable prior information and parameter constraint conditions; The principal component feature dimensions in the geotechnical engineering time sequence state feature data set are classified and mapped according to the spatial position and the parameter type, and a time sequence feature mapping set is obtained; After obtaining the geotechnical engineering time sequence state feature data set, all principal component feature dimensions are preliminarily classified according to the spatial coordinate system according to the spatial position information recorded by each principal component feature dimension in the data set, and a feature dimension spatial grouping based on spatial position is obtained; According to the parameter type definition in the geotechnical engineering model, the principal component feature dimensions in each spatial grouping are classified again according to the specific parameter type of the geotechnical engineering model, the specific geotechnical parameter type corresponding to each feature dimension is determined, and a one-to-one mapping relationship between the feature dimension and the geotechnical parameter type is generated; After the classification is completed, an index mapping table among the spatial position, the feature dimension, and the parameter type is established, and each mapping relationship is labeled and marked to obtain a time sequence feature mapping set taking the spatial position as the main index and the parameter type as the secondary index; According to the unified structure of the initial computable prior information and the parameter constraint conditions of the geotechnical engineering model and in combination with the time sequence feature mapping set, an initial particle state set is generated by using a random sampling method.

[0025] In the embodiment, after the initial particle state set is updated by using the particle flow filtering algorithm, the updated geotechnical engineering model parameter posterior distribution is obtained, and specifically, determine an initial particle state distribution of the particle flow filtering algorithm according to the initial particle state set, and calculate statistical characteristic parameters of the initial particle state distribution; The calculation process of the statistical characteristic parameters is as follows: After determining the initial particle state distribution of the particle flow filtering algorithm according to the initial particle state set, each particle state parameter in the initial particle state set is vectorized and arranged according to the parameter dimension, forming a parameter dimension matrix; An initial mean vector of each parameter dimension is obtained by calculating the arithmetic mean of all particle state values in the parameter dimension matrix; After subtracting the corresponding initial mean vector from all particle state values of each parameter dimension, the sum of squares is calculated and divided by the number of particles minus one to obtain a variance vector of each parameter dimension; The covariance matrix between different parameter dimensions is obtained by summing the product of the deviations between different parameter dimensions and dividing by the number of particles minus one; The initial mean vector, variance vector and covariance matrix are combined to form the statistical characteristic parameters of the initial particle state distribution; Collect the latest standardized time series monitoring data, and determine the mapping relationship between the real-time observation data and the particle state parameters according to the time series feature mapping set; The particle flow path between the particle state parameters from the initial state distribution to the posterior state distribution under the condition of real-time observation data is calculated by using the particle flow filtering algorithm, and the change of the weight of each particle on the particle flow path is determined; The calculation process of the particle flow path is as follows: After obtaining the initial particle state distribution and its statistical characteristic parameters, a conditional likelihood function is established between each particle state parameter in the initial particle state set and the real-time observation data, and a particle flow continuous transformation equation is constructed, and the probability density transformation between the initial state distribution and the posterior state distribution is described as a continuous evolution process of the time variable from 0 to 1; The state transition matrix, observation matrix and control matrix parameters are set in the particle flow continuous transformation equation, and the drift term and diffusion term of the transformation equation are dynamically updated according to the real-time observation data; Each particle state parameter is differentiated with respect to the time variable, the path trajectory point of the particle state in the continuous evolution process is calculated, and the intermediate value of each particle state parameter on the path trajectory point is recorded; At each path trajectory point, the particle weight change rate is calculated according to the conditional likelihood function and the current drift term and diffusion term, and the particle weight is updated, forming a particle flow path containing the continuous evolution trajectory of the particle state parameter and the particle weight change curve; The numerical values of the particle state parameters are updated by the particle flow path and the particle weight change, and the updated particle state parameters are checked and adjusted according to the parameter constraint conditions to generate corrected particle state parameters meeting the constraint conditions; After updating the numerical values of the particle state parameters by the particle flow path and the particle weight change, the updated particle state parameters are compared with the upper and lower limit values of the parameters in the initial computable prior information and the parameter constraint conditions of the unified structure geotechnical engineering model in the parameter dimension, to check whether each particle state parameter exceeds the corresponding parameter allowable range; The particle state parameters exceeding the upper limit value are adjusted to the corresponding upper limit value by the truncation correction method, and the particle state parameters below the lower limit value are adjusted to the corresponding lower limit value by the truncation correction method; The particle state parameters within the allowable range but violating the physical and mechanical constraint relationship are projected into the feasible region meeting the physical and mechanical constraint conditions by the constraint projection algorithm; After completing the truncation correction and constraint projection, the statistical consistency of all the corrected particle state parameters is recalculated, and the particle state set is updated to generate corrected particle state parameters meeting all the parameter constraint conditions; The statistical characteristic parameters of the updated particle state set are calculated by using the corrected particle state parameters, and the geotechnical engineering model parameter posterior distribution is estimated according to the statistical characteristic parameters; It is judged whether the geotechnical engineering model parameter posterior distribution is stable and convergent, and if not, the particle flow path is recalculated for iteration and update until the stable and convergent geotechnical engineering model parameter posterior distribution is obtained.

