Ground vibration load driven rock mass structural surface strength prediction method

Abstract

The invention provides a rock mass structural surface strength prediction method driven by a ground vibration load, and relates to the technical field of rock mass mechanics. The method comprises the following steps: acquiring and integrating geometric and mechanical parameters, seismic oscillation parameters and test and numerical simulation data of a structural plane to form a standardized database; cleaning and standardizing database data, constructing a dynamic attenuation factor, and screening and enhancing key features to obtain initial input features; constructing a physical experience reference model and a machine learning predictor, and training a hybrid prediction model by combining prediction results of the physical experience reference model and the machine learning predictor; verifying the model and quantizing the output uncertainty; inputting the basic data calling model to complete calculation and outputting a result. According to the method, the reliability and adaptability of rock mass structural surface strength prediction are improved, and support is provided for seismic design of geotechnical engineering.

Description

Technical Field

[0001] This application relates to the field of rock mechanics technology, specifically to a method for predicting the strength of rock mass structural surfaces driven by seismic loads. Background Technology

[0002] In the fields of earthquake engineering and geotechnical engineering, the strength characteristics of rock mass structural surfaces directly determine the overall stability of rock mass engineering projects. Accurately predicting the strength of rock mass structural surfaces under seismic loads is a core aspect of ensuring the seismic safety of engineering projects. Currently, some methods for predicting the strength of rock mass structural surfaces rely on single physical empirical models, making it difficult to fully utilize the effective information from multi-source test and simulation data, and limiting the accuracy of predictions to the applicability of empirical formulas. Some purely data-driven models lack physical mechanism constraints, making them prone to overfitting and becoming detached from engineering realities. Furthermore, most methods fail to systematically quantify the uncertainty of prediction results, resulting in insufficient reliability in major engineering decisions and making it difficult to meet the accurate prediction requirements under complex seismic scenarios. Summary of the Invention

[0003] To solve, or at least partially solve, the above-mentioned technical problems, this application provides a method for predicting the strength of rock mass structural surfaces driven by seismic loads.

[0004] This application provides a method for predicting the strength of rock mass structural surfaces driven by seismic loads, including the following steps:

[0005] S1. Acquire and integrate structural geometric and mechanical parameters, seismic motion parameters, experimental data and numerical simulation data to form a standardized database;

[0006] S2. Clean and standardize the data in the standardized database, construct a dynamic attenuation factor, screen and enhance key feature variables, and obtain initial input features;

[0007] S3. Construct a physical empirical benchmark model based on the key feature variables and the dynamic attenuation factor, and generate physical model prediction features through cross-validation; construct a machine learning predictor based on the initial input features, and train the physical model prediction features and the initial input features trained by the machine learning predictor to obtain a hybrid prediction model.

[0008] S4. Validate and perform sensitivity analysis on the hybrid prediction model to quantify the uncertainty of the output results of the hybrid prediction model;

[0009] S5. By inputting the measured data, the hybrid prediction model is invoked to complete the calculation and output the prediction result.

[0010] Optionally, S1 specifically includes:

[0011] S11. Three-dimensional point cloud data of rock mass structural surfaces are acquired using a non-contact measurement method, and the geometric parameters of the structural surfaces are extracted from the three-dimensional point cloud data; the mechanical parameters of the structural surfaces are obtained through in-situ direct shear tests.

[0012] S12. Based on the geographical location information of the predicted site, obtain the ground motion parameters from an open-source seismic data platform;

[0013] S13. Collect standardized cyclic shear test data, including correlation data of the number of cycles, cyclic stress amplitude, normal stress, shear displacement, and peak shear strength; generate numerical simulation data using the discrete element method or the finite element method.

[0014] S14. Integrate the geometric and mechanical parameters of the structural surface, the seismic motion parameters, the standardized cyclic shear test data and the numerical simulation data to form the standardized database.

[0015] Optionally, S2 specifically includes:

[0016] S21. Clean the data in the standardized database, remove abnormal data and fill in missing data; then standardize the cleaned data.

[0017] S22. Based on the peak parameters, spectral parameters and cyclic load parameters in the ground motion parameters, a dynamic attenuation factor is constructed. The dynamic attenuation factor comprehensively reflects the cumulative damage effect of the cyclic load of the ground motion, the instantaneous impact effect of the peak parameters of the ground motion, and the response effect of the spectral parameters of the ground motion.

[0018] S23. The feature variables in the standardized database are screened using a feature importance evaluation algorithm to obtain the key feature variables; the key feature variables are enhanced by constructing interactive features based on the key feature variables to obtain the initial input features.

[0019] Optionally, S3 specifically includes:

[0020] S31. Based on the key feature variables and the dynamic decay factor, the classic Barton-Bandis model is dynamically improved by incorporating the dynamic decay factor into the classic Barton-Bandis model to construct the physical empirical benchmark model; a K-fold cross-validation strategy is used to generate physical model prediction features.

[0021] S32. Construct the machine learning predictor, which includes a first-layer random forest regression model and a second-layer artificial neural network model, both of which are trained using the initial input features as input. The random forest regression model is used to output the peak shear strength prediction value, and the artificial neural network model is used to output the shear stress-shear displacement curve.

[0022] S33. Extract the physical model prediction features output by the physical experience benchmark model as the new features, combine the new features with the initial input features after training to form extended input features, and input the extended input features into the first layer random forest regression model for secondary training to obtain the hybrid prediction model.

[0023] Optionally, S4 specifically includes:

[0024] S41. Divide the data in the standardized database into a training set, a validation set, and a test set. Use cross-validation to train and validate the hybrid prediction model. Use the coefficient of determination, root mean square error, and mean absolute error to comprehensively evaluate the model's prediction performance. If the evaluation results meet the preset performance requirements, then the hybrid prediction model is confirmed to be usable.

[0025] S42. The global sensitivity analysis method is used to quantify the contribution of each input feature to the variance of the model output result, and the local sensitivity analysis method is used to analyze the nonlinear influence of key input features on the model output result in different intervals. The sensitivity analysis results are obtained, and it is determined that the sensitivity analysis results meet the engineering practice and the preset analysis requirements.

[0026] S43. Treat the input parameters that affect the model's prediction results as random variables and assign them corresponding probability distributions. Use the Monte Carlo simulation method to perform propagation analysis on the uncertainty of the input parameters, obtain the probability density function, cumulative distribution function, and confidence interval of the model's output results, and complete the quantification of the uncertainty of the output results.

[0027] Optionally, S5 specifically includes:

[0028] S51. Input three types of basic data, including structural surface geometry and roughness parameters, site ground motion parameters, and normal stress conditions.

[0029] S52. Calculate the equivalent number of cycles based on the duration and dominant frequency in the site ground motion parameters, calculate the dynamic attenuation factor, run the physical empirical benchmark model and the hybrid prediction model in parallel, obtain the corrected peak shear strength, and generate a complete shear stress-shear displacement curve through the artificial neural network model.

[0030] S53. Output the prediction results, which include the determined value and confidence interval of the peak shear strength, the comparison chart of shear stress-shear displacement curves, the sensitivity analysis chart, the uncertainty cloud map, and the structured technical report.

[0031] Optionally, the following steps are also included:

[0032] S6. When new field monitoring data or experimental data are obtained, the hybrid prediction model is incrementally updated using an online random forest algorithm; wherein, during the update process, the parameters of the hybrid prediction model are adjusted based on the new data.

[0033] Optionally, S23 specifically includes:

[0034] S231. Extract the lithological parameters and stress level parameters from the standardized database as the basis for working condition classification, and divide the data in the standardized database into multiple subsets of different working condition types.

[0035] S232. For each of the aforementioned subsets, a feature importance evaluation algorithm is used to screen the key feature variables that are suitable for the corresponding working conditions, thereby obtaining a set of exclusive key features for each working condition.

[0036] S233. Based on the exclusive key feature set under each working condition, construct interactive features that conform to the mechanical properties of the corresponding working condition. The construction of the interactive features is based on the dominant influence mechanism of the structural surface strength under the corresponding working condition.

[0037] S234. Integrate the exclusive key feature set and corresponding interaction features under each working condition to form the initial input features adapted to all working conditions.

[0038] Optionally, in S43, the step of treating the input parameters affecting the model prediction results as random variables and assigning them corresponding probability distributions specifically includes:

[0039] S431. Collect on-site measured data of the structural surfaces of the target site, regional geological statistics data, and related data of similar projects. The input parameters include normal stress, roughness coefficient, and spectral acceleration.

[0040] S432. Using parameter estimation methods, the collected measured data, statistical data and related data of similar projects are fitted to obtain the initial probability distribution type and corresponding distribution parameters of each input parameter.

[0041] S433. Verify the fit between the initial probability distribution and the measured data through the KS test or AD test. If the fit does not meet the preset fit threshold, adjust the distribution type or optimize the distribution parameters until an input parameter probability distribution that matches the actual situation of the target site is obtained.

[0042] S434. Assign the calibrated probability distribution to the corresponding input parameters, and then perform the subsequent uncertainty propagation analysis using the Monte Carlo simulation method.

[0043] Optionally, S22 specifically includes:

[0044] S221. The acquired ground motion parameters are identified by type. Based on the pulse characteristics, spectral characteristics and propagation distance of the ground motion, the ground motion parameters are divided into two types of ground motion sub-parameters: near-field pulse type and far-field stable type.

[0045] S222. For near-field pulsed ground motion parameters, a first dynamic attenuation factor is constructed based on pulse peak value and pulse duration. The first dynamic attenuation factor focuses on reflecting the instantaneous impact damage effect of pulsed load.

