萤火虫算法作为一种群体智能算法，具有良好的全局寻优特性，可用于解决神经网络在反向传播(Back propagation, BP)算法下易陷入局部极值点、收敛速度慢的问题。但在应用于神经网络时其参数选取缺乏有效依据或经验公式，参数取值不当时，萤火虫神经网络的训练误差无法有效收敛、种群个体之间协同寻优功能易失效。因此，提出一种双层萤火虫改进算法对其参数进行选取。用UCI数据及轴承故障试验对算法进行验证，结果表明，应用所提方法后萤火虫神经网络的训练误差收敛值显著减小，并且在训练效率、故障识别率方面均优于BP神经网络、遗传神经网络、粒子群神经网络。

As a new swarm intelligent algorithm, firefly learning has a good performance of global searching, and it can be used to solve the problems of back propagation (BP) network’s local minima and slow convergence speed. However, there is no guidance for firefly parameter selection when the algorithm is utilized in neural network learning. If parameters are selected improperly, it may lead to that the firefly neural network converging slowly and fail in collaborative optimization. Therefore, a bi-layer firefly algorithm is proposed to improve parameter selection. UCI iris data and bearing fault data are used to validate the proposed approach. Experiment results demonstrate that the converging rate of firefly neural network decreases remarkably by using the proposed method, and considering the training efficiency and fault recognition rate, firefly performs better than BP, particle swarm optimization (PSO) and genetic algorithm (GA).