[0026] In the embodiment, the geotechnical engineering model numerical simulation is carried out based on the updated geotechnical engineering model parameter posterior distribution, and an error evaluation index is calculated, specifically: The parameters of the geotechnical engineering model are assigned according to the updated geotechnical engineering model parameter posterior distribution, and the parameters of the geotechnical engineering model are corresponded with the geotechnical engineering model spatial grid one by one to obtain the geotechnical engineering model spatial grid data after parameter assignment; The initial boundary conditions of the geotechnical engineering model are determined according to the initial physical and mechanical constraint conditions of the geotechnical engineering model, and the initial boundary conditions are applied to the geotechnical engineering model spatial grid data after parameter assignment; After obtaining the initial physical and mechanical constraint conditions of the geotechnical engineering model, various physical quantity constraint parameters included therein are divided into stress constraint, displacement constraint and load constraint according to the mechanical type; Using the geometric structure information of the geotechnical engineering model's spatial mesh, the spatial position of each physical quantity constraint parameter is matched. The stress constraint parameter is mapped to the same boundary element in the spatial mesh as the actual stress area, the displacement constraint parameter is mapped to the same boundary element in the spatial mesh as the actual fixed support or displacement control area, and the load constraint parameter is mapped to the same boundary element in the spatial mesh as the external load application area. The physical quantity constraint parameters in each boundary element are numerically assigned and the boundary condition type is labeled. After all boundary elements have completed the constraint parameter mapping and assignment, the initial boundary conditions containing stress, displacement and load boundary conditions are formed. Numerical simulation calculations were performed on the spatial grid data of the geotechnical engineering model after parameter assignment and initial boundary conditions were applied to obtain the model response results; Based on the model response results, the simulation calculation values ​​at the corresponding locations and times are extracted to form a simulation data sequence that can be directly compared with the actual monitoring data; Calculate the numerical difference between the simulated data sequence and the actual monitoring data at each monitoring point to obtain the difference sequence at each monitoring point at the corresponding time. The calculation process for the difference sequence is as follows: After obtaining the simulation data sequence and the actual monitoring data, the simulation data and the actual monitoring data are matched one by one according to the spatial coordinate index and time series index of the monitoring points, and the simulation calculation values ​​of the same monitoring point under the same time index are paired with the actual monitoring values. For each pair of data, calculate the difference. The difference is the simulation value minus the actual monitoring value, to obtain the single-point difference of each monitoring point under each time index. All single-point differences at the same monitoring point are arranged in time series order to form a difference sequence, and the time index and spatial coordinate information corresponding to each difference are recorded to form the difference sequence of each monitoring point. For all monitoring points, calculate the arithmetic mean of the absolute values ​​of the differences, the variance of the difference sequence, and the root mean square error, and record them as error evaluation indicators for the model simulation results.