[0046] S223. For far-field stable ground motion parameters, a second dynamic attenuation factor is constructed based on the cyclic stress amplitude and the equivalent number of cycles. The second dynamic attenuation factor focuses on reflecting the cumulative damage effect of cyclic load.

[0047] S224. The first dynamic attenuation factor and the second dynamic attenuation factor are associated with and stored with the corresponding seismic sub-parameters, respectively; when constructing the physical empirical benchmark model later, the corresponding dynamic attenuation factor is called according to the seismic motion type of the target site.

[0048] The method provided in this application has the following beneficial effects:

[0049] This method integrates structural surface parameters, seismic motion parameters, and various experimental and simulation data to form a standardized database, effectively solving the problem of scattered data sources in previous rock mass structural surface strength predictions. This provides a unified and reliable data foundation for subsequent model construction, avoiding prediction biases caused by fragmented data. The construction of dynamic attenuation factors accurately captures the influence characteristics of seismic loads on structural surface strength, while the selection and enhancement of key feature variables makes the model input information more targeted, reducing interference from redundant data and enabling the model to focus on core influencing factors. By combining a physical empirical benchmark model with a machine learning predictor, the theoretical support advantages of the former are integrated with the data fitting capabilities of the latter. The resulting hybrid prediction model is more adaptable to complex seismic load scenarios and provides more reliable prediction results compared to a single model. Model validation, sensitivity analysis, and uncertainty quantification steps allow for early verification of model performance and clarification of the reliability of results, avoiding the risk of using unreliable models. Finally, the model output results can be retrieved simply by inputting basic data, making the operation convenient and enabling rapid provision of structural surface strength references for engineering practice, meeting the prediction needs of practical applications. Attached Figure Description

[0050] Figure 1A schematic flowchart of a method for predicting the surface strength of rock mass driven by seismic load, provided in an embodiment of this application;

[0051] Figure 2 A flowchart illustrating the generation process of the hybrid prediction model provided in this application embodiment;

[0052] Figure 3 This is a schematic diagram of a shear stress-shear displacement curve provided for an embodiment of this application. Detailed Implementation

[0053] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0054] See Figures 1 to 2 This application provides a method for predicting the strength of rock mass structural surfaces driven by seismic loads, comprising the following steps:

[0055] S1. Acquire and integrate structural geometric and mechanical parameters, seismic motion parameters, experimental data and numerical simulation data to form a standardized database;

[0056] S2. Clean and standardize the data in the standardized database, construct a dynamic attenuation factor, screen and enhance key feature variables, and obtain initial input features;

[0057] S3. Construct a physical empirical benchmark model based on the key feature variables and the dynamic attenuation factor, and generate physical model prediction features through cross-validation; construct a machine learning predictor based on the initial input features, and train the physical model prediction features and the initial input features trained by the machine learning predictor to obtain a hybrid prediction model.

[0058] S4. Validate and perform sensitivity analysis on the hybrid prediction model to quantify the uncertainty of the output results of the hybrid prediction model;

[0059] S5. By inputting the measured data, the hybrid prediction model is invoked to complete the calculation and output the prediction result.

[0060] Rock mass structural surfaces are geological interfaces such as joints and fissures within a rock mass, and their strength directly determines the stability of rock mass engineering. Seismic loads are dynamic loads acting on the rock mass caused by earthquakes, and their characteristics vary with magnitude and focal depth. Geometric and mechanical parameters of structural surfaces are indicators describing their morphology and performance. Geometric parameters include attitude and roughness, while mechanical parameters include compressive and shear strength. Experimental data comes from indoor or field tests, while numerical simulation data is obtained by simulating the stress on structural surfaces using methods such as the finite element method. A standardized database is a collection of these data processed in a unified format, providing a reliable foundation for subsequent analysis. Initial input features are the set of parameters affecting the prediction results. A physical empirical benchmark model is constructed by combining physical principles and engineering experience, a machine learning predictor relies on algorithm training, and a hybrid prediction model is a comprehensive model that integrates both. Sensitivity analysis is used to clarify the degree of influence of input parameters on the output, and uncertainty quantification is used to analyze the error range of the prediction results.

[0061] Traditional prediction methods either rely on empirical formulas, whose accuracy is limited by the data range, or they use only machine learning models, which lack physical mechanism support and are unstable. Therefore, this approach combines the advantages of both: first, integrate the data to build a database; then, process the data and extract features; finally, construct a hybrid prediction model; and after verification and analysis, output the results to meet engineering requirements.

[0062] First, data acquisition and integration were carried out. Geometric parameters of structural surfaces were measured through on-site investigation, and mechanical parameters were obtained through mechanical tests. Seismic motion parameters were collected from seismic monitoring stations. Indoor shear and cyclic loading test data were collected, and simulation data were obtained using numerical methods. These data were then standardized in a unified format, and redundancy was removed to form a standardized database, ensuring data consistency and usability.

[0063] Subsequently, the database data was processed, eliminating abnormal data caused by measurement or experimental anomalies, and using interpolation to fill in missing data. Data with different dimensions were converted to a unified dimension for standardization to avoid interference from unit differences. Simultaneously, based on peak ground motion, frequency spectrum, and cyclic load parameters, a dynamic attenuation factor reflecting the impact of ground motion on structural surface strength was constructed. Then, an algorithm was used to screen key feature variables that significantly affect prediction, and initial input features were obtained by constructing feature association enhancement variables.

[0064] Next, a hybrid prediction model was constructed, incorporating rock mechanics theory and integrating the dynamic attenuation factor to build a physical empirical benchmark model. A machine learning predictor was trained using data samples to capture data patterns. The prediction results of the physical empirical benchmark model were extracted as new features, combined with the initial input features, and fed into the machine learning predictor for secondary training. After adjusting the parameters, the hybrid prediction model was obtained.

[0065] After the hybrid prediction model is built, the database data is divided into training, validation, and test sets. The model is trained using the training set, its parameters are tuned using the validation set, and its performance is evaluated using the test set to determine if the preset accuracy has been achieved. By changing the input feature parameters and observing the output changes, a sensitivity analysis is performed to clarify the degree of influence of each feature. The input parameters affecting the prediction are treated as random variables, assigned corresponding probability distributions, and simulation methods are used to analyze the propagation of uncertainty and determine the range of prediction results.

[0066] In practical applications, by inputting basic data such as structural surface parameters of the engineering site and regional seismic motion parameters, and calling the hybrid prediction model for calculation, the output structural surface strength prediction results can provide support for rock engineering design, construction and seismic assessment.

[0067] This method solves the problems of scattered data and low utilization efficiency in traditional methods by integrating multi-source data into a database. Data processing and feature optimization reduce interference and improve input effectiveness. The hybrid model combines the support of physical mechanisms with the advantages of machine learning, resulting in higher prediction accuracy. Validation, sensitivity analysis, and uncertainty quantification ensure the reliability and stability of the model, making the prediction results more realistic, providing a basis for engineering decision-making, reducing risks, and improving the safety and economy of rock mass engineering.

[0068] In some implementations, S1 specifically includes:

[0069] S11. Three-dimensional point cloud data of rock mass structural surfaces are acquired using non-contact measurement methods, and the geometric parameters of the structural surfaces are extracted from the three-dimensional point cloud data; the mechanical parameters of the structural surfaces are obtained through in-situ direct shear tests.

[0070] S12. Based on the geographical location information of the predicted site, obtain seismic motion parameters from an open-source seismic data platform;

[0071] S13. Collect standardized cyclic shear test data, which includes correlation data of the number of cycles, cyclic stress amplitude, normal stress, shear displacement, and peak shear strength; generate numerical simulation data using the discrete element method or the finite element method.

[0072] S14. Integrate the structural surface geometry and mechanical parameters, seismic motion parameters, standardized cyclic shear test data, and numerical simulation data to form a standardized database.

[0073] Non-contact measurement is a technique that obtains morphological information of rock mass structures without direct contact with the rock mass surface. Commonly used equipment includes 3D laser scanners and photogrammetric equipment. This method avoids disturbing the structural surface while efficiently acquiring comprehensive morphological data. 3D point cloud data is a massive collection of spatial point coordinates acquired through such equipment. Each point contains the location information of the structural surface and is the basis for extracting geometric parameters. In-situ direct shear tests are shear tests conducted in the natural environment of the rock mass structural surface. This method can reproduce the stress state of the structural surface to the greatest extent. During the test, normal pressure and horizontal shear force are applied to measure shear displacement and shear stress, thereby obtaining mechanical parameters.

[0074] Predicted sites refer to specific engineering sites where rock mass structural strength prediction is required, such as dam foundations for water conservancy projects and surrounding rock areas for tunnels. Geographical location information includes key spatial information such as the site's latitude and longitude, its geological structural unit, and its distance from active fault zones. This information is the core basis for locating the seismic environment of the site. Open-source seismic data platforms are publicly accessible seismic data sharing platforms that include observational data of historical earthquake events and regional seismic ground motion parameter zoning data, providing a reliable source for obtaining targeted seismic ground motion parameters.

[0075] Standardized cyclic shear test data are indoor test data processed according to unified specifications. The test object is a specimen with the same properties as the on-site structural surface. Periodic shear loads are applied on a servo testing machine, and the system records various key parameters during the test process to ensure the comparability and usability of the data. Discrete element method (DEM) and finite element method (FEM) are two commonly used numerical simulation methods. DEM is suitable for simulating the stress and deformation of discontinuous media such as rock masses, reflecting the mechanical behavior of the structural surface by defining the contact relationship between particles. FEM treats the rock mass as a continuous medium and performs mechanical analysis by dividing it into element meshes. The two methods can be selected according to the specific characteristics of the structural surface, and the generated numerical simulation data can supplement the deficiencies of the experimental data.

