基于萤火虫BP神经网络的轴承故障诊断研究

针对BP神经网络训练过程易陷入局部极值导致训练误差收敛速度慢的问题,提出将具有全局寻优的萤火虫算法,结合BP算法共同训练神经网络。在本质上,萤火虫BP神经网络利用萤火虫算法对神经网络进行早期训练,避开局部极值点,得到优化后的神经网络初始权值后,利用BP算法的局部寻优特性对网络做进一步精细训练。轴承故障实验表明,萤火虫BP神经网络的训练误差收敛速度相比BP神经网络、萤火虫神经网络显著提升,故障识别率最高达到99.47%。

Bearing fault diagnosis based on FA-BP neural network

Due to the slow rate of error convergence in training BP neural network which was easy to be trapped in local minimums,firefly algorithm （FA） with a global optimization capacity was proposed to combine BP algorithm to train the neural network together. In essence, FA-BP neural network utilized firefly algorithm to train the network preliminarily to avoid to be trapped in local minimums, optimized the initial weights, and then utilized BP＇s local optimization capacity to make a further training. Test on bearing fault diagnosis showed that FA-BP neural network was superior to FA neural network and BP neural network in the converging rate of training error obviously, with a 99.47% rate at most.