深度交叉遗传神经网络在液体火箭发动机故障检测中的应用

Application of deep cross-hybrid genetic neural network to fault detection of liquid rocket engines

将遗传算法与BP神经网络深度交叉融合,即采用遗传算法对BP神经网络的权值和阈值进行多点优化,而在进化的每一代中随机取少量染色体进行单一BP网络训练,训练结果再返回染色体,经过若干代的进化后得到稳定的权值和阈值,再将它们赋给BP神经网络,作为初始值,按误差前向反馈算法沿负梯度搜索重新训练,最终得到最优解。这种算法既避免BP算法易陷入局部最优解的不足,又克服遗传算法以类似穷举的形式寻找最优解而引起的搜索时间长、速度慢的缺点。并且经过仿真分析,深度交叉遗传BP神经网络的收敛性和故障诊断能力优于传统BP神经网络和单一使用遗传算法,可有效应用于液体火箭发动机故障检测中。

This paper proposes a new hybrid algorithm based on genetic algorithm and BP neural network. First, multi-point optimization of the BP neural network＇s weights and threshold values in GA algorithm is carried out, and some chromosomes that are random sampled in each generation perform single BP neural network training. The result gained above is returned to the chromosomes. Second, stable weights and threshold values are obtained after the evolution of some generations, then they are used as the initial value to train the BP neural network by seeking along negative grads in error forward feedback algorithm, and finally the global optimum is gained. The proposed algorithm can avoid the deficiency of BP algorithm that may easily be steeped in local optimums, and can also overcome GA＇s shortcomings of long seeking time and low seeking due to the method of enumerating. The results of simulation indicates that the ability of convergence and diagnosis of the proposed algorithm is better than that of traditional BP neural network or only using GA, and the algorithm can be effectively applied to the fault detection of liquid rocket engine.