[0076] In practice, the first step is to acquire the geometric and mechanical parameters of the structural surfaces. A high-precision 3D laser scanner is used to scan the rock mass structural surfaces at the predicted site, covering the entire exposed area of ​​the structural surfaces to ensure that the acquired 3D point cloud data is free of significant defects. The 3D point cloud data is imported into specialized processing software, and through steps such as denoising, stitching, and registration, environmental interference points and measurement error points are eliminated. Then, geometric parameters are extracted based on the point cloud data, including the orientation of the structural surfaces, surface roughness, and structural surface opening. After the geometric parameters are extracted, a representative structural surface area at the same site is selected for in-situ direct shear tests. Before the test, the test area is cleaned, and loading devices and displacement measuring instruments in the normal and shear directions are installed. Normal pressure is applied in stages, and shear force is applied uniformly at each pressure level until shear failure occurs. The displacement data corresponding to each load level is recorded. Through data processing, mechanical parameters such as the shear strength, internal friction angle, and cohesion of the structural surfaces are obtained.

[0077] Next, we acquire seismic motion parameters. We organize the geographical location information of the predicted site, clarifying its latitude and longitude coordinates and regional geological background. Based on this information, we use an open-source seismic data platform to filter historical earthquake records, seismic motion parameter zoning results, and acceleration time-history data of typical earthquake events in the area where the site is located. According to the project's design reference period and seismic fortification level, we determine the types of seismic motion parameters to be acquired, mainly including peak ground acceleration, peak ground velocity, seismic motion duration, and spectral characteristic parameters. These parameters are exported in a unified format to form a seismic motion parameter dataset, as shown in Table 1.

[0078] Table 1

[0079]

[0080] Then, experimental data collection and numerical simulation were carried out. When collecting data from standardized indoor cyclic shear tests, specimens with lithology and structural features consistent with the predicted site rock mass were prioritized. Different normal stress levels and cyclic load parameters were set during the tests, and the number of cycles, cyclic stress amplitude, normal stress, shear displacement, and corresponding peak shear strength were recorded to ensure the data reflected the strength variation of the structural surfaces under different stress conditions. During numerical simulation, if the structural surfaces were relatively fragmented, the discrete element method was used. Based on the measured lithological parameters and geometric parameters of the structural surfaces, a particle flow model was established to simulate the relative motion between particles and the strength degradation process of the structural surfaces under cyclic shear loads. If the structural surfaces were relatively intact, the finite element method was used to construct a three-dimensional rock mass model including the structural surfaces, define the constitutive relations of the structural surfaces, apply load conditions consistent with the tests, and calculate numerical simulation data such as shear stress and shear displacement.

[0081] Finally, data integration is performed. The extracted structural geometric and mechanical parameters, acquired seismic motion parameters, collected standardized cyclic shear test data, and generated numerical simulation data are aggregated into a single data platform. The formats of various data types are standardized, for example, the units of mechanical parameters are standardized to Pascals and the units of geometric parameters are standardized to meters. At the same time, the validity of the data is verified, data with excessive experimental errors or simulation results that do not conform to physical laws are removed, and missing key information is supplemented, ultimately forming a complete and reliable standardized database.

[0082] In some implementations, S2 specifically includes:

[0083] S21. Clean the data in the standardized database, remove abnormal data and fill in missing data; then standardize the cleaned data.

[0084] S22. Based on the peak parameters, spectral parameters and cyclic load parameters in the ground motion parameters, a dynamic attenuation factor is constructed. The dynamic attenuation factor comprehensively reflects the cumulative damage effect of the cyclic load of the ground motion, the instantaneous impact effect of the peak parameters of the ground motion, and the response effect of the spectral parameters of the ground motion.

[0085] S23. Use a feature importance evaluation algorithm to filter feature variables in the standardized database to obtain key feature variables; by constructing interactive features based on key feature variables, complete the enhancement processing of key feature variables to obtain initial input features.

[0086] Peak ground motion parameters directly reflect the intensity of seismic loads, including peak ground acceleration and peak ground velocity. Spectral parameters reflect the frequency characteristics of seismic motions, such as dominant frequency and spectral acceleration. Cyclic load parameters describe the repetitive action characteristics of seismic motions, encompassing the number of cycles and cyclic stress amplitude. The dynamic attenuation factor is a key indicator that integrates these parameters to quantify the impact of seismic loads on the strength attenuation of rock mass surfaces. Feature importance assessment algorithms are tools that analyze the correlation between parameters and surface strength to identify key feature variables that dominate the prediction results. Interactive features combine multiple key feature variables according to their inherent relationships to form new parameters, which can more comprehensively reflect the synergistic effect between parameters. The initial input features are the final set of parameters used for model training after selection and enhancement.

[0087] In actual data processing, various types of raw data are first extracted from a standardized database. These data cover multiple types, including structural geometric and mechanical parameters, seismic motion parameters, experimental data, and numerical simulation data. During data cleaning, the identification of abnormal data is mainly based on the basic laws of rock mechanics and experimental common sense. For example, when the shear strength data of a certain set of structural surfaces is much lower than the conventional range of similar lithological structural surfaces, and there is no special geological condition to support it, it can be judged as abnormal data. It needs to be confirmed and removed after combining the test records of the same period and the field investigation. For missing data caused by test equipment failure, record omissions, etc., if the missing amount is small, it can be supplemented by interpolation using other data with similar working conditions. If the missing amount is large, the data gap needs to be marked and considered in subsequent analysis. Even after cleaning, the data still has different units. For example, geometric parameters may be in meters, mechanical parameters in Pascals, and seismic motion parameters in meters per second squared. Standardization is to transform these parameters to the same order of magnitude through data transformation. A common transformation method is to subtract the mean of the parameter value and then divide by the standard deviation, so that the processed data is more in line with the calculation requirements of subsequent algorithms.

[0088] The construction of the dynamic attenuation factor requires focusing on the peak ground motion parameters, spectral parameters, and cyclic loading parameters, combined with the strength degradation law of rock mass structural surfaces under dynamic loading, and achieving quantitative integration of each parameter through formulas. The corresponding formula is as follows: ;in Represents the dynamic attenuation factor; This refers to the peak ground acceleration (PGA). This is the acceleration due to gravity, used to convert peak acceleration into a dimensionless form; The dominant frequency of seismic motion. The ratio of the two frequencies reflects the degree of matching between the seismic frequency and the natural frequency of the structural surface. The higher the degree of matching, the more significant the strength attenuation of the structural surface. The number of cyclic loads represents the number of times the seismic action is repeated. and This is an empirical coefficient, determined based on a large amount of experimental data from similar projects, used to adjust the weight of each parameter's influence on the attenuation factor. In actual calculations, the parameter values ​​are first extracted from the processed data and substituted into the formula to obtain the dynamic attenuation factor, which comprehensively reflects the instantaneous impact and cyclic cumulative damage effects of seismic motion. The larger the factor value, the weaker the attenuation effect of seismic load on the structural surface strength, and vice versa.

[0089] The selection and enhancement of key feature variables rely on feature importance assessment algorithms. Commonly used algorithms include random forest and decision tree algorithms. In practice, standardized data are used as input, with the peak shear strength of the structural surface as the output target. The feature importance assessment algorithm is run, and parameters with higher scores are retained as key feature variables. These variables typically include parameters that play a dominant role in structural surface strength, such as surface roughness coefficient, normal stress, dynamic attenuation factor, and peak ground acceleration. The construction of interactive features needs to be based on rock mechanics theory to explore the intrinsic relationships between key feature variables. For example, the synergistic effect of surface roughness coefficient and normal stress significantly affects shear strength; therefore, their product can be used as an interactive feature. Similarly, the combination of dynamic attenuation factor and seismic cycle number more accurately reflects strength changes under long-term dynamic action and can also be constructed as an interactive feature. During the construction process, it is crucial to ensure that the interactive features closely reflect the changing patterns of structural surface strength, avoiding meaningless parameter combinations.

[0090] After completing the above steps, the selected key feature variables are integrated with the constructed interaction features to form the final initial input features. During the integration process, the features need to be validated again to ensure that each feature is unique, conflict-free, and comprehensively covers the main factors affecting the structural strength. If the key feature variables already include the dynamic attenuation factor, there is no need to separately include the peak ground motion parameter to avoid feature redundancy. If a certain interaction feature has a strong correlation with a key feature variable, the construction method of the interaction feature needs to be readjusted to ensure the independence and representativeness of the initial input features.

[0091] The dynamic attenuation factor was constructed to quantify the impact of seismic loads through a formula, overcoming the shortcomings of previous methods that relied solely on qualitative descriptions. Feature selection and enhancement focused on key factors, reducing the interference of irrelevant parameters on the model. These approaches collectively improved the effectiveness and relevance of the initial input features, enhancing the practicality of the entire prediction method.

[0092] In some implementations, S22 specifically includes:

[0093] S221. The acquired ground motion parameters are identified by type. Based on the pulse characteristics, spectral characteristics and propagation distance of the ground motion, the ground motion parameters are divided into two types of ground motion sub-parameters: near-field pulse type and far-field stationary type.

[0094] S222. For near-field pulsed ground motion parameters, a first dynamic attenuation factor is constructed based on pulse peak value and pulse duration. The first dynamic attenuation factor focuses on reflecting the instantaneous impact damage effect of pulsed load.

