Landslide prediction method based on improved random forest in underground engineering scene

Abstract

The invention discloses a landslide prediction method based on an improved random forest in an underground engineering scene, belongs to the technical field of underground engineering, and aims to solve the problem that good stability cannot be achieved while good precision is ensured when landslide prediction is performed by using an existing random forest. According to the method, firstly, data are collected and cleaned in a real underground engineering scene, CART is trained by using a training set, each CART is used for prediction on a plurality of test sets to obtain average classification precision, and descending sorting is carried out; calculating and storing inner product values among the decision trees by adopting a vector inner product method, reserving the decision trees with vector inner products smaller than or equal to an inner product threshold value, and otherwise, marking the decision trees as deletable decision trees; sequentially deleting the decision trees marked as deletable according to the classification accuracy from low to high until the number of the remaining decision trees is N, and determining a decision result of the improved random forest in combination with a majority voting rule; and landslide prediction is carried out by using the obtained random forest classifier.

Description

Technical Field

[0001] The invention belongs to the technical field of underground engineering, and in particular relates to a landslide prediction method in an underground engineering scenario. Background Art

[0002] Underground engineering construction plays a vital role in modern urbanization and infrastructure development. Especially in the construction of tunnels, underground spaces, subways and other projects, geological conditions and natural disasters have become one of the main factors affecting engineering safety and construction progress. Landslides, as a typical geological disaster, pose a serious threat to the safety of underground engineering. The occurrence of landslides not only causes huge economic losses, but may also endanger life safety. Therefore, the prediction and prevention of landslides are particularly important in underground engineering construction. Effective landslide prediction can provide timely warnings for engineering construction and help engineering personnel take appropriate protective measures, thereby reducing disaster losses and ensuring the safety and stability of underground engineering.

[0003] Traditional landslide prediction methods are mainly based on geological surveys, empirical rules, statistical analysis and other means, relying on expert knowledge and a large amount of field data to infer the probability of landslide occurrence. Early landslide prediction methods generally use geological survey data, combined with geological and meteorological factors such as slope, geological type, soil moisture, and precipitation, to predict landslides by establishing risk assessment models. For example, Wang et al. proposed a landslide displacement time series prediction model based on mean low-order autoregressive tensor completion (MLATC). Xue et al. used a data-driven landslide susceptibility assessment method and a physics-based landslide stability assessment method to predict reservoir landslides. Niu et al. developed a GIS-based landslide geological disaster prediction method for the soil-rock contact zone in southern Shaanxi, and established a landslide geological disaster information value prediction model. Junwei et al. proposed an ILF-FFT prediction and warning model for reservoir landslides under rainfall conditions, and realized a four-stage quantitative division method for landslide evolution based on the limiting curvature of velocity and acceleration time history functions.

[0004] However, traditional landslide prediction methods have certain limitations. First, these methods usually rely on expert experience and intuition, lack systematic scientific basis and data support, and easily lead to subjectivity and deviation in prediction results. Secondly, traditional landslide prediction methods mostly rely on local data of geological background, which is difficult to effectively deal with complex multi-dimensional and multi-variable problems in underground engineering environments. More importantly, the prediction accuracy of traditional landslide prediction methods is often limited by data quality. For large-scale landslide prediction, the adaptability and generalization ability of traditional landslide prediction methods are weak and cannot meet the high requirements of practical engineering applications. Therefore, it is urgent to develop new landslide prediction methods to improve prediction accuracy, applicability and reliability.

[0005] With the development of artificial intelligence technology, intelligent landslide prediction methods based on machine learning have gradually become a hot topic of research. The principle is to build a model to predict the risk of new landslides by analyzing the relationship between historical landslide events and related environmental factors (such as topography, geology, meteorology, etc.). The process first collects and processes a large amount of data to extract key features, and then uses classifiers (such as support vector machines, decision trees, random forests, etc.) to train the data to learn the laws and patterns of landslide occurrence. The trained classifier can predict the possibility of landslides by inputting corresponding feature data when facing new geological conditions. Machine learning can handle complex nonlinear relationships, mine potential patterns and associations, and provide more accurate and dynamic landslide predictions than traditional landslide prediction methods. Many scholars have carried out related research. For example, Zhang et al. proposed a new fuzzy deep learning (FuDL) model for near-real-time earthquake-induced landslide spatial prediction. Feng et al. proposed a landslide displacement prediction method based on the chaotic Gaussian mutation sparrow search algorithm to optimize the BP neural network (CG-SSA-BP). Ma et al. used machine learning algorithms to establish prediction models for similar types of landslide deformation, mixed types of landslide deformation, and single landslide deformation. In particular, random forest, as an integrated learning method, has also been widely used in landslide prediction due to its advantages in processing high-dimensional data, especially showing strong prediction ability when dealing with nonlinear problems. For example, Wang et al. used Bayesian hyperparameters to optimize the hyperparameters of random forests and then selected the optimal hyperparameters for landslide susceptibility mapping. Cheng et al. used random forests to predict landslide areas.

