2023 Vol. 48, No. 5
Display Method:
2023, 48(5): 1657-1674.
doi: 10.3799/dqkx.2022.419
PDF 5639KB
Abstract:
Landslide is prone to occur frequently with many aspects, wide coverage, and serious threats. Therefore, based on "satellite-aerial-surface-deep" collaborative integrated big data observations, it is of great significance to carry out early landslide identification, susceptibility assessment, and prediction to ensure the safety of people's lives and property, and to promote the modernization of landslide disaster prevention capability. At present, the landslide early identification relying on manual interpretation is time-consuming and labor-intensive, the landslide susceptibility assessment using a heuristic model cannot better prove the nonlinear relationship between environmental factors, and the landslide prediction accuracy based on single monitoring data is low. The machine learning algorithm is gradually widely used in landslide disaster prevention and mitigation because of its strong nonlinear processing ability and robustness. Based on this, the study systematically expounds on the specific application of machine learning in early identification, susceptibility assessment, and prediction of landslide disasters, summarizes the advantages and disadvantages of various machine learning algorithms in the above three fields, and finally prospects the future development trend of machine learning in landslide disasters.
2023, 48(5): 1675-1685.
doi: 10.3799/dqkx.2023.034
Abstract:
Many loess landslides were caused by the agricultural irrigation in Heifangtai terrace, Yongjing County, Gansu Province. Their stability analysis and critical slip surface identification are particularly important, as it can provide a good support to disaster prevention. The loess landslides located on the margin of Heifangtai terrace have the characteristics of progressive backward failure, and the occurred and potential landslides are highly similar. The results from back analysis can provide important data basis for future landslide stability prediction. In this paper, the finite difference strength reduction method was used to calibrate the cohesion and internal friction angle of loess based on the NSGA-II genetic algorithm by setting three objective optimization functions (i.e., mean value of soil strength parameters, slip surface and factor of safety). Taking Dangchuan 2# landslide in Heifangtai terrace as a study case, based on the slip surface observed after the first slide and assuming its factor of safety was equal to 1, the back analysis results show that the cohesion and internal friction angle of natural loess were 28.20 kPa and 25.16°; and the effective cohesion and internal friction angle of saturated loess were 16.59 kPa and 16.11°. Based on the computed results, the factor of safety and critical slip surface of the three subsequent slides were predicted, with their comparison with in-site observation information. The results show that a more reasonable estimation of loess strength parameters can be obtained by using the multi-objective constraint optimization algorithm, which provides a new solution for the stability analysis and quantitative risk assessment of landslides in Heifangtai terrace.
Many loess landslides were caused by the agricultural irrigation in Heifangtai terrace, Yongjing County, Gansu Province. Their stability analysis and critical slip surface identification are particularly important, as it can provide a good support to disaster prevention. The loess landslides located on the margin of Heifangtai terrace have the characteristics of progressive backward failure, and the occurred and potential landslides are highly similar. The results from back analysis can provide important data basis for future landslide stability prediction. In this paper, the finite difference strength reduction method was used to calibrate the cohesion and internal friction angle of loess based on the NSGA-II genetic algorithm by setting three objective optimization functions (i.e., mean value of soil strength parameters, slip surface and factor of safety). Taking Dangchuan 2# landslide in Heifangtai terrace as a study case, based on the slip surface observed after the first slide and assuming its factor of safety was equal to 1, the back analysis results show that the cohesion and internal friction angle of natural loess were 28.20 kPa and 25.16°; and the effective cohesion and internal friction angle of saturated loess were 16.59 kPa and 16.11°. Based on the computed results, the factor of safety and critical slip surface of the three subsequent slides were predicted, with their comparison with in-site observation information. The results show that a more reasonable estimation of loess strength parameters can be obtained by using the multi-objective constraint optimization algorithm, which provides a new solution for the stability analysis and quantitative risk assessment of landslides in Heifangtai terrace.
2023, 48(5): 1686-1695.
doi: 10.3799/dqkx.2023.029
Abstract:
Rock compressive strength is an important mechanical parameter to evaluate the stability of rock mass engineering. The traditional statistical regression method has some limitations on the prediction of rock compressive strength. To this end, in this paper it proposes a method for intelligent prediction of rock compressive strength using simple rock mechanics parameters. Firstly, 620 sets of triaxial test data containing different types of rocks were collected and preprocessed. Then, three main stream ensemble learning algorithms, Random forest, XGBoost and LightGBM, were used to establish a rock compressive strength prediction model, and Bayesian optimization algorithm was used to optimize the hyperparameters during model training. Finally, the coefficient of determination (R2), mean absolute percentage error (MAPE) and root mean square error (RMSE) were used to evaluate and compare the generalization ability of the optimized model. In addition, the importance of input features was analyzed by LGB model, to evaluate the importance of input features on the generalization ability of the model. The results show that the three models have achieved good prediction results for rock compressive strength. And the generalization ability of the LGB model is slightly better than that of the other two models (R2 = 0.978, RMSE=5.58, MAPE=9.70%), and the running time is relatively minimum. Elastic modulus (E), confining pressure (σ3) and density (ρ) have great influence on generalization ability of model, while Poisson's ratio(v)has little influence. The prediction model has good applicability to rock strength prediction, and provides a new idea for the combination of machine learning and geotechnical engineering.
Rock compressive strength is an important mechanical parameter to evaluate the stability of rock mass engineering. The traditional statistical regression method has some limitations on the prediction of rock compressive strength. To this end, in this paper it proposes a method for intelligent prediction of rock compressive strength using simple rock mechanics parameters. Firstly, 620 sets of triaxial test data containing different types of rocks were collected and preprocessed. Then, three main stream ensemble learning algorithms, Random forest, XGBoost and LightGBM, were used to establish a rock compressive strength prediction model, and Bayesian optimization algorithm was used to optimize the hyperparameters during model training. Finally, the coefficient of determination (R2), mean absolute percentage error (MAPE) and root mean square error (RMSE) were used to evaluate and compare the generalization ability of the optimized model. In addition, the importance of input features was analyzed by LGB model, to evaluate the importance of input features on the generalization ability of the model. The results show that the three models have achieved good prediction results for rock compressive strength. And the generalization ability of the LGB model is slightly better than that of the other two models (R2 = 0.978, RMSE=5.58, MAPE=9.70%), and the running time is relatively minimum. Elastic modulus (E), confining pressure (σ3) and density (ρ) have great influence on generalization ability of model, while Poisson's ratio(v)has little influence. The prediction model has good applicability to rock strength prediction, and provides a new idea for the combination of machine learning and geotechnical engineering.
2023, 48(5): 1696-1710.
doi: 10.3799/dqkx.2022.247
Abstract:
To address the problems of landslide susceptibility prediction (LSP) modeling including possible errors in landslide and non-landslide samples, complex non-linear relationships between environmental factors and unaddressed machine learning interpretability, a deep learning-based Self-screening Bi-directional Long Short-Term Memory and Conditional Random Fields (SBiLSTM-CRF) model is proposed to reduce the impact of these problems on LSP and improve its confidence. The SBiLSTM-CRF model has the advantages of deep learning network with deep layers, wide width and iterative modeling, which can predict the non-linear relationship between environmental factors and automatically screen out the wrong landslide samples; it can select non-landslide samples from the initial low/very low landslide susceptibility zone through iterative modeling, and finally reveal the internal mechanism of the coupling of environmental factors to predict landslide susceptibility. The SBiLSTM-CRF model is used to predict landslide susceptibility in Yanchang County of China, and compared with cpLSTM-CRF, random forest (RF), support vector machine (SVM), stochastic gradient descent (SGD) and logistic regression (LR) models. The results show that SBiLSTM-CRF overcomes the problems of sample error and complex nonlinear relationship between factors in traditional machine learning, has superior performance in modeling susceptibility than conventional machine learning, and the interpretability of the model reveals that factors such as slope, elevation and lithology control the development of mounded landslides in Yanchang County.
To address the problems of landslide susceptibility prediction (LSP) modeling including possible errors in landslide and non-landslide samples, complex non-linear relationships between environmental factors and unaddressed machine learning interpretability, a deep learning-based Self-screening Bi-directional Long Short-Term Memory and Conditional Random Fields (SBiLSTM-CRF) model is proposed to reduce the impact of these problems on LSP and improve its confidence. The SBiLSTM-CRF model has the advantages of deep learning network with deep layers, wide width and iterative modeling, which can predict the non-linear relationship between environmental factors and automatically screen out the wrong landslide samples; it can select non-landslide samples from the initial low/very low landslide susceptibility zone through iterative modeling, and finally reveal the internal mechanism of the coupling of environmental factors to predict landslide susceptibility. The SBiLSTM-CRF model is used to predict landslide susceptibility in Yanchang County of China, and compared with cpLSTM-CRF, random forest (RF), support vector machine (SVM), stochastic gradient descent (SGD) and logistic regression (LR) models. The results show that SBiLSTM-CRF overcomes the problems of sample error and complex nonlinear relationship between factors in traditional machine learning, has superior performance in modeling susceptibility than conventional machine learning, and the interpretability of the model reveals that factors such as slope, elevation and lithology control the development of mounded landslides in Yanchang County.