[0095] S223. For far-field stable ground motion parameters, a second dynamic attenuation factor is constructed based on the cyclic stress amplitude and the equivalent number of cycles. The second dynamic attenuation factor focuses on reflecting the cumulative damage effect of cyclic load.

[0096] S224. Associate and store the first dynamic attenuation factor and the second dynamic attenuation factor with the corresponding seismic sub-parameters respectively; when constructing the physical empirical benchmark model later, call the corresponding dynamic attenuation factor according to the seismic motion type of the target site.

[0097] The pulse characteristics of seismic ground motion refer to the short-duration, high-intensity pulse signals present in some seismic ground motions. These signals can cause violent instantaneous impacts on rock mass structures. Spectral characteristics are the distribution features of seismic ground motion energy at different frequencies, directly related to the resonance response of structural surfaces. Propagation distance is the straight-line distance from the earthquake source to the target site, and is one of the core indicators for distinguishing between near-field and far-field ground motions. Near-field pulse-type ground motions mostly occur within 10 kilometers of the source, characterized by significant pulse signals, and the damage to structural surfaces is mainly instantaneous impact damage. Far-field stable ground motions typically propagate over 50 kilometers, with relatively regular signal fluctuations, and the damage mechanism focuses on the cumulative effect of cyclic loading. These two types of seismic ground motion sub-parameters are detailed data obtained by breaking down the original seismic ground motion parameters according to damage characteristics.

[0098] Far-field ground vibrations, after propagating over long distances, have their high-frequency components filtered out, resulting in relatively stable waveforms and longer durations. Their destructive effect on structural surfaces is similar to low-cycle fatigue. Under cyclic shear stress, the surface roughness undergoes repeated ramp-shear processes, leading to particle wear and debris generation, which in turn causes a gradual decrease in cohesion and friction angle. This is a process of energy accumulation and gradual damage, the core of which lies in the number of cycles and the stress amplitude.

[0099] Near-field ground motions, especially pulse-type ground motions directly caused by fault rupture, are characterized by containing one or more velocity or displacement pulses, with high peak values ​​and short durations. Their damage to structural surfaces is more akin to high-speed impacts. Under instantaneous, massive impact loads, structural surfaces may not have sufficient time for plastic adjustment and energy dissipation, leading to:

[0100] Strain rate effect: The strength of a material changes with increasing loading rate. The shear strength of rocks and structural surfaces may be significantly different under impact loading than under static loading.

[0101] Inertial effect: The rock mass on both sides of the structural plane generates a huge difference in acceleration in an instant, which leads to violent fluctuations in normal stress or even transient tensile stress, which can easily cause loss of cohesion and overall slippage.

[0102] Impact fracture: Strong energy input can cause brittle fracture of structural protrusions rather than progressive wear, resulting in a sudden deterioration of strength. This is a process of instantaneous energy injection and sudden damage, the core of which lies in the peak pulse intensity and pulse duration. Based on the differences in the above physical mechanisms, this embodiment designs two different dynamic attenuation factor formulas to more accurately quantify its attenuation effect on the structural strength.

[0103] (1) The first dynamic attenuation factor representing the dynamic attenuation of far-field steady-state ground motion.

[0104] This factor should focus on reflecting the cumulative damage effect of cyclic loading.

[0105]

[0106] This formula explicitly links the attenuation effect to two core physical quantities: the cyclic stress ratio and the number of cycles, consistent with the fatigue cumulative damage theory. Specifically: As the first dynamic attenuation factor, ; This represents the amplitude of the cyclic shear stress. The static shear strength is calculated using the classic Barton-Bandis model. N represents the equivalent number of cycles, which can be calculated based on the duration of the ground motion and the dominant frequency. , , The empirical coefficients for calibration of far-field stable ground motion reflect the damage evolution of materials under cyclic loading.

[0107] (2) The second dynamic attenuation factor representing the dynamic attenuation of near-field pulse-type ground motion:

[0108]

[0109] This formula incorporates the three core physical elements of instantaneous impact damage—impact intensity, energy input, and frequency response—into a mathematical model by introducing the ratio of peak velocity and duration to the natural period. The second dynamic attenuation factor can capture the strain rate effect and energy impact effect of instantaneous impacts.

[0110] The dynamic attenuation factor under near-field pulse-type ground motion. is the peak ground acceleration of the seismic pulse, and g is the gravitational acceleration, used for dimensionless conversion; This refers to the peak velocity of the ground motion pulse. For reference speed, it can be taken as the order of 0.1 m / s for dimensionless conversion. For pulse duration; The inherent period of the structural surface is related to the size of the structural surface and the wave velocity. , These are empirical coefficients for calibration of near-field pulse-type ground motion.

[0111] In practice, the first step is to identify the type of ground motion parameters. Raw ground motion data for the target site and surrounding area are extracted from a standardized database, including acceleration time history curves, frequency spectra, and source location information. Signal processing techniques are used to analyze the acceleration time history curves. If the curves show obvious single or multiple pulse signals, and the difference between the pulse peak and subsequent fluctuation amplitude exceeds three times, it is preliminarily determined to have pulse characteristics. Frequency spectrum analysis reveals that if the energy is concentrated within a narrow frequency band, the spectral characteristics are significant. Combining the propagation distance, when the propagation distance from the source to the site is less than 10 kilometers, and both obvious pulse characteristics and narrow frequency band spectral characteristics are present, the corresponding ground motion parameters are classified as near-field pulse-type ground motion sub-parameters. When the propagation distance is greater than 50 kilometers, the acceleration time history curve is stable with no significant pulses, and the spectral energy distribution is uniform, it is classified as far-field stable ground motion sub-parameters. After classification, type labels are added to the two types of sub-parameters to facilitate subsequent correlation processing.

[0112] Subsequently, differentiated dynamic attenuation factors were constructed for different types of sub-parameters. In processing near-field pulse-type ground motion sub-parameters, the pulse peak value and pulse duration were extracted from the sub-parameters. Combined with the impact resistance characteristics of the structural surface, a first dynamic attenuation factor was constructed. The larger the pulse peak value and the longer the duration, the stronger the instantaneous impact; the smaller the value of the first dynamic attenuation factor, the more significant the strength attenuation of the corresponding structural surface. The construction logic of this factor is entirely designed around the impact damage mechanism, ensuring accurate quantification of the weakening effect of pulse loads on strength. In processing far-field stationary ground motion sub-parameters, the cyclic stress amplitude was obtained through time-history curve conversion. Combined with the ground motion duration and dominant frequency, the equivalent number of cycles was calculated. Based on the cumulative damage theory, a second dynamic attenuation factor was constructed. The higher the cyclic stress amplitude and the more equivalent cycles, the more severe the cumulative damage; the smaller the value of the second dynamic attenuation factor. Its design core is to conform to the strength degradation law of the structural surface under cyclic loading. During the construction process, both types of factors were dimensionless to ensure dimensional consistency when subsequently integrated with the physical empirical benchmark model.

[0113] After factor construction is completed, the associated storage and retrieval rules are set. A seismic motion type-dynamic attenuation factor association table is established in the data management system, binding the near-field pulse type label to the first dynamic attenuation factor and the far-field stationary type label to the second dynamic attenuation factor. Simultaneously, the construction parameters and applicable conditions of each factor are stored, forming a complete associated data chain. An interface is set up for the physical empirical benchmark model. When the model requires dynamic attenuation factor input, the interface automatically reads the seismic motion parameter type label of the target site, matches the corresponding dynamic attenuation factor from the association table based on the label, and imports it into the model. If the target site is affected by both types of seismic motion, the interface will weightedly fuse the first and second dynamic attenuation factors based on the energy proportion of the two types of seismic motion to generate a comprehensive attenuation factor, ensuring adaptability to complex seismic environments.

[0114] In some implementations, S23 specifically includes:

[0115] S231. Extract lithological parameters and stress grade parameters from the standardized database as the basis for working condition classification, and divide the data in the standardized database into multiple subsets of different working condition types.

[0116] S232. For each subset of data, a feature importance evaluation algorithm is used to select key feature variables that are suitable for the corresponding working conditions, so as to obtain a set of exclusive key features for each working condition.

[0117] S233. Based on the exclusive key feature set under each working condition, construct interactive features that conform to the mechanical properties of the corresponding working condition. The construction of interactive features is based on the dominant influence mechanism of the structural surface strength under the corresponding working condition.

[0118] S234. Integrate the exclusive key feature set and corresponding interaction features under each working condition to form the initial input features adapted to all working conditions.

[0119] Lithological parameters are core indicators characterizing the composition and structural properties of rock masses. Subsets are detailed datasets formed by splitting a standardized database according to working conditions. Each subset corresponds to a specific lithology-stress combination working condition, accurately reflecting the parameter correlation patterns under that condition. The feature importance evaluation algorithm quantifies the correlation between parameters and structural surface strength, selecting a set of exclusive key features that dominate the prediction of specific working conditions, avoiding interference from irrelevant parameters. Interactive features are new parameters formed by combining multiple exclusive key features according to their inherent correlations based on the mechanical properties of the working conditions. This allows for a more comprehensive capture of the synergistic effects between parameters. The initial input features are the final parameter set adapted for prediction across all working conditions, integrating exclusive features and interactive features from each working condition. Through the logic of working condition classification, exclusive feature selection, and targeted interactive feature construction, feature engineering is made more closely aligned with the mechanical essence of each working condition.