[0006] Although intelligent landslide prediction methods have higher accuracy and adaptability than traditional landslide prediction methods, there are still some problems in existing research. First, most of the existing machine learning classifiers are verified in laboratory environments, and the data sets used are often public data sets, which lack specific adaptability to underground engineering scenarios, resulting in poor results of these classifiers in practical applications. For example, many studies use public data sets in geological research, and the environmental conditions of these data sets are quite different from the complex situations faced in actual underground engineering. Therefore, classifiers used for landslide prediction need to be combined with complex data in actual engineering environments to better meet the needs of engineering applications. Secondly, most of the existing studies focus on selecting the most advanced classifiers, or optimizing them based on existing classifiers, especially conducting a lot of exploration in parameter tuning to find the best combination. However, these methods ignore the theoretical depth and optimization space of the classifiers, and often lack optimization of the classifiers from the construction principle. For example, some studies fine-tune existing models through deep learning methods or reinforcement learning. Although good results have been achieved on some data sets, the lack of systematic optimization has led to poor generalization ability and stability of the classifiers. Therefore, classifier selection and optimization should focus on the coordination between theoretical basis and actual data, rather than simply relying on parameter tuning. Finally, many intelligent landslide prediction methods are relatively simple in scheme design. Usually, a certain classifier is selected for experiment, but sufficient comparative analysis is not carried out, and the basis for selecting the optimal classifier is not given. Although many studies have made preliminary attempts on different classifiers, they usually do not conduct detailed performance comparisons, nor do they clearly give the applicability of different algorithms in specific engineering environments.

[0007] In addition, when the existing random forest predicts landslides, since the correlation measurement between features is the same, and the same type of features will have different differences in different environments, when a random forest is established for a certain data set, it can achieve relatively good prediction results, but when predicting landslides in different underground engineering scenarios, it cannot guarantee good recognition effects, that is, there is a problem that it cannot ensure good accuracy while having good stability. Summary of the invention

[0008] In order to solve the problem that the existing random forest cannot ensure good accuracy and good stability when predicting landslides.

[0009] A landslide prediction method based on improved random forest in underground engineering scenarios, firstly obtains geological parameter characteristics, and then uses the improved random forest model to predict landslides;

[0010] The construction process of the improved random forest model includes the following steps:

[0011] Step 1: Insert the integrated sensor module connected to the parameter acquisition terminal into the slope to collect samples. Each sample report includes multiple geological parameter characteristics, thereby obtaining the original sample set;

[0012] K sample subsets are selected from the original sample set by random sampling without replacement, where K is greater than or equal to 3, as test sets for evaluating the performance of each CART classification;

[0013] Step 2: Determine the number of CARTs N that need to be trained and the number of features in the corresponding feature subset; use bagging to perform (1+m)×N random sampling with replacement on the remaining samples of the original sample set to construct (1+m)×N training sets; on this basis, train (1+m)×M CARTs; where m×N CARTs means training a certain number of CARTs more than the preset number;

[0014] Step 3: Apply each CART to make predictions on K sample subsets, and express the classification accuracy obtained as Where, i=1,2,...,(1+m)×N, represents the i-th CART, j=1,2,3,...,K, represents the j-th sample subset;

[0015] Step 4: Calculate the average classification accuracy achieved by the i-th CART

[0016] Step 5: Sort the CARTs in descending order according to the average classification accuracy obtained;

[0017] Step 6: Use the vector dot product method shown in formula (1) to calculate and save the inner product value between CARTs; use the overall classification accuracy obtained by random forest as the evaluation index, and combine the grid search method to find the optimal inner product threshold t; retain the pair of CARTs whose inner product value is less than or equal to the inner product threshold; for the pair of CARTs whose inner product value is greater than the inner product threshold, mark the CART with the low average classification accuracy as deletable;