2023, 48(5): 1711-1729.
doi: 10.3799/dqkx.2023.006
Abstract:
Due to the complex interaction between geological environment and human activities, the Chinese Loess Plateau (CLP) is prone to frequent landslides. It is urgent to carry out landslide vulnerability assessment, selecting suitable influencing factors and training models. In this study, the CLP was taken as the study area. Based on field landslide survey and data collection, an evaluation system including topography, basic geological environment, meteorology and hydrology, human activities, soil physical and chemical properties, and vegetation coverage was built. The information model (IV) was used to connect the random forest model (RF) and convolutional neural network model (CNN) to build coupling models IV-RF and IV-CNN, and landslide susceptibility evaluation research was carried out. The results show that the accuracy of the coupling model (IV-RF, IV-CNN) is higher than that of the independent model (RF, CNN), and the AUC values of the four models are 0.916, 0.938, 0.878, and 0.853, respectively. The IV-CNN has stronger prediction ability and accuracy. The areas of extremely high, high, medium, low, and extremely low vulnerability areas in the IV-CNN model account for 8.78%, 7.47%, 15.34%, 19.82%, and 47.87% respectively, which are mainly distributed in the mountainous and loess hilly areas with complex geological environment and strong human activities in the south and east of the loess plateau. Slope, erosion type, landform type, clay content and distance from the road rank in the top five in the contribution rate analysis, and are the main control factors affecting the landslide development. The purpose of this study is to provide reliable scientific basis for the prediction and prevention of landslide disasters in the CLP, deepen the modeling idea for landslide vulnerability evaluation research, and optimize the uncertainty of independent model evaluation results.
Due to the complex interaction between geological environment and human activities, the Chinese Loess Plateau (CLP) is prone to frequent landslides. It is urgent to carry out landslide vulnerability assessment, selecting suitable influencing factors and training models. In this study, the CLP was taken as the study area. Based on field landslide survey and data collection, an evaluation system including topography, basic geological environment, meteorology and hydrology, human activities, soil physical and chemical properties, and vegetation coverage was built. The information model (IV) was used to connect the random forest model (RF) and convolutional neural network model (CNN) to build coupling models IV-RF and IV-CNN, and landslide susceptibility evaluation research was carried out. The results show that the accuracy of the coupling model (IV-RF, IV-CNN) is higher than that of the independent model (RF, CNN), and the AUC values of the four models are 0.916, 0.938, 0.878, and 0.853, respectively. The IV-CNN has stronger prediction ability and accuracy. The areas of extremely high, high, medium, low, and extremely low vulnerability areas in the IV-CNN model account for 8.78%, 7.47%, 15.34%, 19.82%, and 47.87% respectively, which are mainly distributed in the mountainous and loess hilly areas with complex geological environment and strong human activities in the south and east of the loess plateau. Slope, erosion type, landform type, clay content and distance from the road rank in the top five in the contribution rate analysis, and are the main control factors affecting the landslide development. The purpose of this study is to provide reliable scientific basis for the prediction and prevention of landslide disasters in the CLP, deepen the modeling idea for landslide vulnerability evaluation research, and optimize the uncertainty of independent model evaluation results.
Efficient Identification Method of Coastal Soft Soil Stratum Based on Hierarchical Bayesian Learning
2023, 48(5): 1730-1741.
doi: 10.3799/dqkx.2022.503
Abstract:
Underground stratification based on the cone penetration test (CPT) data relies on empirical charts and subjective judgments. Uncertainties in soil profile identified from CPT data are unavoidable. In this paper it proposes a fast underground stratigraphy identification method based on soil classification index Ic, which also quantifies the identification uncertainty. Under the hierarchical Bayesian learning framework, the proposed method uses full Gaussian random field model to characterize the spatial variability of soils, and efficiently calculates the likelihood function of soil layer thicknesses given a soil layer number by introducing the Normal-Inverse Wishart conjugate distribution. By this means, the model evidence can be solved efficiently, and the most possible profile is identified. The proposed method automatically identifies the most probable soil layer profile considering the statistical characteristics of Ic, and improves the reliability of the identification results.
Underground stratification based on the cone penetration test (CPT) data relies on empirical charts and subjective judgments. Uncertainties in soil profile identified from CPT data are unavoidable. In this paper it proposes a fast underground stratigraphy identification method based on soil classification index Ic, which also quantifies the identification uncertainty. Under the hierarchical Bayesian learning framework, the proposed method uses full Gaussian random field model to characterize the spatial variability of soils, and efficiently calculates the likelihood function of soil layer thicknesses given a soil layer number by introducing the Normal-Inverse Wishart conjugate distribution. By this means, the model evidence can be solved efficiently, and the most possible profile is identified. The proposed method automatically identifies the most probable soil layer profile considering the statistical characteristics of Ic, and improves the reliability of the identification results.
2023, 48(5): 1742-1754.
doi: 10.3799/dqkx.2022.502
Abstract:
Flood disasters can cause great damage to buildings, and identification of post-disaster house damage grade is very important to ensure people's safety. However, the traditional method of artificial identification costs a lot of manpower, financial resources, and time. Based on the data of rural housing damage caused by the "7.20" heavy rainstorm in Zhengzhou, Henan Province, the deep Convolutional neural network (CNN) theory was used to obtain the intelligent classification model of post-disaster house damage grade. Four classic deep CNN architectures, AlexNet, VGGNet, GoogleNet, and ResNet, were used to train, validate and test the data sets, and four intelligent classification models of post-disaster house damage grade were obtained. The weights of the pre-training model were fine-tuned to improve the generalization of the model, then the ResNet-50 with the transfer learning was selected as the main model of the classification. Finally, the influence of hyperparameters in CNN architecture was analyzed. The results show that when the learning rate of the ResNet network was 0.000 5, the epoch was 50, and the batch_size was 16, the network training result was optimal, and the prediction accuracy of the test set reached 95.5%; the visual analysis of the characteristics of the housing risk level clarified the mechanism and accuracy of the model classification. Experimental results show that the ensemble model had a high accuracy rate, which provided an idea and explored an approach for classification of damaged grade rural houses after flood disaster.
Flood disasters can cause great damage to buildings, and identification of post-disaster house damage grade is very important to ensure people's safety. However, the traditional method of artificial identification costs a lot of manpower, financial resources, and time. Based on the data of rural housing damage caused by the "7.20" heavy rainstorm in Zhengzhou, Henan Province, the deep Convolutional neural network (CNN) theory was used to obtain the intelligent classification model of post-disaster house damage grade. Four classic deep CNN architectures, AlexNet, VGGNet, GoogleNet, and ResNet, were used to train, validate and test the data sets, and four intelligent classification models of post-disaster house damage grade were obtained. The weights of the pre-training model were fine-tuned to improve the generalization of the model, then the ResNet-50 with the transfer learning was selected as the main model of the classification. Finally, the influence of hyperparameters in CNN architecture was analyzed. The results show that when the learning rate of the ResNet network was 0.000 5, the epoch was 50, and the batch_size was 16, the network training result was optimal, and the prediction accuracy of the test set reached 95.5%; the visual analysis of the characteristics of the housing risk level clarified the mechanism and accuracy of the model classification. Experimental results show that the ensemble model had a high accuracy rate, which provided an idea and explored an approach for classification of damaged grade rural houses after flood disaster.
2023, 48(5): 1755-1768.
doi: 10.3799/dqkx.2022.491
Abstract:
In order to distinguish the sensitivity factors affecting rockburst and construct a rockburst prediction method under the condition of incomplete data cases, a large sample database is established on the basis of collecting 429 groups of rockburst cases at home and abroad, and the distribution characteristics and regulation of rockburst disaster-inducing factors were summarized. Six evaluation indexes, including buried depth, uniaxial compressive strength of rock, uniaxial tensile strength of rock, maximum tangential stress of surrounding rock, rock elastic energy index and integrity coefficient of rock mass, are selected to establish a rockburst probability prediction model based on large and incomplete data set by using Bayesian network, and the sensitivity analysis and engineering application are carried out. Through analysis, it is found that the maximum tangential stress of surrounding rock and the integrity coefficient of rock mass have a great influence on rockburst. The model has a prediction coincidence rate of 83.3% for rockburst cases with information loss rate of 20%, and the prediction effect is better than the commonly used empirical criterion of rockburst. The results show that the prediction indexes selected in this paper can comprehensively consider the influencing factors of rockburst, and the established model has applicability and reliability for the prediction of deep rockburst disasters.
