A Gray Wolf Optimization-Based Improved Probabilistic Neural Network Algorithm for Surrounding Rock Squeezing Classification in Tunnel Engineering

Abstract

Surrounding rock squeezing deformation is a common and prominent hazard in tunnel engineering projects, which often induces the shield jamming disaster during the TBM tunneling process. Based on the 139 groups of historical squeezing deformation cases, this study developed a hybrid PCA-IWGO-PNN model for squeezing classification. According to the influencing factors and characteristics of squeezing deformation, the strength-stress ratio, tunnel burial depth, tunnel equivalent diameter, rock mass quality index, and support stiffness were selected to establish the prediction index system of squeezing level. Because the probabilistic neural network (PNN) requires that the input variables are independent, principal component analysis (PCA) was used to preprocess the original data to eliminate the correlation between prediction indexes and achieve dimensionality reduction. The spread coefficient was the critical hyper-parameter in the PNN, and the improved gray wolf optimization (IGWO) algorithm was used to realize its efficient automatic optimization. Then, the PNN model was applied to engineering practice. Only 1 of 20 test samples was misjudged, achieving the 95% prediction accuracy. Finally, the comparison analysis with the artificial neural network (ANN) model, support vector machine (SVM) model, and random forest (RF) model was conducted. Among them, the PNN model achieved the highest prediction accuracy, followed by the artificial neural network (85%), RF (85%), and SVM (80%). In addition, the PNN model had the fastest running speed, which only consumed 5.6350 s, while the running time of ANN, SVM, and RF was 8.8340, 6.2290, and 6.9260 s, respectively. The hybrid PCA-IWGO-PNN model developed in this research provides an effective method for surrounding rock squeezing classification, and it has superiorities in both prediction accuracy and running speed.