[0027] In this embodiment, the step of judging whether the posterior distribution of the geotechnical engineering model parameters has reached the predetermined optimization convergence condition based on the error evaluation index, and obtaining the final optimized geotechnical engineering model after the predetermined optimization convergence condition is met, specifically involves: Based on the error evaluation index of the numerical simulation results of the geotechnical engineering model, the absolute value, variance and root mean square error of each monitoring point and time series are comprehensively scored by multiple indicators, and the error comprehensive score sequence corresponding to each iteration is generated. The rate of change of the comprehensive error score between two adjacent iterations is calculated by indexing the time series in the comprehensive error score sequence to obtain the error convergence trend curve. The calculation process of the change rate of the error comprehensive score of the adjacent two iterations is: After obtaining the error comprehensive score sequence corresponding to each iteration, the error comprehensive score sequence is sequentially arranged according to the time sequence index to form a time sequence matrix containing each iteration time index and the corresponding error comprehensive score; The error comprehensive scores of adjacent two iterations in the time sequence matrix are subjected to difference operation, the difference is the error comprehensive score of the latter iteration minus the error comprehensive score of the former iteration, and the difference is divided by the absolute value of the error comprehensive score of the former iteration to obtain the change rate of the error comprehensive scores of adjacent two iterations; According to the preset optimization convergence threshold and the error convergence trend curve, the difference between the current iteration error comprehensive score and the threshold is calculated, and it is judged whether the error convergence trend curve remains below the threshold fluctuation range at consecutive multiple time indexes; The calculation method of the difference between the current iteration error comprehensive score and the threshold is: After obtaining the error convergence trend curve, a one-to-one correspondence is established between the error comprehensive score corresponding to each iteration time index in the trend curve and the preset optimization convergence threshold to construct a control matrix containing iteration time index, error comprehensive score and convergence threshold; The error comprehensive score of each row in the control matrix and the corresponding convergence threshold are subjected to difference calculation, the difference is the error comprehensive score minus the convergence threshold, and the positive and negative signs and absolute values of the difference are recorded; The calculated difference, positive and negative signs and absolute values are sequentially arranged into a difference sequence, and the deviation degree of the error comprehensive score from the threshold at each iteration time index is marked in the difference sequence to form a complete error comprehensive score and threshold difference dataset; When the error convergence trend curve remains below the threshold fluctuation range within a continuous preset number of iterations, the current geotechnical engineering model parameter posterior distribution is determined to satisfy the predetermined optimization convergence condition, and the geotechnical engineering model parameter posterior distribution of the current iteration is output as the final optimized geotechnical engineering model parameter posterior distribution; When the error convergence trend curve does not meet the predetermined optimization convergence condition, the error comprehensive score of the current iteration is recorded and returned to execute the particle flow filtering algorithm to continue the iterative update of the parameter posterior distribution until the predetermined optimization convergence condition is met.

[0028] A geotechnical engineering model optimization training system based on a multi-source knowledge base, comprising: A multi-source knowledge base construction module for constructing a multi-source knowledge base of geotechnical engineering; A sensor data acquisition and preprocessing module for acquiring sensor monitoring data of a geotechnical engineering site and preprocessing to obtain standardized time series monitoring data; The time sequence feature extraction module is configured to extract time sequence features from the standardized time sequence monitoring data by using a Mamba model to generate a time sequence state feature dataset. The initial particle state set generation module is configured to generate initial computable prior information and parameter constraint conditions of the geotechnical engineering model according to the geotechnical engineering multi-source knowledge base, and establish an initial particle state set. The particle flow filtering iterative updating module is configured to iteratively update the initial particle state set by using a particle flow filtering algorithm to obtain an updated geotechnical engineering model parameter posterior distribution. The geotechnical engineering model numerical simulation module is configured to perform numerical simulation of the geotechnical engineering model and calculate an error evaluation index. The optimization convergence judgment module is configured to judge whether the geotechnical engineering model parameter posterior distribution meets a predetermined optimization convergence condition, and output a final optimized and trained geotechnical engineering model parameter posterior distribution when the predetermined optimization convergence condition is met.