[0018] Sim(D i ,D j )=acos(W i ·W j ) (1)

[0019] Where D i and D j Represent two CARTs, W i represents the feature subset corresponding to CARTi, W j represents the feature subset corresponding to CARTj, acos(·) represents the arccosine function;

[0020] W i ·W j The detailed calculation formula is as follows:

[0021]

[0022] Where W im Represents the feature subset W i The mth feature in W jn Represents the feature subset W j The mth feature in; I is the indicator function, only when W im =W jn When I(W im =W jn )=1, otherwise I(W im =W jn )=0;W im =W jn Indicates that the same features are selected in different feature subsets;

[0023] Step 7: Arrange all CARTs marked as deletable in ascending order according to the average classification accuracy obtained, and delete them one by one until the number of remaining CARTs is N; that is, the number of remaining CARTs reaches the preset number;

[0024] Step 8: Use the retained N CARTs to construct a random forest; combine the majority voting rule to determine the decision result of the improved random forest.

[0025] Furthermore, the geological parameters include: inclination angle, surface displacement, horizontal displacement, vertical displacement, water pressure, soil deformation, soil stress, soil moisture, soil temperature, shear strength and electrical conductivity.

[0026] Furthermore, in Step 7, if all CARTs marked as deletable are deleted and the number of remaining unmarked CARTs is still greater than N, then the CARTs are deleted in sequence according to the order given in Step 5 until the number of remaining CARTs is N.

[0027] Furthermore, in Step 1, data cleaning is performed on the data set corresponding to the collected samples to obtain the original sample set.

[0028] Furthermore, K=5.

[0029] Beneficial effects:

[0030] The improved random forest based on multi-decision tree correlation measurement proposed in the present invention can improve the accuracy and stability of landslide prediction. At the same time, the overall and local classification performance of different classifiers and neural networks is comprehensively evaluated, and the basis for selecting the optimal classifier is provided to ensure the final classification effect. In addition, the present invention collects data in real underground engineering application scenarios and performs data cleaning to construct a high-quality and highly reliable data set, thereby ensuring the prediction effectiveness and accuracy of the final random forest. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Building a flowchart for improving random forests.

[0032] Figure 2 Precision plots obtained for the six classifiers.

[0033] Figure 3 Recall plots obtained for the six classifiers.

[0034] Figure 4 Plot of the F1-scores obtained for the six classifiers.

[0035] Figure 5 This is a graph showing the relationship between the inner product threshold value and the classification accuracy achieved by random forest.

[0036] Figure 6 Precision plots obtained for the three improved random forests.

[0037] Figure 7 Recall graphs obtained for the three improved random forests.

[0038] Figure 8 F1-score plots obtained for the three improved random forests.

[0039] Fig. 9 Precision plots achieved for four improved random forests.

[0040] Fig.10 Recall plots obtained for four improved random forests.

[0041] Fig.11 Plot of the F1-score obtained for the four improved random forests. DETAILED DESCRIPTION

[0042] Underground engineering construction is crucial in modern urbanization and infrastructure development. Landslides, as a typical geological disaster, pose a serious threat to the safety of underground engineering. Landslide prediction has become an important means to ensure the stability of engineering. Traditional landslide prediction methods rely on expert experience and geological surveys. Although they have achieved certain results in specific scenarios, due to the lack of systematic data support and scientific basis, their prediction results often have problems of subjectivity and lack of accuracy. In addition, traditional landslide prediction methods are difficult to cope with the complexity and multidimensionality of underground engineering environments. When dealing with large-scale landslide predictions, their adaptability and generalization capabilities are poor. To this end, more and more studies have turned to intelligent landslide prediction methods based on machine learning. By analyzing the relationship between historical landslide data and geological, topographic and meteorological factors, data-driven models are established to predict the possibility of landslides. However, most of the existing intelligent landslide prediction methods are verified using public data sets, lacking adaptability to underground engineering scenarios, which limits their practical application effects.