In order to distinguish the sensitivity factors affecting rockburst and construct a rockburst prediction method under the condition of incomplete data cases, a large sample database is established on the basis of collecting 429 groups of rockburst cases at home and abroad, and the distribution characteristics and regulation of rockburst disaster-inducing factors were summarized. Six evaluation indexes, including buried depth, uniaxial compressive strength of rock, uniaxial tensile strength of rock, maximum tangential stress of surrounding rock, rock elastic energy index and integrity coefficient of rock mass, are selected to establish a rockburst probability prediction model based on large and incomplete data set by using Bayesian network, and the sensitivity analysis and engineering application are carried out. Through analysis, it is found that the maximum tangential stress of surrounding rock and the integrity coefficient of rock mass have a great influence on rockburst. The model has a prediction coincidence rate of 83.3% for rockburst cases with information loss rate of 20%, and the prediction effect is better than the commonly used empirical criterion of rockburst. The results show that the prediction indexes selected in this paper can comprehensively consider the influencing factors of rockburst, and the established model has applicability and reliability for the prediction of deep rockburst disasters.
2023, 48(5): 1769-1779.
doi: 10.3799/dqkx.2022.481
Abstract:
Since the existing earthquake damage prediction methods cannot make rapid predictions for brick masonry structures. A rapid prediction method for earthquake damage of brick masonry structures is proposed. The method uses a machine learning model, considering the ground motion characteristics and structural characteristics. 12 ground motion parameters that represent the ground motion characteristics and 7 structural parameters that have a strong correlation with the damage of brick masonry structures are selected. The ground motion parameters are considered in four aspects: time domain, frequency domain, response spectrum and holding time, and the structural parameters are considered in terms of bearing capacity and stiffness. Three machine learning models based on support vector machine, random forest and artificial neural network are given for fast prediction of seismic damage of brick masonry structures. The input parameters were further optimized using correlation analysis, and the optimal model after optimizing the input parameters was given. The results show that the ANN model has the highest prediction accuracy of 91.56% when 19 input parameters were used. The prediction accuracy of the RF model-based earthquake damage prediction method was higher when 12 optimized parameters were used as inputs, reaching 90.01%. The prediction performance of the RF-based model was more stable when the input parameters were gradually reduced. The optimized input parameters of the RF model-based prediction method can achieve rapid prediction of seismic damage to brick masonry structures. The method that considers both structural and ground vibration parameters as input greatly improves the accuracy of prediction compared to the method that considers only structural parameters or only ground vibration parameters as input.
Since the existing earthquake damage prediction methods cannot make rapid predictions for brick masonry structures. A rapid prediction method for earthquake damage of brick masonry structures is proposed. The method uses a machine learning model, considering the ground motion characteristics and structural characteristics. 12 ground motion parameters that represent the ground motion characteristics and 7 structural parameters that have a strong correlation with the damage of brick masonry structures are selected. The ground motion parameters are considered in four aspects: time domain, frequency domain, response spectrum and holding time, and the structural parameters are considered in terms of bearing capacity and stiffness. Three machine learning models based on support vector machine, random forest and artificial neural network are given for fast prediction of seismic damage of brick masonry structures. The input parameters were further optimized using correlation analysis, and the optimal model after optimizing the input parameters was given. The results show that the ANN model has the highest prediction accuracy of 91.56% when 19 input parameters were used. The prediction accuracy of the RF model-based earthquake damage prediction method was higher when 12 optimized parameters were used as inputs, reaching 90.01%. The prediction performance of the RF-based model was more stable when the input parameters were gradually reduced. The optimized input parameters of the RF model-based prediction method can achieve rapid prediction of seismic damage to brick masonry structures. The method that considers both structural and ground vibration parameters as input greatly improves the accuracy of prediction compared to the method that considers only structural parameters or only ground vibration parameters as input.
2023, 48(5): 1780-1792.
doi: 10.3799/dqkx.2022.450
Abstract:
Accurate assessment of soils' preconsolidation stress (PS) is important in geotechnical engineering practice. In this paper it analyzes the influence of soils' preconsolidation stress, uses ensemble learning algorithms (XGBoost, RF) to capture the relationship between soil parameters and establishes prediction models. A Bayesian optimization method was used to determine the optimal parameters of the models, three machin elearning algorithms, namely SVR, KNN, and MLP, are introduced for comparison, and the models were statistically analyzed by three error metrics, including correlation coefficient(R2), root mean square error (RMSE) and mean absolute percentage error (MAPE). And finally, the prediction accuracy and generalization of each model were evaluated under 5-fold cross-validation (CV). The XGBoost-based prediction accuracy is the highest, with RMSE and MAPE of 20.80 kPa and 18.29%, respectively, followed by RF with 24.532 kPa and 19.15%, respectively. Meanwhile, in the case of PS as a regression variable, its characteristic importance is USS > VES > w > LL > PL. It shows that the ensemble learning algorithms (XGBoost, RF) are better than other algorithms in terms of prediction accuracy and generalization in the case of small-scale data sets, and can be used as an effective method for sensitivity analysis of geotechnical parameters.
Accurate assessment of soils' preconsolidation stress (PS) is important in geotechnical engineering practice. In this paper it analyzes the influence of soils' preconsolidation stress, uses ensemble learning algorithms (XGBoost, RF) to capture the relationship between soil parameters and establishes prediction models. A Bayesian optimization method was used to determine the optimal parameters of the models, three machin elearning algorithms, namely SVR, KNN, and MLP, are introduced for comparison, and the models were statistically analyzed by three error metrics, including correlation coefficient(R2), root mean square error (RMSE) and mean absolute percentage error (MAPE). And finally, the prediction accuracy and generalization of each model were evaluated under 5-fold cross-validation (CV). The XGBoost-based prediction accuracy is the highest, with RMSE and MAPE of 20.80 kPa and 18.29%, respectively, followed by RF with 24.532 kPa and 19.15%, respectively. Meanwhile, in the case of PS as a regression variable, its characteristic importance is USS > VES > w > LL > PL. It shows that the ensemble learning algorithms (XGBoost, RF) are better than other algorithms in terms of prediction accuracy and generalization in the case of small-scale data sets, and can be used as an effective method for sensitivity analysis of geotechnical parameters.
2023, 48(5): 1793-1806.
doi: 10.3799/dqkx.2022.439
Abstract:
The hydrodynamic pressure-driven landslides in the Three Gorges reservoir area have the characteristics of stepped deformation,and it is difficult to complete the analysis and prediction of landslides accurately and reasonably under the condition of insufficient monitoring data. In view of insufficient monitoring data,a state affine transfer learning method (SATLM) was designed in this paper to analyze the state of landslides with insufficient data by learning similar landslide knowledge. In order to verify the effectiveness of SATLM in landslide state analysis,a state similarity analysis method was designed in this paper. After learning the knowledge of multiple landslides in the reservoir area,another landslide displacement prediction with insufficient data was realized.The results show that compared with BPNN and SVM,the mean absolute error and root mean square error of landslide displacement prediction of Wanzhou Tangjiao No.1 landslide are greatly reduced after state affine migration.The successful knowledge transfer of Baijiabao landslide,Baishuihe landslide,Bazimen landslide proves that the state affine transfer learning method has a good effect on the knowledge transfer of similar hydrodynamic pressure-driven landslides.
The hydrodynamic pressure-driven landslides in the Three Gorges reservoir area have the characteristics of stepped deformation,and it is difficult to complete the analysis and prediction of landslides accurately and reasonably under the condition of insufficient monitoring data. In view of insufficient monitoring data,a state affine transfer learning method (SATLM) was designed in this paper to analyze the state of landslides with insufficient data by learning similar landslide knowledge. In order to verify the effectiveness of SATLM in landslide state analysis,a state similarity analysis method was designed in this paper. After learning the knowledge of multiple landslides in the reservoir area,another landslide displacement prediction with insufficient data was realized.The results show that compared with BPNN and SVM,the mean absolute error and root mean square error of landslide displacement prediction of Wanzhou Tangjiao No.1 landslide are greatly reduced after state affine migration.The successful knowledge transfer of Baijiabao landslide,Baishuihe landslide,Bazimen landslide proves that the state affine transfer learning method has a good effect on the knowledge transfer of similar hydrodynamic pressure-driven landslides.
