Convolutional neural networks optimized with an improved butterfly optimization algorithm for fault diagnosis

Abstract

Abstract The butterfly optimization algorithm (BOA) is a novel swarm intelligence optimization algorithm, which simulates the process of butterfly foraging and courtship. However, BOA suffers from low search accuracy, slow convergence, easily to fall into local optima. To overcome this shortcoming, this paper proposes an improved butterfly optimization algorithm (IBOA). The main idea is to balance the exploration and exploitation of the algorithm by improving the update method of butterfly position. IBOA adopts dynamic switching probability, and balances the global search and local search of a single butterfly by adding an adjustment operator in the global search phase and a sine-cosine operator in the local search phase. This takes full advantage of BOA's global and local searches and enhances communication between butterflies. In order to prove the effectiveness of the IBOA, some benchmark functions are used to verify it. It turns to that the IBOA algorithm is superior to other algorithms. On this basis, IBOA is used to optimize the hyperparameters of convolutional neural network (CNN), and a fault diagnosis model is established. The experimental results of Paderborn bearing data set and continuous stirred tank reactor(CSTR) process data set show that IBOA-CNN model can effectively diagnose industrial data with high diagnosis accuracy, and has obvious advantages compared with other optimization algorithms combined with CNN model.