[0043] The present invention aims to solve the limitations of the existing traditional and intelligent landslide prediction methods in underground engineering scenarios. First, in a real underground engineering scenario, special equipment is used to collect the key parameters of landslide occurrence on site, and the collected raw data is used to construct a preliminary data set. Subsequently, the data set is cleaned, including missing value and outlier processing, to ensure the high quality and reliability of the data, laying a solid foundation for subsequent modeling. Secondly, a variety of latest or classic classifiers and neural networks suitable for data classification are trained on the data set, and their overall classification performance on the data set and their local classification performance on data with different labels are compared and analyzed in detail. Through comprehensive evaluation, random forest is finally selected as the optimal classifier, and the scientific basis of the selection and its adaptability in a specific engineering environment are elaborated in detail. Finally, based on the in-depth understanding of the construction principle of random forest, an improved random forest method based on multi-decision tree correlation measurement is proposed, and a full comparative study is carried out with the classic random forest and other improved versions. The results show that the improved random forest performs better in landslide prediction, further improves the prediction accuracy and stability, and effectively solves the limitations of the application of traditional and intelligent landslide prediction methods in underground engineering scenarios. The present invention effectively makes up for the deficiencies in existing landslide prediction methods, not only providing a new solution for landslide prediction in underground engineering scenarios, but also providing important reference and lessons for the theoretical and practical development in the field of landslide prediction. Specific implementation method one:

[0045] This implementation is a landslide prediction method based on improved random forest in underground engineering scenarios. First, an improved random forest model is constructed, and then the improved random forest model is used to predict landslides. Random forest is a combined classifier based on ensemble learning, which is composed of multiple CART decision trees. The process of constructing the improved random forest model is as follows: Figure 1 As shown, including:

[0046] Step 1: The integrated sensor module connected to the parameter acquisition terminal is inserted into the slope to collect samples. Each sample report includes multiple geological parameter characteristics, including: tilt angle, surface displacement, horizontal displacement, vertical displacement, water pressure, soil deformation, soil stress, soil moisture, soil temperature, shear strength and conductivity. Subsequently, the original sample set is obtained by cleaning the data set.

[0047] Five sample subsets were selected from the original sample set by random sampling without replacement as test sets for evaluating the classification performance of each CART.

[0048] It should be noted that the five sample subsets contain the same number of samples, but the specific samples they contain are different, and the labels of these samples are known.

[0049] Step 2: Determine the number of CARTs N that need to be trained and the number of features in the corresponding feature subset. Use bagging to perform (1+m)×N random sampling with replacement on the remaining samples of the original sample set to construct (1+m)×N training sets. On this basis, train (1+m)×N CARTs. Among them, m×N CARTs means training a certain number of CARTs on top of the preset number.

[0050] It should be noted that m is called the weighting coefficient, and its value is usually 0.1 or 0.15.

[0051] Step 3: Apply each CART to make predictions on five sample subsets, and express the classification accuracy obtained as Among them, i=1,2,...,(1+m)×N, represents the i-th CART, j=1,2,3,4,5, represents the j-th sample subset.

[0052] Step 4: Calculate the average classification accuracy achieved by the i-th CART. The formula is:

[0053] Step 5: Arrange the CARTs in descending order according to the average classification accuracy obtained.

[0054] Step 6: Use the vector dot product method shown in formula (1) to calculate and save the inner product value between CARTs. Use the overall classification accuracy obtained by random forest as the evaluation index, and combine the grid search method to find the optimal inner product threshold t. In this way, for those pairs of CARTs whose inner product values ​​are less than or equal to the inner product threshold, no processing is performed. For those pairs of CARTs whose inner product values ​​are greater than the inner product threshold, the CART with low average classification accuracy is marked as deletable.

[0055] Sim(D i ,D j )=acos(W i ·W j ) (1)

[0056] Where D i and D j Represent two CARTs, W i represents the feature subset corresponding to CARTi, W j represents the feature subset corresponding to CARTj, W i ·W j The detailed calculation formula is as follows:

[0057]

[0058] Where W im Represents the feature subset W i The mth feature in W jn Represents the feature subset W j The nth feature in . I is the indicator function. im =W jn When I(W im =W jn )=1, otherwise I(W im =W jn )=0. im =W jn Indicates that the same features are selected in different feature subsets;

[0059] From formula (2), we can see that the denominator is used to calculate the feature subset W i With feature subset W j The sum of the values ​​of all features in the numerator is the calculation of the feature subset W. i With feature subset W jThe sum of the values ​​between the same features in . It can be seen that the key to measuring the correlation between two CARTs lies in the numerator of formula (2). It is worth noting that in the present invention, the feature subset corresponding to CART is calculated, which is actually the value between the features to measure the correlation between the two. After studying the landslide and the geological parameters corresponding to it, the present invention determines that by evaluating the role of the same geological parameters in different feature subsets to assist in the judgment of directional consistency, the present invention decides to use the dot product method of formula (1) to evaluate the correlation between CARTs. After subsequent verification, the present invention can improve the accuracy and stability of landslide prediction, and can have good stability while ensuring good accuracy.