[0029] Embodiment: In order to verify the feasibility of the present application in implementation, the present application is applied to the safety risk monitoring scene of a certain large foundation pit construction of geotechnical engineering. In the actual construction process, various sensor devices are buried on site to monitor key data such as foundation pit displacement, settlement, stress, and groundwater level changes in real time. Generally, only single-source monitoring data is used for model parameter optimization, and multi-source data resources are not fully utilized, resulting in low data utilization rate, insufficient model accuracy, and poor convergence stability, which leads to large fluctuations in risk assessment and prediction results and affects the accuracy of construction safety management decisions.

[0030] In the specific implementation process, first, collect the geotechnical engineering multi-source knowledge base by collecting geotechnical engineering project geological survey data, geophysical exploration data, historical engineering case data, construction log data, test data, and specification experience data. The system dynamically evaluates the confidence of the multi-source data, automatically filters out data with low confidence, and further constructs a standardized data set through semantic analysis and dynamic standardization processing. Then, through the spatio-temporal correlation between data, a multi-layer spatio-temporal index library and multi-source dynamic correlation data set are constructed, and a structured geotechnical engineering multi-source knowledge base is formed.

[0031] At the same time, the sensors installed on site continuously collect real-time monitoring data. For the collected data, the system first performs integrity check, uses statistical interpolation method to complete the missing data, and then identifies and removes abnormal data. Subsequently, semantic standardization processing, data smoothing processing, resampling, and numerical range standardization are performed to generate high-quality standardized time sequence monitoring data.

[0032] The system then uses the Mamba model to extract time series features from standardized time series monitoring data. First, a multi-scale time series sequence group is constructed, the optimal scale is determined using the built-in adaptive scale learning mechanism of the model, and then the state space is converted to generate dynamic state feature data. By similarity matching with historical time series feature patterns in the multi-source knowledge base, the matching weight is obtained, and further using the dynamic sparse attention mechanism to highlight the key features, the principal component features with higher importance are selected to form the geotechnical engineering time series state feature data set.

[0033] In the model optimization phase, the system uses the multi-source knowledge base of geotechnical engineering to determine the experience range of the initial structure and parameters of the model, establishes a spatial parameter partition, and combines physical and mechanical constraints to form initial computable prior information and parameter constraints. Then, according to the above conditions and combined with the time series state feature data set, the initial particle state set required by the particle filter algorithm is generated by random sampling.

[0034] After using the particle flow filter algorithm, the system determines the initial particle state distribution, updates the particle flow path and particle weight in real time, and iteratively updates the particle state through boundary checking and constraint adjustment until the model parameter posterior distribution converges stably.

[0035] After convergence, the system assigns the model parameters to the geotechnical engineering model space grid, applies clear initial boundary conditions, and performs detailed numerical simulation calculations to extract simulation data sequences and compare them with actual monitoring data to obtain a set of error evaluation indicators for the model.

[0036] To ensure the stability and accuracy of the model parameter optimization, the system forms an error convergence trend curve through multi-index comprehensive scoring to determine whether the predetermined optimization convergence condition is met. When the error trend curves of multiple consecutive iterations are within the fluctuation threshold range, the system determines convergence and outputs the final optimized geotechnical engineering model parameter posterior distribution.

[0037] The following table is the comparison data of displacement prediction results and actual monitoring values for some monitoring points in specific construction scenarios: Table 1 Comparison of geotechnical engineering model predicted displacement and actual monitoring displacement

[0038] According to the data in Table 1, the absolute error between the predicted displacement value and the actual measured displacement value of each monitoring point is controlled within 0.16 mm, and the relative error is less than 1% after using the optimization training method of the present application, which shows high prediction accuracy and reliability. Specifically, the actual monitoring displacement value of measuring point 01 is 12.53 mm, the predicted displacement value is 12.42 mm, and the relative error is only 0.88%, effectively overcoming the large fluctuation problem commonly seen in single-source data method prediction at this position. In addition, the relative errors of measuring point 03 and measuring point 05 are 0.74% and 0.72% respectively, which reflects the high stability of the system in the small displacement prediction scenario.