[0120] In practice, lithological parameters and stress level parameters are first extracted from a standardized database as classification dimensions. Lithological parameters are divided into soft rock and hard rock based on uniaxial compressive strength. Soft rock typically refers to rock masses with compressive strength below 30 MPa, such as mudstone and shale, whose structural strength is significantly affected by hydrogeology and cohesion. Hard rock refers to rock masses with compressive strength above 60 MPa, such as granite and limestone, whose structural strength mainly depends on roughness and wall compressive strength. Stress level parameters are divided into three levels based on normal stress: low stress (less than 5 MPa), medium stress (5-15 MPa), and high stress (greater than 15 MPa). In low-stress conditions, structural failure is mainly due to surface friction, while in high-stress conditions, wall fracture can easily lead to a sudden drop in strength. Combining these two dimensions, six basic working condition subsets are formed, such as soft rock-low stress and hard rock-high stress. During the classification process, it is necessary to ensure that the sample size of each subset meets the requirements of subsequent algorithms. If the sample size for a certain working condition is insufficient, it can be supplemented using data augmentation techniques to avoid bias in feature selection.

[0121] Next, specific key features were selected for each subset of data. A random forest algorithm was used as the feature importance evaluation tool, as shown in Table 2. Various parameters in the subset were used as input variables, and the peak shear strength of the structural surface was used as the output variable. After running the algorithm, the importance score of each parameter was obtained. Taking the soft rock-low stress subset as an example, the selection results typically used cohesion, normal stress, number of cycles, and dynamic decay factor as specific key features. These parameters are directly related to the cohesive failure and cyclic degradation characteristics of soft rock under low stress. The specific key features of the hard rock-high stress subset were mainly roughness coefficient, wall compressive strength, peak ground acceleration, and spectral acceleration, which are consistent with the friction-fracture mixed failure mechanism of hard rock under high stress. For each subset, only the top 6 parameters with the highest importance scores were retained as the specific key feature set, and non-key parameters with low scores were removed to reduce feature redundancy.

[0122] Table 2

[0123]

[0124] The construction of interactive features must strictly adhere to the strength-dominant influence mechanism of the corresponding working condition, avoiding meaningless parameter combinations. In soft rock-low stress conditions, the synergistic effect of cohesion and normal stress directly affects shear strength; therefore, their product is constructed as a cohesion-normal stress interactive term. The combination of cycle number and dynamic attenuation factor reflects the cumulative degradation of soft rock, and is constructed as a cycle-attenuation interactive term. In hard rock-high stress conditions, the combination of roughness coefficient and peak ground acceleration reflects the frictional strengthening effect under impact loads, and is constructed as a roughness-peak value interactive term. The combination of wall compressive strength and spectral acceleration reflects the degree of wall fragmentation under high-frequency loads, and is constructed as a wall-spectrum interactive term. Other working conditions follow the same logic, constructing 2-3 targeted interactive features based on mechanical laws to ensure that interactive features can supplement the information deficiencies of single features.

[0125] Finally, feature integration is performed, summarizing the exclusive key feature sets and corresponding interactive features for the six working conditions. Duplicate features are removed, and the normal stress parameter, which is present in all working conditions, is retained only once. Feature standardization is then used to eliminate dimensional differences, forming initial input features suitable for all working conditions. The integrated feature set not only includes the core influencing parameters of each working condition but also reflects the specific laws of each working condition through interactive features, avoiding the problem of insufficient adaptability of uniform features.

[0126] In some implementations, such as Figure 2 As shown, S3 specifically includes:

[0127] S31. Based on the key feature variables and the dynamic decay factor, the classic Barton-Bandis model is dynamically improved by incorporating the dynamic decay factor into the classic Barton-Bandis model to construct the physical empirical benchmark model; a K-fold cross-validation strategy is used to generate physical model prediction features.

[0128] S32. Construct the machine learning predictor, which includes a first-layer random forest regression model and a second-layer artificial neural network model, both of which are trained using the initial input features as input. The random forest regression model is used to output the peak shear strength prediction value, and the artificial neural network model is used to output the shear stress-shear displacement curve.

[0129] S33. Extract the physical model prediction features output by the physical experience benchmark model as the new features, combine the new features with the initial input features after training to form extended input features, and input the extended input features into the first layer random forest regression model for secondary training to obtain the hybrid prediction model.

[0130] In the prediction of static shear strength of rock mass structural surfaces, the classic Barton-Bandis model is the core model established based on the laws of rock mechanics and engineering experience, and achieves strength prediction by correlating key parameters of the structural surfaces. The dynamic improvement model incorporates the influence factors of dynamic loads to adapt it to the strength prediction scenario under seismic action. The machine learning predictor is a data analysis tool based on algorithms, which captures the intrinsic relationship between parameters and prediction results through data training. The random forest regression model consists of multiple independent decision trees, which improves the prediction stability through result fusion and is suitable for numerical result output. The artificial neural network model simulates the connection structure of biological neurons, has strong nonlinear fitting ability, and can generate complex curves. The new feature is to transform the prediction results of the physical experience benchmark model into supplementary parameters and expand the dimension of input information. The expanded input feature is a combination of the original initial input features and the new features.

[0131] The physical empirical benchmark model is constructed based on the classic Barton-Bandis model, whose original formula is: ,in The static peak shear strength of the structural surface. For normal stress, The surface roughness coefficient is the structural surface roughness coefficient. The compressive strength of the structural wall surface. The formula, which uses the residual friction angle of the structural surface as an example, predicts static strength through the mechanical correlation of key parameters. The core of the dynamic improvement is the integration of a dynamic attenuation factor to quantify the attenuation effect of seismic loads on strength. The improved formula is: ,in The peak dynamic shear strength of the structural surface. Using the previously constructed dynamic attenuation factor, this formula incorporates the instantaneous impact and cyclic cumulative damage effects of seismic motion into the model, making the prediction results suitable for dynamic load scenarios. In practice, the parameter values ​​are extracted from the processed dataset and substituted into the formula to obtain the prediction results of the physical empirical benchmark model.

[0132] Preferably, when constructing the physical empirical benchmark model, the corresponding dynamic attenuation factor needs to be called according to the input seismic motion type. That is: for far-field stable types: For near-field pulsed types: .

[0133] The machine learning predictor is constructed using a two-layer model design: a random forest regression model and an artificial neural network model. During training, the initial input features obtained earlier are used as common input data for both models. For the random forest regression model, the peak shear strength of the structural surface is used as the output target. By adjusting parameters such as the number of decision trees and the maximum depth, the model learns the mapping relationship between input features and peak shear strength in the training set, improving prediction accuracy. For the artificial neural network model, shear stress-shear displacement data pairs are used as the output target. Utilizing the model's nonlinear fitting capability, it learns the shear deformation patterns of the structural surface under different load conditions, enabling it to generate complete shear stress-shear displacement curves to meet the engineering requirements for deformation characteristic analysis.

[0134] The construction of a hybrid prediction model integrates the advantages of physical empirical benchmark models and machine learning predictors. First, the physical empirical benchmark model's output prediction features (i.e., the predicted peak shear strength) are extracted using a K-fold cross-validation strategy and used as new features. These new features are combined with the initial input features trained by the machine learning predictor to form more comprehensive extended input features. Subsequently, these extended input features are fed into a pre-trained random forest regression model for secondary training. During this secondary training, the model simultaneously learns the correlation patterns of the original input features and the physical constraints of the new features. The new features effectively reduce the risk of overfitting, making the prediction results more consistent with the fundamental laws of rock mechanics. By repeatedly adjusting the model parameters and optimizing the learning effect, a hybrid prediction model that balances physical mechanism support and data fitting ability is finally obtained.

[0135] In specific embodiments, the key parameters of the random forest regression model are shown in Table 3, and the key parameters of the artificial neural network model are shown in Table 4.

[0136] Table 3

[0137]

[0138] Table 4

[0139]

[0140] In some implementations, S4 specifically includes:

[0141] S41. Divide the data in the standardized database into training set, validation set and test set, and use cross-validation to train and validate the hybrid prediction model. Use the coefficient of determination, root mean square error and mean absolute error to comprehensively evaluate the model's prediction performance. If the evaluation results meet the preset performance requirements, the hybrid prediction model is confirmed to be usable.

[0142] S42. Use global sensitivity analysis to quantify the contribution of each input feature to the variance of the model output, and use local sensitivity analysis to analyze the nonlinear impact of key input features on the model output in different intervals. Obtain the sensitivity analysis results and determine that the sensitivity analysis results meet the engineering practice and the preset analysis requirements.

[0143] S43. Treat the input parameters that affect the model's prediction results as random variables and assign them corresponding probability distributions. Use the Monte Carlo simulation method to perform propagation analysis on the uncertainty of the input parameters, obtain the probability density function, cumulative distribution function, and confidence interval of the model's output results, and complete the quantification of the uncertainty of the output results.

[0144] When conducting model validation, data is first extracted from a standardized database and randomly divided into training, validation, and test sets in a 7:2:1 ratio. During this division, it is crucial to ensure that the distribution of operating conditions in each dataset is consistent with the original database, avoiding excessive concentration or absence of any particular type of operating condition in the dataset. Cross-validation employs a 5-fold cross-validation method, merging the training and validation sets and then randomly dividing them into five subsets of similar size. Four subsets are selected as training data each time, and one subset is used as validation data. This process is repeated five times. After each training iteration, the parameters of the hybrid prediction model are adjusted based on the prediction results of the validation set, such as the number of decision trees in the random forest and the number of hidden layer nodes in the neural network.