[0060] Step 7: Arrange all CARTs marked as deletable in ascending order according to the average classification accuracy obtained, and delete them one by one until the number of remaining CARTs is N. That is, the number of remaining CARTs reaches the preset number. It should be noted that if all CARTs marked as deletable are deleted, and the number of remaining unmarked CARTs is still greater than N, they will be deleted one by one according to the order of CARTs given in Step 5 until the number of remaining CARTs is N.

[0061] Step 8: Use the retained N CARTs to construct a random forest. Combined with the majority voting rule, determine the decision result of the improved random forest.

[0062] Example

[0063] The specific implementation process of this embodiment is as follows:

[0064] 1. Complete Dataset:

[0065] First, the acquisition terminal is placed on multiple slopes in the underground engineering scene. The integrated sensor module connected to the parameter acquisition terminal is inserted into the slope to collect geological parameters, including: inclination angle, surface displacement, horizontal displacement, vertical displacement, water pressure, soil deformation, soil stress, soil moisture, soil temperature, shear strength and conductivity, a total of 11 geological parameters, called 11 features. In this way, during the operation of the parameter acquisition terminal, each acquisition will obtain the values ​​of 11 geological parameters and form a piece of data. Multiple acquisitions will eventually obtain multiple pieces of data. According to the actual needs of landslide prediction in underground engineering scenarios, these data are divided into four categories, including: normal, small soil erosion, soil erosion and landslide, quantified into four labels of "0", "1", "2" and "3", and corresponding labels are added to these data. Taking an underground project in a province in southern China as an example, a large amount of data was collected using the parameter acquisition terminal from January 2022 to October 2023, and the above labels were added to the data according to the actual situation. A preliminary data set was constructed, and its detailed description is shown in Table 1. It is worth noting that before constructing the preliminary data set, the original large amount of data collected was artificially balanced to ensure that the number of data with the four labels was roughly equal. Specifically, the number of data with the label "3" was used as the reference number, and approximately equal numbers of data were selected from the data of the other three labels.

[0066] Table 1 Detailed description of the preliminary dataset

[0067]

[0068] After statistics, the preliminary data set contains 30,389 data in total. Referring to the 3σ principle of normal distribution, the preliminary data set is tested for outliers. Specifically, with 11 features as the basic unit, 30,389 values ​​belonging to the same feature are tested each time, and values ​​exceeding three times the standard deviation are regarded as outliers. In this way, those data with outliers are called outliers. Similarly, missing value detection is performed on the preliminary data set. Specifically, those values ​​with special symbols such as "NULL" and "?" are regarded as missing values. In this way, those data with missing values ​​are called missing data. On this basis, abnormal data and missing data are uniformly referred to as incomplete data. It is worth noting that there may be some data that contains both outliers and missing values ​​in incomplete data. After statistics, there are 43 incomplete data in the data labeled "0", 51 incomplete data in the data labeled "1", 45 incomplete data in the data labeled "2", and 56 incomplete data in the data labeled "3", totaling 195 incomplete data. The research of Sun et al. shows that for a data set, if the proportion of abnormal data does not exceed 2%, or the proportion of missing data does not exceed 3%, these abnormal data or missing data can be directly discarded because they do not pose a threat to the completeness and diversity of the data set.

[0069] After calculation, the proportion of incomplete data with label "0" in the data with label "0" is about 0.57%, the proportion of incomplete data with label "1" in the data with label "1" is about 0.67, the proportion of incomplete data with label "2" in the data with label "2" is about 0.59, the proportion of incomplete data with label "3" in the data with label "3" is about 0.74%, and the proportion of all incomplete data in the data set is about 0.64%. It can be seen that whether in the data of each label or in the entire data set, the proportion of incomplete data does not exceed 1%, not to mention the proportion of abnormal data or missing data. Therefore, the incomplete data in the preliminary data set is directly removed, thereby obtaining a high-quality complete data set in the underground engineering scenario for subsequent experiments, and its detailed description is shown in Table 2.