[0039] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change within the technical range disclosed by the present application according to the technical solution and inventive concept of the present application, which should be covered within the protection scope of the present application.

Claims

1. A geotechnical engineering model optimization training method based on a multi-source knowledge base, characterized in that, The application relates to a method for constructing a geotechnical engineering model based on a multi-source knowledge base. The method comprises the following steps: constructing a geotechnical engineering multi-source knowledge base; collecting real-time sensor monitoring data of a geotechnical engineering site and pre-processing the data to obtain standardized time-series monitoring data; extracting time-series features from the standardized time-series monitoring data by using a Mamba model to generate a time-series state feature data set; generating initial computable prior information and parameter constraints of a geotechnical engineering model according to the geotechnical engineering multi-source knowledge base, and combining the time-series state feature data set to establish an initial particle state set; iteratively updating the initial particle state set by using a particle flow filtering algorithm to obtain an updated geotechnical engineering model parameter posterior distribution; performing numerical simulation on the geotechnical engineering model based on the updated geotechnical engineering model parameter posterior distribution, and calculating an error evaluation index; 2. The method of claim 1, wherein, judging whether the geotechnical engineering model parameter posterior distribution reaches a predetermined optimization convergence condition based on the error evaluation index, and obtaining a final optimized and trained geotechnical engineering model when the predetermined optimization convergence condition is met. The method for constructing the geotechnical engineering multi-source knowledge base comprises the following steps: performing dynamic confidence evaluation on geological survey data, geophysical survey data, historical engineering case data, construction log data, test data and specification experience data, and automatically eliminating low-confidence data according to a confidence threshold to obtain a high-confidence screening data set; performing semantic analysis on the high-confidence screening data set to obtain a semantic constraint mapping table; performing dynamic standardization processing on the high-confidence screening data set according to the semantic constraint mapping table to generate a standardized data set; constructing a space-time topology index library by using the spatial position, time stamp and semantic label of each data record in the standardized data set; performing multi-layer space-time association processing on the historical engineering case data and the construction log data according to the space-time topology index library to obtain a multi-source dynamic association data set; 3. The method of claim 1, wherein, integrating the standardized data set and the multi-source dynamic association data set in a unified structure to obtain the geotechnical engineering multi-source knowledge base. The method for collecting real-time sensor monitoring data of a geotechnical engineering site and pre-processing the data to obtain standardized time-series monitoring data comprises the following steps: performing data integrity checking on the real-time collected sensor monitoring data, and performing interpolation completion on missing data by using the statistical distribution characteristics of the data before and after the time series to obtain sensor monitoring data after missing data completion; performing abnormal data identification and elimination on the sensor monitoring data after missing data completion to obtain sensor monitoring data after abnormal data elimination; corresponding each data record of the sensor monitoring data after abnormal data elimination to a predetermined standard data field by using the semantic constraint mapping table to obtain semantically standardized sensor monitoring data; performing data smoothing processing on the semantically standardized sensor monitoring data to obtain denoised sensor monitoring data; performing time-series data resampling on the denoised sensor monitoring data to obtain sensor monitoring data with uniform time-series sampling frequency; performing numerical range standardization operation on the sensor monitoring data with uniform time-series sampling frequency to convert the values in the data into a unified standard value range to generate standardized time-series monitoring data.

