CN-121562429-B - Tunnel large deformation prediction method based on geological feature adaptation and PSO-RF algorithm

Abstract

The invention discloses a tunnel large deformation prediction method based on geological feature adaptation and a PSO-RF algorithm, and belongs to the technical field of tunnel engineering safety monitoring. Aiming at the problems of poor adaptability to geological conditions and insufficient prediction precision in the prior art, the method constructs a full-flow technical framework of data preprocessing, geological scene clustering, PSO parameter optimization, RF prediction and result verification. Firstly 9 core indexes are collected, after missing value processing and standardization, a K-means cluster is adopted to divide a geological scene into three categories of a soft rock water-rich section, a fault breaking section and a hard rock stabilizing section, then an improved PSO algorithm is adopted to synchronously optimize RF super parameters and characteristic weights, parameter self-adaption is realized through adjustment of scene inertia weights and constraint of fitness functions, finally a weighted characteristic matrix is constructed to train a PSO-RF model, and I-V large deformation grade prediction results and characteristic importance sequencing are output. Experiments show that the method has high accuracy and low misjudgment rate.

Inventors

• SHUAI YONGJIAN
• SHANG MIAOMIAO
• SONG YU
• ZHAO XUEFENG
• SUN ZHIQIANG
• HE GUOLONG
• YU FEI
• LI XIN
• CHEN KEPU
• LI HUAILIANG
• CHEN TAO
• WANG JIE

Assignees

• 中国水利水电第七工程局有限公司
• 成都理工大学

Dates

Publication Date

20260508

Application Date

20260121

Claims (6)

1. 1. A tunnel large deformation prediction method based on geological feature adaptation and PSO-RF algorithm is characterized by comprising the following steps: The data preprocessing stage comprises the steps of collecting core index data of tunnel engineering, performing standardized processing including missing value filling, outlier correction and data conversion, and then outputting a standardized feature matrix; The geological scene clustering and dividing stage is used for carrying out clustering analysis by adopting a K-means algorithm based on the lithology and the ground stress after coding as core clustering characteristics, eliminating dimension differences to determine optimal clustering numbers, dividing geological scenes into three types of soft rock water-rich sections, fault breaking sections and hard rock stabilizing sections, and distributing scene labels for each sample; constructing multidimensional particle vectors, corresponding to the super parameters of the RF model and the characteristic weights of core index data, and outputting optimal super parameters and optimal characteristic weight vectors through adjustment of scene inertia weights and optimization of fitness functions; Model training and large deformation prediction, namely performing point multiplication on the preprocessed standardized feature matrix and an optimal feature weight vector to construct a weighted feature matrix, training an RF model by adopting optimal super parameters and cross verification to obtain a PSO-RF model, predicting the large deformation level of a tunnel by using the trained PSO-RF model, and outputting feature importance ranking; the missing value filling in the data preprocessing stage comprises the following steps of filling the missing value of the classification variable as unknown and encoding the missing value as 0; abnormal value correction adopts 3 sigma criterion to identify abnormal value, and calculates average value of the j-th numerical characteristic index X j in similar lithology section And standard deviation of When the sample value satisfies The numerical characteristic index is adopted to replace 95% of the numerical values in the similar lithology section, and the numerical value is adopted to replace 5% of the numerical values in the similar lithology section; the data conversion comprises the following steps of adopting single-heat coding to convert the classified variable into a numerical vector, and adopting min-max normalization mapping to a [0,1] interval; in the geological scene clustering dividing stage, carrying out standardized treatment on the lithology coding value and the ground stress normalization value by STANDARDSCALER to eliminate dimension differences; adopting an elbow method to analyze the change trend of the square sum of the clustering errors along with the clustering number K, when K=3, obvious inflection points appear on the SSE curve, and the optimal clustering number is determined to be 3; The scene clustering adopts a K-means algorithm, and geological scenes are divided according to lithology codes and ground stress values of the cluster center and named as a soft rock water-rich section, a fault breaking section and a hard rock stabilizing section; in the improved PSO combined optimization stage, the particle vector is 13 dimensions, the first 4 dimensions correspond to super parameters of a random forest model, and the second 9 dimensions correspond to characteristic weights of 9 core index data; The fitness function is: wherein MSE is the predicted mean square error; Is the number of decision trees; And the scene core feature weight is obtained.
2. 2. The tunnel large deformation prediction method based on the geological feature adaptation and the PSO-RF algorithm as set forth in claim 1, wherein the core index data comprises lithology, burial depth, integrity, rock strength, weathering degree, rock mass strength, ground stress, strength stress ratio and deformation rate.
3. 3. The tunnel large deformation prediction method based on the geological feature adaptation and the PSO-RF algorithm according to claim 1, wherein the classification variables comprise lithology, integrity and wind-age, and the numerical variables comprise burial depth, rock strength, rock mass strength, ground stress, strength stress ratio and deformation rate.
4. 4. The method for predicting large tunnel deformation based on geological feature adaptation and PSO-RF algorithm as set forth in claim 1, wherein the super parameters of the random forest model include the number of decision trees, the maximum depth of the decision trees, the minimum number of split nodes samples and the maximum feature number of each tree.
5. 5. The tunnel large deformation prediction method based on the geological feature adaptation and the PSO-RF algorithm according to claim 1, wherein the large deformation grades are divided into a plurality of grades according to deformation in the model training and large deformation prediction stage.
6. 6. The tunnel large deformation prediction method based on geological feature adaptation and PSO-RF algorithm according to any one of claims 1 to 5, further comprising a model verification stage, wherein the prediction result is visually evaluated by using a confusion matrix, and the true grade label is taken as an ordinate, the prediction grade is taken as an abscissa, and the number of correctly predicted samples of a main diagonal and the misjudgment of a non-diagonal are analyzed.