[0070] Table 2 Detailed description of the complete dataset

[0071]

[0072] 2. Optimal classifier:

[0073] Classifiers in machine learning are usually divided into three types: classifiers based on mathematical theory, classifiers based on tree models, and deep learning models. Among the classifiers based on mathematical theory, k nearest neighbor and support vector machine are the most commonly used. Among the classifiers based on tree models, decision trees and random forests are the most commonly used. In deep learning models, multilayer perceptrons and one-dimensional convolutional neural networks are the most commonly used for data classification. Therefore, for the complete data set shown in Table 2, it is divided into training set and test set in a ratio of 7:3. The above six classifiers are trained on the training set respectively, and the parameter configurations when they achieve the highest classification accuracy are obtained, as shown in Table 3, and their classification performance is tested again on the test set, and the classification accuracy, Precision, Recall and F1-score they achieve are counted. Table 4 shows the classification accuracy achieved by the six classifiers on the test set. Figures 2 to 4 The Precision, Recall, and F1-score obtained by the six classifiers on the test set are given.

[0074] Table 3 Optimal parameter configurations of six classifiers

[0075]

[0076] Table 4 Classification accuracy achieved by six classifiers

[0077]

[0078] As can be seen from Table 4, the highest classification accuracy achieved by random forest is 89.84%, which is 4.69%, 1.81%, 3.43%, 10.92% and 2.07% higher than the classification accuracy achieved by k-nearest neighbor, support vector machine, decision tree, multi-layer perceptron and one-dimensional convolutional neural network, respectively. The advantages are obvious. Among them, the support vector machine achieved the second highest classification accuracy, and it has obvious performance advantages compared with the other four classifiers, proving that it can construct the interface of high-dimensional feature space to effectively realize data classification when the kernel function is introduced, but there is still a certain gap compared with the random forest based on ensemble learning. The one-dimensional convolutional neural network also achieved satisfactory classification accuracy, proving its stability as a classic neural network for data classification. Among them, the multi-layer perceptron achieved the lowest classification accuracy, and its classification performance was greatly affected by the initial weights and was prone to fall into local optimal solutions. From Figures 2 to 4It can be seen that the random forest also achieved the highest Precision, Recall and F1-score on the data of the four labels, and its advantages over other classifiers are also obvious. This proves that its overall classification performance on the data set is the most significant. In addition, the Precision, Recall and F1-score achieved by the random forest on the data of the four labels are also the most balanced and stable, with little difference between them. This proves that its local classification performance on the data of each label is the most significant and stable.

[0079] 3. Improved Random Forest:

[0080] Based on the improved random forest method of multi-decision tree correlation measurement of the present invention, the random forest with default parameter configuration and the random forest after parameter optimization (optimal parameter configuration in the optimal classifier) ​​are improved respectively, and two improved random forests are obtained, which are respectively called default improved random forest and optimized improved random forest. Among them, the inner product thresholds of the two improved random forests are set to 21 and 25, respectively, which are explained. As mentioned above, it is necessary to determine the inner product threshold to judge the correlation between CART. Specifically, the overall classification accuracy obtained by random forest is used as an evaluation index, and the optimal inner product threshold is found in combination with the grid search method. Train multiple CARTs on the training set, and calculate the inner product value between CARTs. On this basis, according to experience, the value range of the inner product threshold t is set to 5 to 29. In this way, when the inner product threshold t is sequentially taken between 5 and 29, the average classification accuracy of a pair of CARTs higher than the inner product threshold t is deleted, and the remaining CARTs are used to construct a random forest, and the classification accuracy obtained on the test set is calculated. Therefore, by using the grid search method, the value range of the search inner product threshold t is set to 5 to 29. Finally, 25 random forests are constructed and 25 classification accuracies are obtained. At this point, the optimal inner product threshold t can be obtained by comparing which random forest achieves the highest average classification accuracy. Figure 5 The relationship between the value of the inner product threshold and the classification accuracy achieved by random forest is given.

[0081] Depend on Figure 5 It can be seen that when the inner product threshold is 21, the default improved random forest constructed achieves the highest classification accuracy. When the inner product threshold is 25, the optimized improved random forest constructed achieves the highest classification accuracy. Thereafter, when the inner product threshold continues to increase, the average classification accuracy achieved by the two improved random forests decreases slightly and then remains stable. Therefore, in the present invention, the inner product threshold of the default improved random forest is set to 21, and the inner product threshold of the optimized improved random forest is set to 25.