2023, 48(5): 1807-1824.
doi: 10.3799/dqkx.2022.429
Abstract:
In view of the lack of research about landslide intensity prediction in the current quantitative vulnerability analysis of buildings, in this paper it innovatively proposes a quantitative analysis method of the combination of intensity empirical curve based on InSAR technology and spatial displacement prediction of secondary development of ABAQUS.Taking the Shilongmen landslide in the Three Gorges Reservoir area as an example, the PS-InSAR method was adopted to calculate the cumulative displacement of the landslide in 2017‒2020 and obtained the empirical curve of landslide intensity. The ABAQUS software was used to compile the subroutine of load and pore water pressure to calculate the cumulative displacement under extreme scenarios (reservoir water level drop with heavy rainfall) and predicted the vulnerability of buildings. The evaluation system of resistance was constructed by weighting eight indicators of PSO-Fuzzy AHP model, which can be combined with the landslide intensity to quantitatively evaluate the vulnerability of buildings. The results indicate follows: (1) The evaluation system of resistance proposed in this paper can well present the structural characteristics of rural buildings in the Three Gorges Reservoir area, and has high evaluation accuracy. (2) The retrieved upper-intensity curve obtained by PS-InSAR is Ipu=0.065×Dtot0.236 which has higher prediction accuracy and effectively reduces false-negative errors. (3) The landslide intensity of extreme conditions simulated by ABAQUS increases with the increase of rainfall, the predicted vulnerability level of buildings increases, and the buildings with obvious deformation in the previous investigation are successfully warned. It is concluded that the landslide intensity prediction method and vulnerability analysis method proposed in this paper has high spatial identification and early warning accuracy, and real-time vulnerability mapping of buildings can be obtained through landslide intensity information.
In view of the lack of research about landslide intensity prediction in the current quantitative vulnerability analysis of buildings, in this paper it innovatively proposes a quantitative analysis method of the combination of intensity empirical curve based on InSAR technology and spatial displacement prediction of secondary development of ABAQUS.Taking the Shilongmen landslide in the Three Gorges Reservoir area as an example, the PS-InSAR method was adopted to calculate the cumulative displacement of the landslide in 2017‒2020 and obtained the empirical curve of landslide intensity. The ABAQUS software was used to compile the subroutine of load and pore water pressure to calculate the cumulative displacement under extreme scenarios (reservoir water level drop with heavy rainfall) and predicted the vulnerability of buildings. The evaluation system of resistance was constructed by weighting eight indicators of PSO-Fuzzy AHP model, which can be combined with the landslide intensity to quantitatively evaluate the vulnerability of buildings. The results indicate follows: (1) The evaluation system of resistance proposed in this paper can well present the structural characteristics of rural buildings in the Three Gorges Reservoir area, and has high evaluation accuracy. (2) The retrieved upper-intensity curve obtained by PS-InSAR is Ipu=0.065×Dtot0.236 which has higher prediction accuracy and effectively reduces false-negative errors. (3) The landslide intensity of extreme conditions simulated by ABAQUS increases with the increase of rainfall, the predicted vulnerability level of buildings increases, and the buildings with obvious deformation in the previous investigation are successfully warned. It is concluded that the landslide intensity prediction method and vulnerability analysis method proposed in this paper has high spatial identification and early warning accuracy, and real-time vulnerability mapping of buildings can be obtained through landslide intensity information.
2023, 48(5): 1825-1835.
doi: 10.3799/dqkx.2022.392
Abstract:
The northeastern part of Afghanistan is a typical cold and arid region where landslide geological hazards are developed. The landslide development is not only affected by topography, geological structure, human activities, and other factors, but also is controlled by snow cover, snow, and ice melt. In this paper, based on the primary data of remote sensing interpretation, considering the influence of snow cover and glacier activity on landslide development, two evaluation indexes of snow cover and ablation water equivalent were introduced to study the landslide risk in the cold and dry areas of the plateau. The landslide susceptibility evaluation system was established based on the weight of evidence and a fully connected neural network model. Degree-day model and SCS-CN model established the landslide risk evaluation system, and the evaluation model was tested according to the confusion matrix. The hazard assessment results show that the extremely high-risk area accounts for 10.46% of the total area, and the disaster area accounts for 82.71%, mainly distributed in the Kunar-Chitral reach in the east of Nuristan Province, the middle and eastern high mountains of Badakhshan Province except for Wakhan corridor section, and the Helmand Reach in Parwan Province. The high-risk area accounts for 14.83% of the total area, and the disaster area accounts for 12.11%, mainly distributed in the eastern region of Badakhshan Province, the western region of Nuristan Province, and Parwan Province. The test results and statistical results all show that the accuracy of the neural network is significantly improved by taking negative samples with a weight of evidence method. The research results can provide the scientific basis for Afghanistan's early warning and prevention of landslide geological disasters.
The northeastern part of Afghanistan is a typical cold and arid region where landslide geological hazards are developed. The landslide development is not only affected by topography, geological structure, human activities, and other factors, but also is controlled by snow cover, snow, and ice melt. In this paper, based on the primary data of remote sensing interpretation, considering the influence of snow cover and glacier activity on landslide development, two evaluation indexes of snow cover and ablation water equivalent were introduced to study the landslide risk in the cold and dry areas of the plateau. The landslide susceptibility evaluation system was established based on the weight of evidence and a fully connected neural network model. Degree-day model and SCS-CN model established the landslide risk evaluation system, and the evaluation model was tested according to the confusion matrix. The hazard assessment results show that the extremely high-risk area accounts for 10.46% of the total area, and the disaster area accounts for 82.71%, mainly distributed in the Kunar-Chitral reach in the east of Nuristan Province, the middle and eastern high mountains of Badakhshan Province except for Wakhan corridor section, and the Helmand Reach in Parwan Province. The high-risk area accounts for 14.83% of the total area, and the disaster area accounts for 12.11%, mainly distributed in the eastern region of Badakhshan Province, the western region of Nuristan Province, and Parwan Province. The test results and statistical results all show that the accuracy of the neural network is significantly improved by taking negative samples with a weight of evidence method. The research results can provide the scientific basis for Afghanistan's early warning and prevention of landslide geological disasters.
2023, 48(5): 1836-1852.
doi: 10.3799/dqkx.2022.372
Abstract:
To explore the influence of spatial variability of soil parameters with rotated anisotropy on probabilistic large slope deformation characteristics, a multiple response surface-based random material point method is proposed in this study. First, random field theory is employed to simulate the spatial variability of soil parameters with rotated anisotropy. Then, the multiple response surface method is used to evaluate the factor of safety (FS) of each random field sample, based on which the FSs for all random field samples are sorted efficiently in an ascending order. Finally, the random material point method is used to sequentially simulate the large deformation features of a slope for the failed samples with the FS at less than 1. An undrained clay slope is taken as the illustrative example, where the undrained shear strength of the soil is simulated as a rotated anisotropic random field. The effect of the rotational angle β and the autocorrelation length θ2 in the minor principal direction of the rotated anisotropic random field on the large deformation features and failure modes of the slope are systematically studied. The results show that the proposed method can be efficiently conducted for probabilistic analysis of large slope deformation. Both β and θ2 have significant effects on large slope deformation features and failure modes. In terms of large deformation features, the mean and standard deviation of the influence distance, sliding distance, and sliding volume increase with the increase of θ2. There might be four failure modes when considering the rotated anisotropic spatial variability of the undrained strength of the slope, among which the deep sliding mechanism and progressive failure are the two main probabilistic failure modes. Therefore, the proposed method provides an effective way for the probabilistic large slope deformation analysis as well as a good theoretical reference for an accurate risk assessment of slope stability.
To explore the influence of spatial variability of soil parameters with rotated anisotropy on probabilistic large slope deformation characteristics, a multiple response surface-based random material point method is proposed in this study. First, random field theory is employed to simulate the spatial variability of soil parameters with rotated anisotropy. Then, the multiple response surface method is used to evaluate the factor of safety (FS) of each random field sample, based on which the FSs for all random field samples are sorted efficiently in an ascending order. Finally, the random material point method is used to sequentially simulate the large deformation features of a slope for the failed samples with the FS at less than 1. An undrained clay slope is taken as the illustrative example, where the undrained shear strength of the soil is simulated as a rotated anisotropic random field. The effect of the rotational angle β and the autocorrelation length θ2 in the minor principal direction of the rotated anisotropic random field on the large deformation features and failure modes of the slope are systematically studied. The results show that the proposed method can be efficiently conducted for probabilistic analysis of large slope deformation. Both β and θ2 have significant effects on large slope deformation features and failure modes. In terms of large deformation features, the mean and standard deviation of the influence distance, sliding distance, and sliding volume increase with the increase of θ2. There might be four failure modes when considering the rotated anisotropic spatial variability of the undrained strength of the slope, among which the deep sliding mechanism and progressive failure are the two main probabilistic failure modes. Therefore, the proposed method provides an effective way for the probabilistic large slope deformation analysis as well as a good theoretical reference for an accurate risk assessment of slope stability.
