A PSO-ANN hybrid model for predicting factor of safety of slope

Abstract

Stability analyses of slopes are routinely done to identify potential landslide sections. This is done for applying proper mitigation techniques. The stability of slopes is generally defined by using the term factor of safety (FOS). Various analytical and numerical models for calculating the value of FOS of slopes have been given by researchers. They had identified various independent parameters affecting the stability of slopes. The relation between FOS and influencing parameters is non-linear, which makes it a complex, multivariate problem. Recently intelligent systems have gained acknowledgement in solving such highly complex, multivariate problem. Artificial neural network (ANN), have got high degree success in solving such problem. However, ANN has its own limitations. Various researchers identified slow learning rate and getting trapped in local minima as the potential shortcoming of ANN. These shortcomings are associated with optimisation algorithm in the ANN. A number of metaheuristic optimisation algorithms are coupled with ANN to increase its performance. Particle swarm optimisation (PSO) had recently gained acknowledgement for solving continuous and discrete optimisation problems. The present study discusses in detail the application of PSO-ANN hybrid model for predicting the FOS of the slope. A three layer PSO-ANN hybrid model have been developed for the present study. A comprehensive database of 83 natural slope sections which were analysed for circular failure has been compiled from different literatures. The network has been trained, validated and tested using the database developed. An extensive iterative programme has been carried out to obtain the value of parameters associated with PSO-ANN hybrid model. Six different numbers of neuron in hidden layer, six different number of swarm size and five different value of acceleration factor (c1 and c2) have been considered. A total of 900 runs have been made, each having a maximum of 2000 iteration. On the basis of the result of 900 runs, a 6-9-1 network was found to be the most optimum network giving the minimum RMSE and Higher R2 value for both training and testing set of the data. The proposed hybrid model is compared with conventional slope stability methods with the help of a case study.