4. The method of claim 1, wherein, The Mamba model is used to extract time sequence features from standardized time sequence monitoring data to generate a time sequence state feature dataset, specifically as follows: Based on the standardized time sequence monitoring data, a multi-scale time sequence group is constructed, and the adaptive scale learning mechanism built-in the Mamba model is used to determine the optimal convolution scale of each time sequence, thereby generating adaptive scale feature data containing optimal scale information; The adaptive scale feature data is mapped to a dynamic state space by using the state space conversion mechanism in the Mamba model, so as to convert the time sequence features into dynamic state space feature data; According to the historical case time sequence feature mode library in the multi-source knowledge base of geotechnical engineering, the dynamic state space feature data is subjected to time sequence feature mode similarity matching, the matching degree between the current monitoring data features and the historical feature modes is calculated, and state feature data is obtained; The time sequence dynamic sparse attention mechanism is used on the state feature data to highlight the key state features in the geotechnical engineering evolution process, suppress redundant and low-correlation features, and generate sparse and optimized key state feature data; The interpretable time sequence state feature screening mechanism is used to score the importance of the sparse and optimized key state feature data, and a small number of key feature dimensions with high importance in both historical time sequence mode matching and dynamic state space mapping are selected for retention, thereby obtaining time sequence state principal component feature data; Based on the time sequence state principal component feature data, a geotechnical engineering time sequence state feature dataset is constructed.

5. The method of claim 1, wherein, The initial computable prior information and parameter constraint conditions of the geotechnical engineering model are generated according to the multi-source knowledge base of geotechnical engineering, and an initial particle state set is established in combination with the time sequence state feature dataset, specifically as follows: The experience value range of the initial structure and parameters of the geotechnical engineering model is determined according to the historical engineering case data and test data in the multi-source knowledge base of geotechnical engineering, thereby forming a parameter experience constraint set; The spatial parameter partition of the geotechnical engineering model is established according to the spatial distribution characteristics and geological attribute classification information in the geophysical prospecting data and geological survey data, thereby obtaining initial parameter spatial partition information; The initial physical and mechanical constraint conditions of the geotechnical engineering model are extracted from the specification experience data and construction log data, thereby obtaining a physical constraint set containing initial stress boundary conditions and mechanical parameter constraints; The parameter experience constraint set, initial parameter spatial partition information and physical constraint set are fused to form the initial computable prior information and parameter constraint conditions; The principal component feature dimensions in the geotechnical engineering time sequence state feature dataset are classified and mapped according to the spatial position and parameter type, thereby obtaining a time sequence feature mapping set; According to the initial computable prior information and parameter constraint conditions of the unified structure of the geotechnical engineering model and in combination with the time sequence feature mapping set, an initial particle state set is generated by using random sampling.

6. The method of claim 1, wherein, The particle flow filtering algorithm is used to iteratively update the initial particle state set to obtain the updated geotechnical engineering model parameter posterior distribution, specifically as follows: The initial particle state distribution of the particle flow filtering algorithm is determined according to the initial particle state set, and the statistical characteristic parameters of the initial particle state distribution are calculated; Collect the latest standardized time series monitoring data, and determine the mapping relationship between real-time observation data and particle state parameters according to the time series feature mapping set; Calculate the particle flow path between the posterior state distribution of the particle state parameters from the initial state distribution to the real-time observation data condition by using the particle flow filtering algorithm, and determine the weight change of each particle on the particle flow path; Update the numerical value of each particle state parameter through the particle flow path and the weight change of the particle, and perform boundary check and constraint adjustment on the updated particle state parameter according to the parameter constraint condition to generate a corrected particle state parameter that satisfies the constraint condition; Calculate the statistical characteristic parameters of the updated particle state set by using the corrected particle state parameter, and estimate the posterior distribution of the geotechnical engineering model parameters according to the statistical characteristic parameters; Determine whether the posterior distribution of the geotechnical engineering model parameters is stable and convergent, and if not, recalculate the particle flow path and continue iteration until the stable and convergent posterior distribution of the geotechnical engineering model parameters is obtained.