2023, 48(5): 1853-1864.
doi: 10.3799/dqkx.2022.338
Abstract:
Real-time intensity prediction can estimate the maximum possible impact of an earthquake based on P-wave before the arrival of destructive seismic waves. Earthquake early warning targets can take measures to reduce the potential damage. Peak P-wave displacement amplitude is a parameter that effectively estimates the peak ground motion, however, it is difficult to fully characterize the information in ground motion by a single or multiple parameters. Meanwhile, the calculation of the parameter requires the determination of the time window size, and continuous prediction cannot be achieved. To solve the above problems, a prediction model based on long short-term memory network is proposed in this paper. The model is constructed based on K-NET data from 2010‒2021, and the MJMA 7.3 earthquake event in March 2022 is selected as a case to validate the model. The results show that the intensity can be predicted at each time step of the record after the P-wave arrival, and the accuracy in the test set is 96.47% at 3 seconds after P-wave arrival. The LSTM model proposed in this paper improves the accuracy and continuity of intensity prediction and can provide a scientific basis for earthquake early warning and emergency response.
Real-time intensity prediction can estimate the maximum possible impact of an earthquake based on P-wave before the arrival of destructive seismic waves. Earthquake early warning targets can take measures to reduce the potential damage. Peak P-wave displacement amplitude is a parameter that effectively estimates the peak ground motion, however, it is difficult to fully characterize the information in ground motion by a single or multiple parameters. Meanwhile, the calculation of the parameter requires the determination of the time window size, and continuous prediction cannot be achieved. To solve the above problems, a prediction model based on long short-term memory network is proposed in this paper. The model is constructed based on K-NET data from 2010‒2021, and the MJMA 7.3 earthquake event in March 2022 is selected as a case to validate the model. The results show that the intensity can be predicted at each time step of the record after the P-wave arrival, and the accuracy in the test set is 96.47% at 3 seconds after P-wave arrival. The LSTM model proposed in this paper improves the accuracy and continuity of intensity prediction and can provide a scientific basis for earthquake early warning and emergency response.
2023, 48(5): 1865-1874.
doi: 10.3799/dqkx.2022.327
Abstract:
Failure mechanism and reliability analysis of rainfall-induced slopes generally ignore the effects of field observation information, such as the observation that the slope keeps stable in natural conditions or after a historical rainfall event. In this paper, with an infinite slope model as an example, the BUS (Bayesian Updating with Subset simulation) method is adopted for the probabilistic back analysis of spatially variable hydraulic and shear strength parameters based on the field observation that the slope survived from a previous extreme rainfall event. The probabilities of slope failure under different rainfall durations are evaluated within the framework of Monte-Carlo simulation. The influence of ignoring/incorporating the field observation on the estimate of probability of slope failure is also investigated. The results indicate that the possibility of slope failing along the weak zones caused by the spatial variability of soil parameters can be effectively excluded through the probabilistic back analysis incorporating the field observation. Based on this, more realistic probability of slope failure induced by the rainfall can be produced. If the field observation that the slope survived from a previous extreme rainfall event is ignored, the probability of slope failure will be significantly overestimated, especially in the early stage of rainfall. The research outcomes provide a new perspective for interpreting the rainfall-induced slope failure mechanisms in the spatially variable soils.
Failure mechanism and reliability analysis of rainfall-induced slopes generally ignore the effects of field observation information, such as the observation that the slope keeps stable in natural conditions or after a historical rainfall event. In this paper, with an infinite slope model as an example, the BUS (Bayesian Updating with Subset simulation) method is adopted for the probabilistic back analysis of spatially variable hydraulic and shear strength parameters based on the field observation that the slope survived from a previous extreme rainfall event. The probabilities of slope failure under different rainfall durations are evaluated within the framework of Monte-Carlo simulation. The influence of ignoring/incorporating the field observation on the estimate of probability of slope failure is also investigated. The results indicate that the possibility of slope failing along the weak zones caused by the spatial variability of soil parameters can be effectively excluded through the probabilistic back analysis incorporating the field observation. Based on this, more realistic probability of slope failure induced by the rainfall can be produced. If the field observation that the slope survived from a previous extreme rainfall event is ignored, the probability of slope failure will be significantly overestimated, especially in the early stage of rainfall. The research outcomes provide a new perspective for interpreting the rainfall-induced slope failure mechanisms in the spatially variable soils.
2023, 48(5): 1875-1891.
doi: 10.3799/dqkx.2022.313
Abstract:
Knowledge graph is a knowledge system that formally describes entities and their interrelationships, and it plays an important role in emergency disaster relief and spatio-temporal prediction and decision making. In order to obtain geological hazard information implied in multi-source heterogeneous texts and investigate the impact caused by disasters, in this paper it analyzes many elements associated with the development process of geological hazards from the perspective of hazard chains, and proposes a top-down and bottom-up method of constructing a knowledge graph of geological hazard chains.Firstly, the complex formation mechanism and chain formation law among geohazards are analyzed from the perspective of hazard chain, and a unified semantic expression framework of geohazard chain ontology for information extraction is established based on the top-down method, including geohazard event ontology describing the geohazard knowledge system, geoenvironment ontology carrying the occurrence of geohazards, geographic object ontology under the effect of geohazards, and geohazards in disaster. Secondly, the data layer is constructed by combining the bottom-up method, and the concepts, instances and attributes and their associated relations are identified through knowledge fusion and knowledge storage.The results show that this method can effectively identify the four types of elements and their relationships, realize the conversion of data-information-knowledge, and provide a technical reference for the construction of knowledge graphs in the field of geological disasters.
Knowledge graph is a knowledge system that formally describes entities and their interrelationships, and it plays an important role in emergency disaster relief and spatio-temporal prediction and decision making. In order to obtain geological hazard information implied in multi-source heterogeneous texts and investigate the impact caused by disasters, in this paper it analyzes many elements associated with the development process of geological hazards from the perspective of hazard chains, and proposes a top-down and bottom-up method of constructing a knowledge graph of geological hazard chains.Firstly, the complex formation mechanism and chain formation law among geohazards are analyzed from the perspective of hazard chain, and a unified semantic expression framework of geohazard chain ontology for information extraction is established based on the top-down method, including geohazard event ontology describing the geohazard knowledge system, geoenvironment ontology carrying the occurrence of geohazards, geographic object ontology under the effect of geohazards, and geohazards in disaster. Secondly, the data layer is constructed by combining the bottom-up method, and the concepts, instances and attributes and their associated relations are identified through knowledge fusion and knowledge storage.The results show that this method can effectively identify the four types of elements and their relationships, realize the conversion of data-information-knowledge, and provide a technical reference for the construction of knowledge graphs in the field of geological disasters.
2023, 48(5): 1892-1907.
doi: 10.3799/dqkx.2022.271
Abstract:
The Tianshan Mountain and its surrounding areas will become the deployment areas of national important strategic transportation, oil and gas resources pipelines, and urban settlement construction in the future. The risk prediction and assessment of debris flow disasters in the region will make the monitoring layout and prevention of major potential debris flow disaster points more targeted. The ensemble learning algorithm can avoid the difficulty of algorithm selection in disaster susceptibility assessment and significantly improve the modeling accuracy. However, its application in debris flow susceptibility assessment is still limited and its reliability needs to be tested. In this paper, the stacking ensemble algorithm was used to evaluate and predict the susceptibility of debris flow disasters in the Tianshan Mountain. Considering 14 characteristic variables such as drought degree and steepness index, the prediction performance of the stacking ensemble algorithm and the independent heterogeneous algorithm was compared. Finally, the control factors of debris flow disasters in the Tianshan area are discussed. The results show follows: (1) The areas with high debris flow disaster and extremely high susceptibility to debris flow in the Tianshan area account for 17.06% and 19.75%, respectively, and are concentrated on the northern slope of the North Tianshan and the southern slope of the South Tianshan. (2) The AUC value of the prediction rate curve of the stacked ensemble algorithm is 0.87, which is significantly higher than that of the independent machine learning algorithm (0.79-0.81) and has better prediction performance than the independent machine learning algorithm. (3) In addition to conventional topography and rainfall, which have significant control on the formation of debris flows in the Tianshan area, drought and uplift have important effects on the formation of debris flow in the Tianshan area. The results of this paper not only contribute to the risk management of debris flow disasters in the Tianshan area but also have implications for the assessment of debris flow susceptibility in arid mountainous areas.