[0082] On this basis, the classification accuracy, Precision, Recall and F1-score obtained by the two improved random forests were statistically analyzed. Among them, the classification accuracy obtained by the default improved random forest and the optimized improved random forest were 91.68% and 93.05%, respectively. Compared with the classification accuracy of 89.84% obtained by the random forest after parameter optimization, they were improved by 1.84% and 3.21%, respectively. This preliminarily illustrates the feasibility, practicality and superiority of the improved random forest method proposed in the present invention. Among them, the classification accuracy advantage of the default improved random forest compared to the random forest after parameter optimization fully demonstrates that only fine-tuning the parameters of the classifier cannot improve its generalization ability and stability, but system optimization based on the construction principle can effectively solve this problem. This fully illustrates the necessity of carrying out the present invention. The most significant classification accuracy achieved by the optimized improved random forest fully demonstrates that simultaneously performing parameter fine-tuning and system optimization can synchronously improve the classification performance of the original classifier, and is an integrated optimization solution worthy of consideration. In addition, Figures 6 to 8 The Precision, Recall, and F1-score obtained by three improved random forests (including random forests after parameter optimization) on the test set are given.

[0083] from Figures 6 to 8 It can be seen that the optimized improved random forest achieves the highest Precision, Recall and F1-score on the data of the four labels, which is significantly better than the default improved random forest and the parameter optimization random forest. This proves that the overall classification performance achieved by the optimized improved random forest on the data set is the most significant. At the same time, the Precision, Recall and F1-score achieved by the default improved random forest are higher than those achieved by the parameter optimization random forest. This effectively illustrates the feasibility and superiority of the improved random forest method proposed in the present invention. Moreover, the Precision, Recall and F1-score achieved by the optimized improved random forest on the data of the four labels are also the most balanced and stable, with the smallest difference between them. This proves that the local classification performance achieved on the data of each label is the most significant and stable. At the same time, the difference in Precision, Recall and F1-score achieved by the default improved random forest is smaller than that achieved by the parameter optimization random forest. This effectively illustrates the stability of the improved random forest method proposed in the present invention. In summary, the improved random forest method proposed in the present invention can simultaneously improve the overall classification performance and local classification performance of random forests, which is significantly better than the existing parameter optimization methods.

[0084] Improved random forests in the latest literature are used for landslide prediction comparison, including the Bayesian hyperparameter optimization random forest (BH-RF) proposed by Wang et al., the Bayesian optimized random forest combined with Kalman filter (BORF-KF) proposed by Zhang et al., and the cascade forest random forest framework (CF-RF) proposed by Chen et al. The three improved random forests are trained on the training set, and their classification performance is tested on the test set. The classification accuracy, Precision, Recall and F1-score achieved by them are statistically compared with the optimized improved random forest. Table 5 shows the classification accuracy achieved by the four improved random forests on the test set. Figures 9 to 11 The Precision, Recall, and F1-score achieved by four improved random forests on the test set are given.

[0085] Table 5 Classification accuracy achieved by four improved random forests

[0086]

[0087] As can be seen from Table 5, the optimized improved random forest constructed by the present invention still achieves the highest classification accuracy of 93.05%, which is 2.69%, 2.77% and 3.13% higher than the classification accuracy of 90.36%, 90.28% and 89.92% achieved by BH-RF, BORF-KF and CF-RF, respectively, with obvious advantages. It is worth noting that the classification accuracy achieved by the constructed default improved random forest is 91.68%, which is also higher than the classification accuracy achieved by BH-RF, BORF-KF and CF-RF. This fully demonstrates the superiority of the improved random forest method proposed by the present invention, and achieves the best overall classification performance on the data set. In fact, BH-RF is also an improved random forest method for parameter optimization, and BORF-KF and CF-RF are both improved random forest methods that combine random forests with other classifiers. As mentioned in the introduction, these methods focus on optimizing existing classifiers, especially conducting extensive exploration on parameter tuning, or using reinforcement learning to fine-tune existing models, and fail to conduct systematic optimization research based on the construction principle of random forests. This fully proves the correctness and necessity of carrying out the present invention.