[0145] Model performance is evaluated by calculating the coefficient of determination (COP). ), root mean square error ( Mean absolute error ( The three indicators are achieved, and the corresponding formulas are as follows: ,in These are measured values. For predicted values, This is the average of the measured values. The closer the value is to 1, the better the model fit. This reflects the dispersion of the prediction error; the smaller the value, the more concentrated the error. , representing the average deviation between the predicted and measured values, directly reflects the prediction accuracy. The test set is input into the cross-validation optimized hybrid prediction model, and the values ​​of the three indicators are calculated. If... If the error values ​​are greater than the preset threshold and the latter two error indicators are less than the preset upper limit, then the model is confirmed to be usable.

[0146] Sensitivity analysis first requires identifying key input features through global sensitivity analysis. A commonly used method is the Sobol method, which is based on the principle of variance decomposition. It quantifies the contribution of each input feature to the variance of the output result by calculating the Sobol exponent of each feature. The larger the Sobol exponent, the more significant the impact of that feature on the prediction result. In the calculation process, a hybrid prediction model is used as the analysis object, and the initial input features are used as analysis variables. The output result is the peak shear strength of the structural surface. The Sobol exponent of each feature is obtained through extensive sampling calculation, and the top 5 features with the highest exponents are selected as key input features. These typically include the surface roughness coefficient, normal stress, and dynamic attenuation factor.

[0147] Local sensitivity analysis targets these key input features using a single-factor variable method. This involves keeping other input features constant while allowing the target key input feature to vary within a reasonable range at certain step increments. The model output for each variation is recorded, and the relationship between feature values ​​and output results is plotted to analyze the feature's influence across different ranges. For example, when normal stress is in a lower range, increasing its value leads to a significant increase in peak shear strength. However, when normal stress exceeds a certain threshold, the rate of increase in strength gradually slows down. This nonlinear relationship must be consistent with the mechanical properties of the rock mass structure. If the analysis results show a decrease in strength despite increased normal stress, which contradicts engineering principles, it is necessary to retrospectively check the model or data until the sensitivity analysis results meet the preset requirements.

[0148] The core of uncertainty quantification lies in assigning probability distributions to input parameters and conducting Monte Carlo simulation analysis. First, the input parameters affecting the model's prediction results are identified. Combining engineering experience and previous sensitivity analysis results, normal stress, roughness coefficient, and spectral acceleration are determined as the main random variables. For each random variable, measured data and regional statistical data from the target site are collected. Initial probability distributions are obtained through parameter estimation methods. For example, normal stress often follows a normal distribution, while roughness coefficients mostly follow a log-normal distribution.

[0149] In the Monte Carlo simulation, 10,000 random samples are taken based on the probability distribution of each input parameter. Each sample yields a set of parameter combinations, which are then input into the hybrid prediction model to calculate the corresponding output results. Statistical analysis is performed on the 10,000 output results, and probability density function curves and cumulative distribution function curves are plotted. The peak value of the probability density function curve corresponds to the most probable value of the output result, while the cumulative distribution function curve allows direct lookup of the result range at a specific confidence level. For example, the peak shear strength range corresponding to the 95% confidence interval provides a reference range for engineering applications, thus quantifying the uncertainty of the output results.

[0150] In some implementations, in S43, the input parameters affecting the model prediction results are treated as random variables and assigned corresponding probability distributions, specifically including:

[0151] S431. Collect on-site measured data of the structural surfaces of the target site, regional geological statistics data, and related data of similar projects. Input parameters include normal stress, roughness coefficient, and spectral acceleration.

[0152] S432. Using parameter estimation methods, the collected measured data, statistical data and related data of similar projects are fitted to obtain the initial probability distribution type and corresponding distribution parameters of each input parameter.

[0153] S433. Verify the fit between the initial probability distribution and the measured data through the KS test or AD test. If the fit does not meet the preset fit threshold, adjust the distribution type or optimize the distribution parameters until an input parameter probability distribution that matches the actual situation of the target site is obtained.

[0154] S434. Assign the calibrated probability distribution to the corresponding input parameters, and then perform the subsequent uncertainty propagation analysis using the Monte Carlo simulation method.

[0155] The target site refers to the specific engineering area where rock mass structural surface strength prediction is carried out. Its geological conditions and seismic environment are unique, and all parameter acquisition and analysis must be based on the actual situation of this area. Field measured data of the structural surface are first-hand data obtained through direct exploration and testing at the target site, which most accurately reflects the local structural surface characteristics. Regional geological statistics are comprehensive data covering the geological unit where the target site is located, including lithological distribution and structural development patterns, used to supplement the limitations of single-point measurements. Related data from similar projects are data from existing projects with similar lithology and working conditions to the target site, providing a reference for parameter analysis. Normal stress is the vertical pressure borne by the structural surface, directly affecting shear strength; roughness coefficient reflects the surface undulation of the structural surface and is a core influencing factor of frictional resistance; spectral acceleration reflects the energy intensity of seismic motion at a specific period. All three are input parameters that significantly affect the prediction results.

[0156] Parameter estimation methods are mathematical approaches that deduce the probability distribution type and parameters of random variables from sample data, transforming discrete data into a continuous distribution model. The initial probability distribution is a distribution form obtained based on a preliminary fit of the data, and its rationality needs to be verified through testing. The KS test and AD test are two commonly used distribution fit testing methods, which determine whether the distribution model fits reality by calculating the deviation between the sample data and the theoretical distribution. The preset fit threshold is an upper limit of deviation set according to engineering accuracy requirements; exceeding this value necessitates adjusting the distribution model. The calibrated probability distribution is the final distribution form that has been tested and optimized, highly matching the actual parameter characteristics of the target site, providing a reliable foundation for subsequent uncertainty analysis.

[0157] In practice, the first step is to collect multi-source data to form a comprehensive parameter sample library. The acquisition of on-site measured data for structural surfaces needs to be combined with the engineering layout of the target site. Three to five representative measuring points are selected in the exposed area of ​​the structural surface. Normal stress is measured using pressure sensors, and surface morphology data is acquired using a 3D laser scanner to calculate the roughness coefficient. Spectral acceleration data is extracted from historical records of seismic monitoring stations around the target site, ensuring that the data spans nearly 30 years of typical seismic events. Regional geological statistics are retrieved from regional geological survey reports of the local geological exploration institute, focusing on extracting statistical characteristics related to the input parameters, such as the range of normal stress variation and the mean roughness coefficient in the region. Related data from similar projects are obtained by consulting industry engineering databases, selecting engineering cases with consistent lithology and similar seismic resistance levels to extract their structural surface parameter records as supplementary samples.

[0158] After data collection, distribution fitting was performed using parameter estimation methods. First, statistical analysis was conducted on the sample data of the three types of input parameters, and frequency histograms were plotted to preliminarily determine the possible distribution types. If the histogram of the normal stress sample showed a symmetrical shape with a high center and low ends, it was preliminarily judged to be a normal distribution; if the roughness coefficient sample showed a right-skewed distribution, it tended to be a log-normal distribution; the spectral acceleration sample often conformed to an extreme value type I distribution. Then, the maximum likelihood estimation method was used to fit the sample data, obtaining the specific type of the initial probability distribution of each parameter and its corresponding parameters, such as the mean and standard deviation of the normal distribution, and the location and scale parameters of the log-normal distribution. Substituting these parameters into the distribution function yielded the initial probability distribution model.

[0159] After fitting, the fit is verified using either the KS test or the AD test. Either test method can be chosen, or both can be used simultaneously to improve the reliability of the results. During the verification process, the sample data of the input parameters is compared with the theoretical data of the initial probability distribution, and the test statistic is calculated. The KS test statistic is the maximum deviation between the sample cumulative distribution and the theoretical cumulative distribution, while the AD test statistic is a weighted sum of the deviations, focusing more on the fitting effect of the tail data. The test statistic is compared with a preset fit threshold. If the statistic is less than the threshold, it indicates that the initial probability distribution fits the sample data well; if the statistic exceeds the threshold, an adjustment and optimization process needs to be initiated: if there is a deviation in the distribution type judgment, the distribution type needs to be changed by combining histogram features, such as changing the normal distribution to a gamma distribution; if only the parameters are deviated, the distribution parameters can be adjusted through an iterative algorithm until the test statistic meets the requirements, resulting in a calibrated probability distribution that matches the actual parameter characteristics of the target site.

[0160] Finally, the calibrated probability distributions are assigned to their corresponding input parameters one by one: the calibrated normal distribution parameter is assigned to the normal stress, the log-normal distribution parameter to the roughness coefficient, and the extreme value type I distribution parameter to the spectral acceleration. After the assignment is completed, these input parameters become random variables with clear probabilistic characteristics, which can be directly used for uncertainty propagation analysis in subsequent Monte Carlo simulations, ensuring that the random sampling of parameters during the simulation process can truly reflect the parameter fluctuation patterns of the target site.

[0161] See Figure 2 In some implementations, S5 specifically includes:

[0162] S51. Input three types of basic data, including structural surface geometry and roughness parameters, site ground motion parameters, and normal stress conditions.

[0163] S52. Calculate the equivalent number of cycles based on the duration and dominant frequency in the site ground motion parameters, calculate the dynamic attenuation factor, run the physical empirical benchmark model and the hybrid prediction model in parallel, obtain the corrected peak shear strength, and generate a complete shear stress-shear displacement curve through the artificial neural network model.

[0164] S53. Output the prediction results, which include the determined value and confidence interval of the peak shear strength, a comparison chart of shear stress-shear displacement curves, a sensitivity analysis chart, an uncertainty cloud map, and a structured technical report.