The Tianshan Mountain and its surrounding areas will become the deployment areas of national important strategic transportation, oil and gas resources pipelines, and urban settlement construction in the future. The risk prediction and assessment of debris flow disasters in the region will make the monitoring layout and prevention of major potential debris flow disaster points more targeted. The ensemble learning algorithm can avoid the difficulty of algorithm selection in disaster susceptibility assessment and significantly improve the modeling accuracy. However, its application in debris flow susceptibility assessment is still limited and its reliability needs to be tested. In this paper, the stacking ensemble algorithm was used to evaluate and predict the susceptibility of debris flow disasters in the Tianshan Mountain. Considering 14 characteristic variables such as drought degree and steepness index, the prediction performance of the stacking ensemble algorithm and the independent heterogeneous algorithm was compared. Finally, the control factors of debris flow disasters in the Tianshan area are discussed. The results show follows: (1) The areas with high debris flow disaster and extremely high susceptibility to debris flow in the Tianshan area account for 17.06% and 19.75%, respectively, and are concentrated on the northern slope of the North Tianshan and the southern slope of the South Tianshan. (2) The AUC value of the prediction rate curve of the stacked ensemble algorithm is 0.87, which is significantly higher than that of the independent machine learning algorithm (0.79-0.81) and has better prediction performance than the independent machine learning algorithm. (3) In addition to conventional topography and rainfall, which have significant control on the formation of debris flows in the Tianshan area, drought and uplift have important effects on the formation of debris flow in the Tianshan area. The results of this paper not only contribute to the risk management of debris flow disasters in the Tianshan area but also have implications for the assessment of debris flow susceptibility in arid mountainous areas.
2023, 48(5): 1908-1922.
doi: 10.3799/dqkx.2022.281
Abstract:
Surrounding rock perception based on TBM construction data is essential to ensuring the safety of TBM construction and improving its construction efficiency, in which the accuracy of TBM tunnelling parameter prediction is crucial to testing the effect of surrounding rock perception. Therefore, in this paper it takes Jilin Yin-Song's project TBM4 bid section as the research object, selects the characteristic parameters of the surrounding rock from the rock breaking data of the loading phase of the TBM as input feature X1, and selects two construction control parameters (the rotation speed and penetration rate) as input feature X2, and constructs a convolutional neural network machine learning model to predict the TBM tunnelling response parameters Y (cutterhead torque and total thrust). According to the different learning objects, the point prediction model that only learns the response behavior of the stable boring phase and the line prediction model that simultaneously learns the response behavior of the loading phase and the stable boring phase are constructed, respectively. The improved results show that the point prediction model cannot describe the influence of control parameters on tunnelling response parameters. Although the line prediction model can describe the influence of control parameters on tunnelling response parameters, the prediction value of driving response in the stable boring phase is low. Considering that the low predictive value of the line prediction model in the stable boring phase is because the number of behavior samples in the stable boring phase only accounts for 9% of the total number of samples, in this paper, a method of adjusting loss function is proposed to improve the weight of behavior samples in the stable boring phase, which significantly improves the prediction accuracy of the line prediction model. The results show that the behavior of the loading boring phase should be studied, and the weight of the behavior of the stable boring phase should be increased to obtain a high-precision prediction model of tunnelling response parameters. The model obtained in this paper can provide a basis for further surrounding rock perception and control parameter optimization.
Surrounding rock perception based on TBM construction data is essential to ensuring the safety of TBM construction and improving its construction efficiency, in which the accuracy of TBM tunnelling parameter prediction is crucial to testing the effect of surrounding rock perception. Therefore, in this paper it takes Jilin Yin-Song's project TBM4 bid section as the research object, selects the characteristic parameters of the surrounding rock from the rock breaking data of the loading phase of the TBM as input feature X1, and selects two construction control parameters (the rotation speed and penetration rate) as input feature X2, and constructs a convolutional neural network machine learning model to predict the TBM tunnelling response parameters Y (cutterhead torque and total thrust). According to the different learning objects, the point prediction model that only learns the response behavior of the stable boring phase and the line prediction model that simultaneously learns the response behavior of the loading phase and the stable boring phase are constructed, respectively. The improved results show that the point prediction model cannot describe the influence of control parameters on tunnelling response parameters. Although the line prediction model can describe the influence of control parameters on tunnelling response parameters, the prediction value of driving response in the stable boring phase is low. Considering that the low predictive value of the line prediction model in the stable boring phase is because the number of behavior samples in the stable boring phase only accounts for 9% of the total number of samples, in this paper, a method of adjusting loss function is proposed to improve the weight of behavior samples in the stable boring phase, which significantly improves the prediction accuracy of the line prediction model. The results show that the behavior of the loading boring phase should be studied, and the weight of the behavior of the stable boring phase should be increased to obtain a high-precision prediction model of tunnelling response parameters. The model obtained in this paper can provide a basis for further surrounding rock perception and control parameter optimization.
2023, 48(5): 1923-1934.
doi: 10.3799/dqkx.2022.274
Abstract:
Railway and highway tunnels located in southwest mountainous areas of China often suffer from tunnel squeezing due to high geo-stress, soft rock and fault fracture zone, which bring huge economic losses. To fulfill the requirements of tunnel alignment and design, considering the accessibility of prediction indexes in this stage, the buried depth, surrounding rock grade, equivalent tunnel diameter and rock strength were taken as prediction indexes. Data containing 151 groups of tunnel squeezing cases were collected and a database was established based on tunnels located in Southwest China. The Bayesian network was used to establish the probability classification prediction model of tunnel squeezing with incomplete data. The accuracy of the model was found to be 76.52% by ten-fold cross validation. Based on this model, the tunnel squeezing classification prediction software platform was established. It was applied in Baima tunnel of Jiu-Mian Expressway. The prediction accuracy was 71.11%. The research results of this paper can provide a technical support for the prediction of tunnel squeezing during investigation and design stage under similar geological conditions in Southwest China.
Railway and highway tunnels located in southwest mountainous areas of China often suffer from tunnel squeezing due to high geo-stress, soft rock and fault fracture zone, which bring huge economic losses. To fulfill the requirements of tunnel alignment and design, considering the accessibility of prediction indexes in this stage, the buried depth, surrounding rock grade, equivalent tunnel diameter and rock strength were taken as prediction indexes. Data containing 151 groups of tunnel squeezing cases were collected and a database was established based on tunnels located in Southwest China. The Bayesian network was used to establish the probability classification prediction model of tunnel squeezing with incomplete data. The accuracy of the model was found to be 76.52% by ten-fold cross validation. Based on this model, the tunnel squeezing classification prediction software platform was established. It was applied in Baima tunnel of Jiu-Mian Expressway. The prediction accuracy was 71.11%. The research results of this paper can provide a technical support for the prediction of tunnel squeezing during investigation and design stage under similar geological conditions in Southwest China.
2023, 48(5): 1935-1947.
doi: 10.3799/dqkx.2023.013
Abstract:
To overcome the shortcoming of insufficient landslide inventories, TrAdaboost-DT and TrAdaBoost-RF models with decision tree and random forest as basic learners respectively in 2013-2015 were built, by taking the Wenchuan-Yingxiu, Sichuan Province as the study area and the landslide inventory in 2011-2013 as an auxiliary data set. The proposed models were used to predict landslide susceptibility and prediction results were compared with those of DT and RF models trained by the landslide inventory in 2013-2015. The comparison results show that areas of under receiver operating characteristic curve (AUC) of TrAdaBoost-DT and TrAdaBoost-RF models were more than 0.03 and 0.01 than those of DT and RF models, respectively. To validate the prediction performance of the proposed models, the landslide inventory in 2013-2015 was used to build LS model in 2015-2018. The results indicate that the AUC of both DT and RF models increased by 0.13 using the proposed model. TrAdaBoost algorithm can improve the prediction performance of LS model based on machine learning algorithm and show significant improvement for those under small data sets.
To overcome the shortcoming of insufficient landslide inventories, TrAdaboost-DT and TrAdaBoost-RF models with decision tree and random forest as basic learners respectively in 2013-2015 were built, by taking the Wenchuan-Yingxiu, Sichuan Province as the study area and the landslide inventory in 2011-2013 as an auxiliary data set. The proposed models were used to predict landslide susceptibility and prediction results were compared with those of DT and RF models trained by the landslide inventory in 2013-2015. The comparison results show that areas of under receiver operating characteristic curve (AUC) of TrAdaBoost-DT and TrAdaBoost-RF models were more than 0.03 and 0.01 than those of DT and RF models, respectively. To validate the prediction performance of the proposed models, the landslide inventory in 2013-2015 was used to build LS model in 2015-2018. The results indicate that the AUC of both DT and RF models increased by 0.13 using the proposed model. TrAdaBoost algorithm can improve the prediction performance of LS model based on machine learning algorithm and show significant improvement for those under small data sets.