Description

Tunnel large deformation prediction method based on geological feature adaptation and PSO-RF algorithm Technical Field The invention relates to the technical field of tunnel engineering safety monitoring, in particular to a tunnel large deformation prediction method based on geological feature adaptation and a PSO-RF algorithm. Background Tunnel engineering is used as a core component of traffic infrastructure, however, large deformation is a common disease under complex geological conditions such as soft rock, high ground stress and the like in tunnel construction and operation, and is easy to cause damage of a supporting structure, construction stagnation and even safety accidents, so that accurate prediction of large deformation level has important significance for risk prevention and control. Early tunnel deformation predictions relied primarily on traditional empirical mathematical models, such as regression analysis and time series analysis. Wang Tao, sun Wenlong and Li Lei in the "combined prediction of surrounding rock deformation based on regression analysis and gray theory" (report of underground space and engineering, 2017, volume 13, S1, pages 48-51, DOI: 10.20174), a method of combining regression analysis and gray theory is adopted to develop combined prediction of surrounding rock deformation of a tunnel in a short-term manner, aiming at improving prediction accuracy, but such methods rely on linear assumption and have limited generalization capability in nonlinear and high-dimensional geological data. With the development of intelligent detection technology, machine learning methods such as Random Forest (RF), support Vector Machine (SVM), and Convolutional Neural Network (CNN) are introduced into the field of tunnel deformation prediction. Huang Agang et al in "analysis of tunnel operation safety conditions based on chaos-RF-SVM deformation prediction model" (mapping engineering, 2022, 31 st volume, 4 th period, pages 52-56, DOI: 10.19349), construct a prediction model based on operation tunnel monitoring data in combination with chaos theory, RF algorithm and SVM, verify reliability by M-K test, and embody the advantage of intelligent algorithm in terms of improving accuracy. Linguangdong further confirms in prediction of initial cumulative settlement amount of tunnel construction based on random forest (computing technology and Automation, 2022, volume 41, 1 st, pages 160-163, DOI: 10.16339), that RF model is superior to deep neural network in prediction of initial cumulative settlement amount of tunnel construction, and highlights nonlinear processing ability and anti-overfitting property. However, these models are still limited by the subjectivity of superparameter selection, and improper parameters can limit prediction accuracy and generalization ability. In recent years, the integration of machine learning and intelligent optimization algorithms has become a research hotspot. Du Junsheng et al (patent CN202510683628.4, "method for determining deformation of surrounding rock of water-rich composite stratum tunnel") classify the deformation of surrounding rock by using an RF model, and analyze the current surrounding rock data to be determined by means of a trained surrounding rock deformation classification model to obtain first deformation data. The preliminary judgment of the deformation tendency of the surrounding rock is realized through the strong classification capability of the random forest model, and the accuracy and the efficiency of judgment are improved. However, when the method is used for a water-rich composite stratum, the weight and the threshold depend on a large amount of historical experience, and the parameter determination difficulty is high. Wu Xianguo et al (patent CN116050603a, "hybrid intelligent method-based method of deformation prediction and optimization of undercut tunnel and apparatus") use Bayesian Optimization (BO) to adjust RF super parameters, but for small-pitch undercut tunnels, the parameter constraints are dependent on manual settings, with strong subjectivity. Yao Zhixiong et al (patent CN 202311096268.5) propose a PSO-LSTM based prediction method that verifies the effectiveness of Particle Swarm Optimization (PSO) in optimizing a timing model, which demonstrates the effectiveness of the PSO algorithm in optimizing the timing prediction model. The primary tone and the like adopt an improved black wing iris algorithm to optimize super parameters, so that the technical trend of a meta heuristic algorithm is reflected. In addition, wang Shudong has pointed out in "research on technique for controlling large deformation of weak surrounding rock of tunnel in complicated ground stress area" ([ D ]. Beijing university of transportation, 2010.) that deformation collapse is likely to occur in tunnel construction in the compressive weak surrounding rock due to various uncertainties in geological conditions. It is seen that geological conditions may cause d