[0165] Structural surface geometry and roughness parameters are core indicators describing the morphology of rock mass structural surfaces. Geometric parameters encompass spatial characteristics such as the orientation and dip angle of the structural surface, while roughness parameters are reflected through statistical data on surface undulations. Together, they reflect the basic physical morphology of the structural surface. Site ground motion parameters are seismic dynamic characteristic data specific to the engineering site, including key information such as ground motion duration, dominant frequency, and peak ground acceleration, directly related to the stress environment of the structural surface under seismic loading. Normal stress conditions represent the vertical pressure borne by the structural surface and are fundamental load parameters for calculating shear strength. Equivalent cycle count is a quantitative indicator that converts the continuous action of ground motion into the number of standard cyclic loads, intuitively reflecting the cumulative effect of ground motion. Peak shear strength is the core evaluation indicator of the shear resistance of the structural surface; the corrected value more closely reflects the actual stress conditions. The shear stress-shear displacement curve fully presents the deformation characteristics of the structural surface from loading to failure, providing a basis for engineering deformation control. Various charts and reports in the prediction results are carriers that transform model calculation data into intuitive engineering information, meeting the application needs of different scenarios.

[0166] In practical applications, the first step is to input and verify three types of basic data. Data input can be achieved in two convenient ways: either by filling out forms one by one in the system's graphical user interface, or by importing external data files in batches through a preset application programming interface. Both methods support common data formats, improving operational flexibility. The input structural surface geometry and roughness parameters must be combined with the site survey report to ensure that the strike, dip, and other data are consistent with the actual site conditions. The site ground motion parameters must reference the results of authoritative seismic data platforms in the region, avoiding the use of outdated or non-specific data. The normal stress conditions must be determined according to the load requirements of the engineering design, covering the stress state at different stages such as the construction and operation phases. After the data input is completed, the system will automatically start the analysis and verification process. The analysis process mainly converts data of different formats into parameter forms that the model can recognize. Verification focuses on the rationality of the data. For example, when the roughness parameters exceed the conventional range of similar lithological structural surfaces, the system will issue a prompt, requiring confirmation based on the site conditions before proceeding to the next step, ensuring the reliability of the input data.

[0167] After data verification, the calculation process begins, focusing on calculating the equivalent cycle count, dynamic attenuation factor, and parallel operation of the two models. The calculation of the equivalent cycle count is based on the duration and dominant frequency of the ground motion parameters at the site. Ground motion duration represents the duration of the seismic action, while the dominant frequency is the frequency range where ground motion energy is concentrated. Combining these two parameters quantifies the number of repeated loads exerted on the structural surface by the ground motion. The longer the duration and the closer the dominant frequency is to the natural frequency of the structural surface, the higher the equivalent cycle count, and the more significant the attenuation of the structural surface strength. The calculation of the dynamic attenuation factor utilizes the previously constructed quantification system, combining the equivalent cycle count with the site ground motion parameters to obtain a factor value suitable for the site, ensuring accurate reflection of the impact of local ground motion on the structural surface. Subsequently, a parallel computing mechanism was initiated, simultaneously running a physical empirical benchmark model and a hybrid prediction model: the physical empirical benchmark model provides basic strength predictions based on rock mechanics principles, while the hybrid prediction model integrates the advantages of machine learning to output more accurate results. The two are compared and fused through a preset correction algorithm to ultimately obtain the peak shear strength that takes into account both mechanism and data support. At the same time, the artificial neural network model generates a complete shear stress-shear displacement curve based on input parameters and deformation patterns learned during training. This curve clearly presents the entire process of the structural surface from elastic deformation, plastic yielding to shear failure, making up for the limitation of only outputting peak strength.

[0168] After the calculations are completed, the results are output, forming a complete set of technical achievements. The core data includes the determined peak shear strength and its confidence interval. The determined value directly serves as a core parameter for engineering design, while the confidence interval provides engineers with a reference for the reliability of the results, facilitating the provision of reasonable safety margins in the design. A shear stress-shear displacement curve comparison chart displays the model's predicted curve alongside experimental data curves for similar working conditions. Figure 3 As shown, the outputs intuitively demonstrate prediction accuracy, helping users quickly assess the reliability of the model results. Sensitivity analysis charts use bar charts and other formats to indicate the impact of each input parameter on strength; for example, when the normal stress has the highest impact, it indicates that stress state needs to be carefully controlled in engineering projects. Uncertainty cloud maps present the probabilistic characteristics of the output results through dense data point distribution, making the abstract uncertainty quantification results easier to understand. Structured technical reports systematically organize the sources of input data, parameter settings during calculation, and the engineering significance of the results, forming a complete technical archive to meet engineering acceptance and archiving requirements. These outputs complement each other, covering all needs from core data to auxiliary analysis.

[0169] In some implementations, the following steps are also included:

[0170] S6. When new field monitoring data or experimental data are obtained, the hybrid prediction model is incrementally updated using the online random forest algorithm; during the update process, the parameters of the hybrid prediction model are adjusted based on the new data.

[0171] On-site monitoring data refers to the real-time data collected by sensors and other equipment during the construction and operation of rock mass engineering, including data on structural surface stress, deformation, and environmental loads. This data directly reflects the changes in structural surface characteristics under actual working conditions. Experimental data, on the other hand, is obtained from subsequent indoor shear tests or in-situ tests, supplementing the existing database with data covering insufficient conditions. The online random forest algorithm is an incremental learning algorithm optimized from the traditional random forest algorithm. Its core feature is the ability to adjust model parameters using only new data without retraining all historical data. Incremental updates in hybrid prediction models refer to adaptively modifying the model's decision logic and output patterns based on newly acquired data, ensuring the model aligns with the latest engineering realities.

[0172] In a preferred embodiment, when new field monitoring data or experimental data is acquired, the system initiates a robust online incremental learning process, specifically including:

[0173] S61. Perform outlier detection and isolation on newly added data, and conduct preliminary verification through data distribution to ensure the reliability of input data and prevent noise from contaminating the model.

[0174] S62. By monitoring the statistical changes in the prediction error of the model, actively identify the gradual or sudden drift of the data distribution.

[0175] S63. Based on the drift detection results, dynamically adjust the update window size and learning rate. Use incremental updates for gradual drifts and aggressive updates for sudden drifts to ensure the model can quickly adapt to new patterns.

[0176] S64. The online random forest algorithm is adopted to adjust the parameters of the hybrid prediction model based on newly added data that has passed quality assurance and an adaptive update strategy.

[0177] S65. Place the updated candidate model in shadow mode and perform performance testing using real-time data. Model replacement will only be performed if the candidate model's performance consistently outperforms the current production model under preset evaluation metrics; otherwise, the system will automatically roll back to the original model to ensure system stability.

[0178] When actually carrying out model updates, the newly acquired field monitoring and experimental data must first be preprocessed. Field monitoring data must first be filtered to remove outliers caused by equipment malfunctions or signal interference. For example, if the structural surface displacement data collected by a certain sensor shows irregular and sudden changes within a short period and does not match the trend of data from adjacent sensors, it must be removed after confirmation based on field inspection records. For experimental data, the format must be standardized and validity verified according to the same standards as the initial database to ensure it matches the model's input requirements. For example, the unit of the shear stress data obtained from the experiment must be standardized to Pascals to maintain consistency with the units used during model training.

[0179] After data preprocessing, the online random forest algorithm is launched for incremental updates of the hybrid prediction model. The core structure of the online random forest algorithm consists of multiple independent decision trees. During incremental updates, the entire forest of decision trees is not reconstructed; instead, the prediction error of each decision tree is calculated based on the new data. For decision trees whose errors exceed a preset range, the algorithm corrects them by adjusting parameters such as the node splitting threshold and the output weights of leaf nodes. For entirely new feature combinations appearing in the new data, the algorithm adaptively generates a small number of new decision trees to add to the forest, capturing patterns in these new scenarios. Throughout the process, model parameter adjustments are entirely based on the new data; historical data serves only as the foundation for the model's original decision logic and does not need to be repeatedly used in the calculations.

[0180] The specific logic for parameter adjustment revolves around the output target of the hybrid prediction model. For example, when new experimental data shows that the shear strength of a structural surface with a specific lithology is higher than the model's original prediction, the online random forest algorithm will locate the decision node in the model related to the lithology parameter. Through feature feedback from the new data, the judgment threshold of that node will be adjusted, enabling the model to output a more realistic strength prediction result when encountering similar lithology parameters in the future. If the new data involves a completely new seismic load condition, the algorithm will establish the correlation between the input features and the output strength under that condition by adding a new decision tree, avoiding prediction bias caused by the original decision tree not covering the scenario.

[0181] After parameter adjustments are completed, the incrementally updated hybrid prediction model needs to be validated. The updated model is applied to the prediction of some newly added data, and the degree of agreement between the prediction results and the measured data is compared. If the degree of agreement meets the preset requirements, it indicates that the model parameter adjustments are reasonable and it can be used in subsequent engineering applications. If the degree of agreement is insufficient, the parameter settings of the data preprocessing process or algorithm need to be re-examined, such as adjusting the error judgment threshold of the decision tree, to ensure that the model update effect meets the engineering requirements.

[0182] Extreme working conditions refer to scenarios beyond the coverage of conventional data, such as strong earthquakes and special lithology. In these scenarios, the strength patterns of structural surfaces tend to deviate from the norm. Transfer learning is an algorithm that uses data from similar working conditions to train a basic model, and then fine-tunes it with a small amount of extreme working condition data, thus compensating for the lack of extreme data. Real-time ground motion monitoring data consists of real-time data collected by stations around the target site, which can capture the dynamic changes in ground motion. The dynamic correction coefficient is an adjustment parameter calculated based on the deviation between real-time data and historical data.