2023, 48(5): 1948-1959.
doi: 10.3799/dqkx.2022.255
Abstract:
Since commonly used first order reliability method (FORM) often yields inaccurate reliability index of pile bearing capacity, support vector machine (SVM) is combined with FORM to calculate the reliability index of bearing capacity of defective piles in this paper. Taking the reduced-diameter defective pile as an example, the vertical loading test of one complete pile and five reduced-diameter piles was carried out. The random weighting method is used to estimate the mean value of the bearing capacity reduction coefficient of the defective piles. The example shows that the shrinkage is filled with soil, which converts the lateral friction resistance of the pile here into the friction of the soil interface, and weakens the bearing capacity of the pile. The larger the shrinkage length, the larger the area filled by the soil at the shrinkage, and the smaller the pile bearing capacity reduction coefficient and reliability index. The closer the shrinkage position is to the top of the pile, the greater the degree to which the shrinkage limits the resistance of the side friction. The pile end resistance compensates for part of the loss of lateral friction resistance, so that when the shrinkage diameter is located in the shallow and middle parts of the pile body, the pile bearing capacity reduction coefficient and reliability index are roughly the same; when located deep, the bearing capacity loss is minimal and the reliability index is the largest.
Since commonly used first order reliability method (FORM) often yields inaccurate reliability index of pile bearing capacity, support vector machine (SVM) is combined with FORM to calculate the reliability index of bearing capacity of defective piles in this paper. Taking the reduced-diameter defective pile as an example, the vertical loading test of one complete pile and five reduced-diameter piles was carried out. The random weighting method is used to estimate the mean value of the bearing capacity reduction coefficient of the defective piles. The example shows that the shrinkage is filled with soil, which converts the lateral friction resistance of the pile here into the friction of the soil interface, and weakens the bearing capacity of the pile. The larger the shrinkage length, the larger the area filled by the soil at the shrinkage, and the smaller the pile bearing capacity reduction coefficient and reliability index. The closer the shrinkage position is to the top of the pile, the greater the degree to which the shrinkage limits the resistance of the side friction. The pile end resistance compensates for part of the loss of lateral friction resistance, so that when the shrinkage diameter is located in the shallow and middle parts of the pile body, the pile bearing capacity reduction coefficient and reliability index are roughly the same; when located deep, the bearing capacity loss is minimal and the reliability index is the largest.
2023, 48(5): 1960-1976.
doi: 10.3799/dqkx.2022.224
Abstract:
Coseismic landslides are imperceptible and widely distributed in the mountainous regions with abundant vegetation influenced by strong earthquakes, which hinders the development of prevention and control for disaster. Landslide susceptibility evaluation can predict the landslide-prone areas. But few studies have obtained the high-precision landslide susceptibility maps, and it was difficult to uncover the patterns of coseismic landslides under the complex geo-environment, because the lack of data and there was no unified standards for quantification of data based on traditional landslide susceptibility evaluation method. Therefore, the multi-source monitoring data, spatial analysis and deep learning method were adopted in this work to find out the development patterns, seismic response mechanism of landslide and to obtain landslide susceptibility maps. The main conclusions are as follows: the seismic response mechanism of landslide was proposed based on the seismic response analysis of landform and rock structure. The CNN (convolutional neural network) and DNN (deep neural network) achieved the relatively good area under the curve (AUC) value (AUC were 0.901 and 0.865, respectively), which could obtain the high-precision landslide susceptibility maps. Extremely high and high susceptibility region of landslides were widely distributed in 13 gullies, such as Danzu gully and so on. And these gullies were more prone to debris flow under rainstorm.
Coseismic landslides are imperceptible and widely distributed in the mountainous regions with abundant vegetation influenced by strong earthquakes, which hinders the development of prevention and control for disaster. Landslide susceptibility evaluation can predict the landslide-prone areas. But few studies have obtained the high-precision landslide susceptibility maps, and it was difficult to uncover the patterns of coseismic landslides under the complex geo-environment, because the lack of data and there was no unified standards for quantification of data based on traditional landslide susceptibility evaluation method. Therefore, the multi-source monitoring data, spatial analysis and deep learning method were adopted in this work to find out the development patterns, seismic response mechanism of landslide and to obtain landslide susceptibility maps. The main conclusions are as follows: the seismic response mechanism of landslide was proposed based on the seismic response analysis of landform and rock structure. The CNN (convolutional neural network) and DNN (deep neural network) achieved the relatively good area under the curve (AUC) value (AUC were 0.901 and 0.865, respectively), which could obtain the high-precision landslide susceptibility maps. Extremely high and high susceptibility region of landslides were widely distributed in 13 gullies, such as Danzu gully and so on. And these gullies were more prone to debris flow under rainstorm.
2023, 48(5): 1977-1988.
doi: 10.3799/dqkx.2022.216
Abstract:
The conventional method to optimize the slope investigation program is usually assigned with complicated concept and arduous computational efforts. Also, the quantitative evaluation of slope failure loss is required, which is not convenient in practice. In this paper it aims to solve the above problem with a suggested method based on training of response surface-based machine learning model with incomplete features. The relationship between the factor of safety and the site investigation data is established. Then a prediction function is imported and calibrated with simulated samples. This method adopts the root mean square error of factor of safety as the indicator to assess the effectiveness of slope borehole program. The algorithm is provided and applied in an illustrative example of an undrained slope. The results accord well with those reported in literatures. The suggested method provides an efficient way to assess the effectiveness of site investigation program for slope. It has the characteristics of clear concept, simple algorithm and convenient calculation. Also the computational efforts are greatly reduced. This method will be more acceptable for practitioners.
The conventional method to optimize the slope investigation program is usually assigned with complicated concept and arduous computational efforts. Also, the quantitative evaluation of slope failure loss is required, which is not convenient in practice. In this paper it aims to solve the above problem with a suggested method based on training of response surface-based machine learning model with incomplete features. The relationship between the factor of safety and the site investigation data is established. Then a prediction function is imported and calibrated with simulated samples. This method adopts the root mean square error of factor of safety as the indicator to assess the effectiveness of slope borehole program. The algorithm is provided and applied in an illustrative example of an undrained slope. The results accord well with those reported in literatures. The suggested method provides an efficient way to assess the effectiveness of site investigation program for slope. It has the characteristics of clear concept, simple algorithm and convenient calculation. Also the computational efforts are greatly reduced. This method will be more acceptable for practitioners.
2023, 48(5): 1989-1999.
doi: 10.3799/dqkx.2023.036
Abstract:
The prediction and stability analysis of landslide disaster have great engineering significance and application value. Machine learning algorithm is mainly used in landslide displacement prediction, but is limited in landslide stability analysis. Therefore, in order to more accurately analyze the stability of bedding rock slope under cyclic seismic load, the strain softening process of sliding zone soil was obtained by combining the research methods of indoor physical model test and the comparison of discrete element numerical simulation software. In addition, a landslide stability prediction model based on machine learning algorithm is proposed by taking advantage of the nonlinear characteristics of landslide deformation. The results show follows: (1) The gradual reduction of shear stress promotes the strain-softening process of soil in the sliding zone. Although confining pressure of soil in the sliding zone can inhibit the increase of cracks in the sliding zone, its inhibition effect on strain softening is limited. (2) The ARIMA(1, 1, 0)(0, 1, 1) model with the standard BIC value of 8.160 was established to accurately predict the time series data of the slope stability coefficient. Based on the field observation of the slope stability coefficient and stress field, two possible landslide-triggering mechanisms are described. Mechanical learning of time series can accurately predict the variation law of slope stability coefficient under cyclic load.
The prediction and stability analysis of landslide disaster have great engineering significance and application value. Machine learning algorithm is mainly used in landslide displacement prediction, but is limited in landslide stability analysis. Therefore, in order to more accurately analyze the stability of bedding rock slope under cyclic seismic load, the strain softening process of sliding zone soil was obtained by combining the research methods of indoor physical model test and the comparison of discrete element numerical simulation software. In addition, a landslide stability prediction model based on machine learning algorithm is proposed by taking advantage of the nonlinear characteristics of landslide deformation. The results show follows: (1) The gradual reduction of shear stress promotes the strain-softening process of soil in the sliding zone. Although confining pressure of soil in the sliding zone can inhibit the increase of cracks in the sliding zone, its inhibition effect on strain softening is limited. (2) The ARIMA(1, 1, 0)(0, 1, 1) model with the standard BIC value of 8.160 was established to accurately predict the time series data of the slope stability coefficient. Based on the field observation of the slope stability coefficient and stress field, two possible landslide-triggering mechanisms are described. Mechanical learning of time series can accurately predict the variation law of slope stability coefficient under cyclic load.