7. The method of claim 1, wherein, The numerical simulation of the geotechnical engineering model is carried out based on the updated posterior distribution of the geotechnical engineering model parameters, and the error evaluation index is calculated, which is specifically: According to the updated posterior distribution of the geotechnical engineering model parameters, the parameters of the geotechnical engineering model are assigned, the parameters of the geotechnical engineering model are corresponded one by one with the spatial grid of the geotechnical engineering model, and the spatial grid data of the geotechnical engineering model after parameter assignment is obtained; According to the initial physical and mechanical constraint conditions of the geotechnical engineering model, the initial boundary conditions of the geotechnical engineering model are determined, and the initial boundary conditions are applied to the spatial grid data of the geotechnical engineering model after parameter assignment; The numerical simulation calculation is carried out on the spatial grid data of the geotechnical engineering model after parameter assignment and application of the initial boundary conditions, and the model response result is obtained; According to the model response result, the simulation calculation values at corresponding positions and times are extracted to form a simulation data sequence that can be directly compared with the actual monitoring data; The numerical difference between the simulation data sequence and the actual monitoring data at each monitoring point is calculated to obtain the difference sequence at the corresponding time of each monitoring point; The arithmetic mean of the absolute value of the difference, the variance of the difference sequence and the root mean square error are calculated for the difference sequence of all monitoring points, and are recorded as the error evaluation index of the model simulation result.

8. The method of claim 1, wherein, The posterior distribution of the geotechnical engineering model parameters is determined based on the error evaluation index of the numerical simulation result of the geotechnical engineering model, and when the predetermined optimization convergence condition is met, the final optimized and trained geotechnical engineering model is obtained, which is specifically: According to the error evaluation index of the numerical simulation result of the geotechnical engineering model, the error absolute value, variance and root mean square error of each monitoring point and time sequence are comprehensively scored, and the error comprehensive score sequence corresponding to each iteration is generated; In the error comprehensive score sequence, the change rate of the error comprehensive score of adjacent two iterations is calculated according to the time sequence index to obtain the error convergence trend curve; According to the preset optimization convergence threshold and the error convergence trend curve, the difference between the current iteration error comprehensive score and the threshold is calculated, and it is determined whether the error convergence trend curve remains below the threshold fluctuation range at consecutive multiple time indexes; When the error convergence trend curve remains below the threshold fluctuation range for a continuous preset number of iterations, the current geotechnical engineering model parameter posterior distribution is determined to meet the predetermined optimization convergence condition, and the geotechnical engineering model parameter posterior distribution of the current iteration is output as the final optimized training geotechnical engineering model parameter posterior distribution; When the error convergence trend curve does not meet the predetermined optimization convergence condition, the error comprehensive score of the current iteration is recorded and the particle flow filtering algorithm is executed to continue the iterative update of the parameter posterior distribution until the predetermined optimization convergence condition is met.

9. A multi-source knowledge base based geotechnical engineering model optimization training system, performing a multi-source knowledge base based geotechnical engineering model optimization training method according to any one of claims 1 to 8, characterized in that, Comprise: A multi-source knowledge base construction module for constructing a geotechnical engineering multi-source knowledge base; A sensor data acquisition and preprocessing module for acquiring sensor monitoring data in a geotechnical engineering site and preprocessing to obtain standardized time series monitoring data; A time series feature extraction module for extracting time series features from standardized time series monitoring data using a Mamba model to generate a time series state feature dataset; An initial particle state set generation module for generating initial computable prior information and parameter constraints of a geotechnical engineering model based on the geotechnical engineering multi-source knowledge base and establishing an initial particle state set; A particle flow filtering iterative update module for iteratively updating the initial particle state set using a particle flow filtering algorithm to obtain an updated geotechnical engineering model parameter posterior distribution; A geotechnical engineering model numerical simulation module for numerical simulation of the geotechnical engineering model and calculation of error evaluation indicators; An optimization convergence judgment module for determining whether the geotechnical engineering model parameter posterior distribution meets the predetermined optimization convergence condition, and outputting the final optimized training geotechnical engineering model parameter posterior distribution when the predetermined optimization convergence condition is met.

Images