[0183] In practice, for adaptation to extreme working conditions, we first select extreme working condition data with similar lithology to the target site from the industry database and train the transfer learning base model. This model is then integrated with the existing hybrid prediction model, and the model parameters are fine-tuned using a small amount of existing extreme data from the target site. This allows the model to quickly grasp the intensity patterns under extreme scenarios, avoiding the inefficiency of training from scratch.

[0184] For real-time data fusion, a real-time data interface was established with seismic monitoring stations around the site, synchronizing real-time ground motion parameters hourly. Deviations between real-time and historical data were compared, and a dynamic correction coefficient was calculated. When the real-time peak ground acceleration was 10% higher than the historical average, the correction coefficient was increased proportionally to correct the dynamic attenuation factor. The corrected attenuation factor was updated to the model in real time, ensuring that changes in ground motion parameters were promptly reflected in the prediction results.

[0185] Meanwhile, a working condition identification module was added to the model to automatically determine whether the current scenario is an extreme working condition. If it is, the model branch optimized by transfer learning is called first. If the real-time data deviation is detected to exceed the threshold, the attenuation factor correction process is triggered immediately. This improvement allows the hybrid prediction model to maintain accuracy in extreme scenarios without relying on a large amount of data. The dynamic integration of real-time data solves the lag problem of static input of seismic motion parameters, making the prediction results more consistent with real-time working conditions and further improving the adaptability of the solution to complex engineering environments.

[0186] The above description is merely a preferred embodiment and the technical principles employed in this application. This application is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions that can be made by those skilled in the art will not depart from the scope of protection of this application. Therefore, although this application has been described in detail through the above embodiments, this application is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of this application.

Claims

1. A method for predicting the strength of rock mass structural surfaces driven by seismic loads, characterized in that, Includes the following steps: S1. Acquire and integrate structural geometric and mechanical parameters, seismic motion parameters, experimental data and numerical simulation data to form a standardized database; S2. Clean and standardize the data in the standardized database, construct a dynamic attenuation factor, screen and enhance key feature variables, and obtain initial input features; S3. Construct a physical empirical benchmark model based on the key feature variables and the dynamic attenuation factor, and generate physical model prediction features through cross-validation; construct a machine learning predictor based on the initial input features, and train the physical model prediction features and the initial input features trained by the machine learning predictor to obtain a hybrid prediction model. S4. Validate and perform sensitivity analysis on the hybrid prediction model to quantify the uncertainty of the output results of the hybrid prediction model; S5. By inputting the measured data, the hybrid prediction model is invoked to complete the calculation and output the prediction result.

2. The method according to claim 1, characterized in that, S1 specifically includes: S11. Three-dimensional point cloud data of rock mass structural surfaces are acquired using a non-contact measurement method, and the geometric parameters of the structural surfaces are extracted from the three-dimensional point cloud data; the mechanical parameters of the structural surfaces are obtained through in-situ direct shear tests. S12. Based on the geographical location information of the predicted site, obtain the ground motion parameters from an open-source seismic data platform; S13. Collect standardized cyclic shear test data, including correlation data of the number of cycles, cyclic stress amplitude, normal stress, shear displacement, and peak shear strength; generate numerical simulation data using the discrete element method or the finite element method. S14. Integrate the geometric and mechanical parameters of the structural surface, the seismic motion parameters, the standardized cyclic shear test data and the numerical simulation data to form the standardized database.

3. The method according to claim 1, characterized in that, S2 specifically includes: S21. Clean the data in the standardized database, remove abnormal data and fill in missing data; then standardize the cleaned data. S22. Based on the peak parameters, spectral parameters and cyclic load parameters in the ground motion parameters, a dynamic attenuation factor is constructed. The dynamic attenuation factor comprehensively reflects the cumulative damage effect of the cyclic load of the ground motion, the instantaneous impact effect of the peak parameters of the ground motion, and the response effect of the spectral parameters of the ground motion. S23. The feature variables in the standardized database are screened using a feature importance evaluation algorithm to obtain the key feature variables; the key feature variables are enhanced by constructing interactive features based on the key feature variables to obtain the initial input features.

4. The method according to claim 1, characterized in that, S3 specifically includes: S31. Based on the key feature variables and the dynamic decay factor, the classic Barton-Bandis model is dynamically improved by incorporating the dynamic decay factor into the classic Barton-Bandis model to construct the physical empirical benchmark model; a K-fold cross-validation strategy is used to generate physical model prediction features. S32. Construct the machine learning predictor, which includes a first-layer random forest regression model and a second-layer artificial neural network model, both of which are trained using the initial input features as input. The random forest regression model is used to output the peak shear strength prediction value, and the artificial neural network model is used to output the shear stress-shear displacement curve. S33. Extract the physical model prediction features output by the physical experience benchmark model as the new features, combine the new features with the initial input features after training to form extended input features, and input the extended input features into the first layer random forest regression model for secondary training to obtain the hybrid prediction model.

5. The method according to claim 1, characterized in that, S4 specifically includes: S41. Divide the data in the standardized database into a training set, a validation set, and a test set. Use cross-validation to train and validate the hybrid prediction model. Use the coefficient of determination, root mean square error, and mean absolute error to comprehensively evaluate the model's prediction performance. If the evaluation results meet the preset performance requirements, then the hybrid prediction model is confirmed to be usable. S42. The global sensitivity analysis method is used to quantify the contribution of each input feature to the variance of the model output result, and the local sensitivity analysis method is used to analyze the nonlinear influence of key input features on the model output result in different intervals. The sensitivity analysis results are obtained, and it is determined that the sensitivity analysis results meet the engineering practice and the preset analysis requirements. S43. Treat the input parameters that affect the prediction results of the mixed prediction model as random variables and assign them corresponding probability distributions. Use the Monte Carlo simulation method to perform propagation analysis on the uncertainty of the input parameters, and obtain the probability density function, cumulative distribution function and confidence interval of the output results of the mixed prediction model, thus completing the quantification of the uncertainty of the output results.

6. The method according to claim 4, characterized in that, S5 specifically includes: S51. Input three types of basic data, including structural surface geometry and roughness parameters, site ground motion parameters, and normal stress conditions. S52. Calculate the equivalent number of cycles based on the duration and dominant frequency in the site ground motion parameters, calculate the dynamic attenuation factor, run the physical empirical benchmark model and the hybrid prediction model in parallel, obtain the corrected peak shear strength, and generate a complete shear stress-shear displacement curve through the artificial neural network model. S53. Output the prediction results, which include the determined value and confidence interval of the peak shear strength, the comparison chart of shear stress-shear displacement curves, the sensitivity analysis chart, the uncertainty cloud map, and the structured technical report.

7. The method according to claim 1, characterized in that, It also includes the following steps: S6. When new field monitoring data or experimental data are obtained, the hybrid prediction model is incrementally updated using an online random forest algorithm; wherein, during the update process, the parameters of the hybrid prediction model are adjusted based on the new data.

8. The method according to claim 3, characterized in that, S23 specifically includes: S231. Extract the lithological parameters and stress level parameters from the standardized database as the basis for working condition classification, and divide the data in the standardized database into multiple subsets of different working condition types. S232. For each of the aforementioned subsets, a feature importance evaluation algorithm is used to screen the key feature variables that are suitable for the corresponding working conditions, thereby obtaining a set of exclusive key features for each working condition. S233. Based on the exclusive key feature set under each working condition, construct interactive features that conform to the mechanical properties of the corresponding working condition. The construction of the interactive features is based on the dominant influence mechanism of the structural surface strength under the corresponding working condition. S234. Integrate the exclusive key feature set and corresponding interaction features under each working condition to form the initial input features adapted to all working conditions.

9. The method according to claim 5, characterized in that, In S43, the step of treating the input parameters affecting the prediction results of the hybrid prediction model as random variables and assigning them corresponding probability distributions specifically includes: S431. Collect on-site measured data of the structural surfaces of the target site, regional geological statistics data, and related data of similar projects. The input parameters include normal stress, roughness coefficient, and spectral acceleration. S432. Using parameter estimation methods, the collected measured data, statistical data and related data of similar projects are fitted to obtain the initial probability distribution type and corresponding distribution parameters of each input parameter. S433. Verify the fit between the initial probability distribution and the measured data through the KS test or AD test. If the fit does not meet the preset fit threshold, adjust the distribution type or optimize the distribution parameters until an input parameter probability distribution that matches the actual situation of the target site is obtained. S434. Assign the calibrated probability distribution to the corresponding input parameters, and then perform the subsequent uncertainty propagation analysis using the Monte Carlo simulation method.

10. The method according to claim 3, characterized in that, S22 specifically includes: S221. The acquired ground motion parameters are identified by type. Based on the pulse characteristics, spectral characteristics and propagation distance of the ground motion, the ground motion parameters are divided into two types of ground motion sub-parameters: near-field pulse type and far-field stable type. S222. For near-field pulsed ground motion parameters, a first dynamic attenuation factor is constructed based on pulse peak value and pulse duration. The first dynamic attenuation factor focuses on reflecting the instantaneous impact damage effect of pulsed load. S223. For far-field stable ground motion parameters, a second dynamic attenuation factor is constructed based on the cyclic stress amplitude and the equivalent number of cycles. The second dynamic attenuation factor focuses on reflecting the cumulative damage effect of cyclic load. S224. The first dynamic attenuation factor and the second dynamic attenuation factor are associated with and stored with the corresponding seismic sub-parameters, respectively; when constructing the physical empirical benchmark model later, the corresponding dynamic attenuation factor is called according to the seismic motion type of the target site.

Images