2023, 48(5): 2000-2010.
doi: 10.3799/dqkx.2022.144
Abstract:
Current research on blasting-induced peak particle velocity prediction in open-pit mines is infeasible for the hole-by-hole blasting operation using digital electronic detonators, and its models lack interpretability. By recording the blasting parameters for each blasthole and measuring the induced triaxial particle velocity, a model to predict the triaxial peak particle velocity is established based on LightGBM, and SHAP is introduced to interpret the variable importance of the model. Regarding the root mean squared error RMSE and goodness of fitting R2 on the test set, the established LightGBM model outperforms the support vector machine and neural network models as its RMSE decreases by 25.9% and 28.9%, while the R2 improves by 12.7% and 9.9%. Compared to the Sadaovsk empirical formula, which is widely applied to the blasting design and safety evaluation, the RMSE of LightGBM declines by 63.4%, 39.5% and 68.3%, while the R2 increases by 18.9%, 27.7%and 42.4% in the longitudinal axis X, transverse axis Y and vertical axis Z, respectively. The SHAP values computed from the model inform that apart from the axis variable, the distance between blasting source and monitoring point, total charge, the minimum row distance, the average charge length, hole diameter and the maximum hole distance are the six most influencing variables that affect the predicted value of peak particle velocity. The distance between blasting source and monitoring point is negatively correlated with the predicted peak particle velocity, while the other five variables are positively correlated with the predicted value.
Current research on blasting-induced peak particle velocity prediction in open-pit mines is infeasible for the hole-by-hole blasting operation using digital electronic detonators, and its models lack interpretability. By recording the blasting parameters for each blasthole and measuring the induced triaxial particle velocity, a model to predict the triaxial peak particle velocity is established based on LightGBM, and SHAP is introduced to interpret the variable importance of the model. Regarding the root mean squared error RMSE and goodness of fitting R2 on the test set, the established LightGBM model outperforms the support vector machine and neural network models as its RMSE decreases by 25.9% and 28.9%, while the R2 improves by 12.7% and 9.9%. Compared to the Sadaovsk empirical formula, which is widely applied to the blasting design and safety evaluation, the RMSE of LightGBM declines by 63.4%, 39.5% and 68.3%, while the R2 increases by 18.9%, 27.7%and 42.4% in the longitudinal axis X, transverse axis Y and vertical axis Z, respectively. The SHAP values computed from the model inform that apart from the axis variable, the distance between blasting source and monitoring point, total charge, the minimum row distance, the average charge length, hole diameter and the maximum hole distance are the six most influencing variables that affect the predicted value of peak particle velocity. The distance between blasting source and monitoring point is negatively correlated with the predicted peak particle velocity, while the other five variables are positively correlated with the predicted value.
2023, 48(5): 2011-2023.
doi: 10.3799/dqkx.2022.251
Abstract:
Accurate prediction of rockburst intensity grade is of great significance for mitigating and eliminating rockburst hazards. Aiming at the problems of uncertain feature selection and low prediction accuracy of rockburst intensity grade prediction model, in this paper it proposes a rockburst grade prediction model based on SSA-SVM-AdaBoost algorithm with ReliefF-Pearson feature selection. The method combines the weight idea of ReliefF and the correlation principle of Pearson coefficient to select feature indexes, and SSA-SVM-AdaBoost algorithm is proposed by using the sparrow search algorithm (SSA) optimized support vector machine(SVM)classifier as the AdaBoost weak classifier to solve the multiclassification problem. First, 7 kinds of feature indicators are selected to form the original feature space by analyzing rockburst case data, then the 4⁃dimension advantage features are selected by ReliefF-Pearson method. The data is processed with random oversampling before input SSA-SVM-AdaBoost prediction model. The research results show that the feature selection method based on ReliefF-Pearson can effectively extract advantage feature indicators. Compared with SSA-SVM and AdaBoost based on single-layer decision tree, the prediction accuracy of SSA-SVM-AdaBoost model is improved by 12.5%, and 31.25% compared with SVM. It shows that SSA-SVM as a weak classifier is better than a single-layer decision tree in classification performance, and the AdaBoost enhancement algorithm integrating multiple single classifiers is better than a single classification model. Data oversampling process does not affect the accuracy of the model prediction set, but improves the prediction accuracy of the training set. It is proved that the proposed model can be effectively applied to rockburst intensity grade prediction, which provides a new perspective for this problem.
Accurate prediction of rockburst intensity grade is of great significance for mitigating and eliminating rockburst hazards. Aiming at the problems of uncertain feature selection and low prediction accuracy of rockburst intensity grade prediction model, in this paper it proposes a rockburst grade prediction model based on SSA-SVM-AdaBoost algorithm with ReliefF-Pearson feature selection. The method combines the weight idea of ReliefF and the correlation principle of Pearson coefficient to select feature indexes, and SSA-SVM-AdaBoost algorithm is proposed by using the sparrow search algorithm (SSA) optimized support vector machine(SVM)classifier as the AdaBoost weak classifier to solve the multiclassification problem. First, 7 kinds of feature indicators are selected to form the original feature space by analyzing rockburst case data, then the 4⁃dimension advantage features are selected by ReliefF-Pearson method. The data is processed with random oversampling before input SSA-SVM-AdaBoost prediction model. The research results show that the feature selection method based on ReliefF-Pearson can effectively extract advantage feature indicators. Compared with SSA-SVM and AdaBoost based on single-layer decision tree, the prediction accuracy of SSA-SVM-AdaBoost model is improved by 12.5%, and 31.25% compared with SVM. It shows that SSA-SVM as a weak classifier is better than a single-layer decision tree in classification performance, and the AdaBoost enhancement algorithm integrating multiple single classifiers is better than a single classification model. Data oversampling process does not affect the accuracy of the model prediction set, but improves the prediction accuracy of the training set. It is proved that the proposed model can be effectively applied to rockburst intensity grade prediction, which provides a new perspective for this problem.
2023, 48(5): 2024-2038.
doi: 10.3799/dqkx.2022.309
Abstract:
The Three Gorges Reservoir Area is the key area for geological disaster management, and the hydraulic effect of the Yangtze River on the slopes along its banks cannot be ignored. Therefore, it is necessary to study the influence of drainage factors on landslide susceptibility. The historical landslides points in Fengjie County and their corresponding features are taken as analysis data. Due to the significant influence of regional water system, the research area is divided into two sub-zones according to hydrographic conditions. Area of 300 meters along the two sides of the rivers is regarded as Sub-Zone Ⅰ, and the remaining area is defined as Sub-Zone Ⅱ. Then, a total 16 influencing factors are selected to establish landslide susceptibility evaluating models, and the landslide susceptibility evaluation results of the whole region and sub-zones were compared and analyzed. The following results of landslide susceptibility analysis based on machine learningalgorithm can be obtained. Because the fluctuation of reservoir water level reduces the effective stress of anti-slip section and the cultivated land has weak conservation effect on slope mass owing to the destruction of the original mountain balance in the process of reclamation, the areas with high and extremely high probability of landslide occurrences in Fengjie County mainly lie on the bank of rivers and in the area of cultivated land. The accuracy of susceptibility assessment of the hydrographic-divided model is better than the whole-range model. Specifically, the accuracy and F-score are improved by 5.1% and 5.2%, which indicates the practicability and validity of conducting zone-dividing susceptibility analysis.
The Three Gorges Reservoir Area is the key area for geological disaster management, and the hydraulic effect of the Yangtze River on the slopes along its banks cannot be ignored. Therefore, it is necessary to study the influence of drainage factors on landslide susceptibility. The historical landslides points in Fengjie County and their corresponding features are taken as analysis data. Due to the significant influence of regional water system, the research area is divided into two sub-zones according to hydrographic conditions. Area of 300 meters along the two sides of the rivers is regarded as Sub-Zone Ⅰ, and the remaining area is defined as Sub-Zone Ⅱ. Then, a total 16 influencing factors are selected to establish landslide susceptibility evaluating models, and the landslide susceptibility evaluation results of the whole region and sub-zones were compared and analyzed. The following results of landslide susceptibility analysis based on machine learningalgorithm can be obtained. Because the fluctuation of reservoir water level reduces the effective stress of anti-slip section and the cultivated land has weak conservation effect on slope mass owing to the destruction of the original mountain balance in the process of reclamation, the areas with high and extremely high probability of landslide occurrences in Fengjie County mainly lie on the bank of rivers and in the area of cultivated land. The accuracy of susceptibility assessment of the hydrographic-divided model is better than the whole-range model. Specifically, the accuracy and F-score are improved by 5.1% and 5.2%, which indicates the practicability and validity of conducting zone-dividing susceptibility analysis.