[0088] from Figures 9 to 11It can be seen that the optimized and improved random forest constructed by the present invention achieves the highest Precision, Recall and F1-score, which are significantly higher than the Precision, Recall and F1-score achieved by BH-RF, BORF-KF and CF-RF. Moreover, the Precision, Recall and F1-score achieved by the optimized and improved random forest on the data of the four labels are also the most balanced and stable, with the smallest difference between them. This once again illustrates that the optimized and improved random forest constructed by the present invention achieves the most significant and stable local classification performance on the data of each label, demonstrating its local classification performance advantage over BH-RF, BORF-KF and CF-RF. In summary, compared with the existing improved random forests for landslide prediction, the overall classification performance and local classification performance of the optimized and improved random forest constructed by the present invention are better. The feasibility and correctness of the improved random forest method proposed by the present invention have been proven, and its feasibility, practicality, stability and superiority have been fully demonstrated.

[0089] The above calculation examples of the present invention are only used to explain the calculation model and calculation process of the present invention in detail, and are not intended to limit the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A landslide prediction method based on improved random forest in underground engineering scenarios, characterized in that: Firstly, the geological parameter characteristics are obtained, and then the improved random forest model is used to predict landslides; The construction process of the improved random forest model includes the following steps: Step 1: Insert the integrated sensor module connected to the parameter acquisition terminal into the slope to collect samples. Each sample report includes multiple geological parameter characteristics, thereby obtaining the original sample set; K sample subsets are selected from the original sample set by random sampling without replacement, where K is greater than or equal to 3, as test sets for evaluating the performance of each CART classification; Step 2: Determine the number of CARTs N that need to be trained and the number of features in the corresponding feature subset; use bagging to perform (1+m)×N random sampling with replacement on the remaining samples of the original sample set to construct (1+m)×N training sets; on this basis, train (1+m)×N CARTs; where m×N CARTs means training a certain number of CARTs more than the preset number; Step 3: Apply each CART to make predictions on K sample subsets, and express the classification accuracy obtained as Where, i=1,2,...,(1+m)×N, represents the i-th CART, j=1,2,3,...,K, represents the j-th sample subset; Step 4: Calculate the average classification accuracy achieved by the i-th CART Step 5: Sort the CARTs in descending order according to the average classification accuracy obtained; Step 6: Use the vector dot product method shown in formula (1) to calculate and save the inner product value between CARTs; use the overall classification accuracy obtained by random forest as the evaluation index, and combine the grid search method to find the optimal inner product threshold t; retain the pair of CARTs whose inner product value is less than or equal to the inner product threshold; for the pair of CARTs whose inner product value is greater than the inner product threshold, mark the CART with the low average classification accuracy as deletable; Sim(D i ,D j )=acos(W i ·W j ) (1) Where D i and D j Represent two CARTs, W i represents the feature subset corresponding to CARTi, W j represents the feature subset corresponding to CARTj, acos(·) represents the arccosine function; W i ·W j The detailed calculation formula is as follows: Where W im Represents the feature subset W i The mth feature in W jn Represents the feature subset W h The nth feature in; I is the indicator function, only when W im =W jn When I(W im =W jn )=1, otherwise I(W im =W jn )=0;W im =W jn Indicates that the same features are selected in different feature subsets; Step 7: Arrange all CARTs marked as deletable in ascending order according to the average classification accuracy obtained, and delete them one by one until the number of remaining CARTs is N; that is, the number of remaining CARTs reaches the preset number; Step 8: Use the retained N CARTs to construct a random forest; combine the majority voting rule to determine the decision result of the improved random forest.

2. The landslide prediction method based on improved random forest in underground engineering scenario according to claim 1 is characterized in that: The geological parameters include: tilt angle, surface displacement, horizontal displacement, vertical displacement, water pressure, soil deformation, soil stress, soil moisture, soil temperature, shear strength and electrical conductivity.

3. The landslide prediction method based on improved random forest in underground engineering scenario according to claim 1 or 2 is characterized in that: In the step 7, if the number of the remaining unmarked CARTs is still greater than N after all the CARTs marked as deletable are deleted, they are deleted in sequence according to the order of the CARTs given in step 5 until the number of the remaining CARTs is N.

4. The landslide prediction method based on improved random forest in underground engineering scenario according to claim 3 is characterized in that: In Step 1, data cleaning is performed on the data set corresponding to the collected samples to obtain the original sample set.

5. The landslide prediction method based on improved random forest in underground engineering scenario according to claim 3 is characterized in that: Said K=5.

Images