Abstract: The storage reliability is an important part of quality monitoring in the military reserve. The scientific and accurate prediction of storage reliability is a necessary requirement for the evaluation of the modern military. For the historical storage data, a prediction model of storage reliability is derived based on period and reliability. A prediction method of storage reliability is provided where BP neural network algorithm is optimized by improved particle swarm optimization algorithm based on evolutionary strategy. Data augmentation can improve the quality and quantity of the samples. The improved PSO algorithm can optimize the initial weights and thresholds in BP neural network, and increase the generalization ability. PSO algorithm has better global search ability, and BP neural network has strong local search ability. The proposed algorithm in this paper can avoid precocious phenomena and improve its convergence speed and its prediction accuracy. The prediction results show that the improved-PSO-BP network method proposed in this paper has better prediction performance than PSO-BP and BP neural networks.

Key words: prediction, storage reliability, BP neural network, particle swarm optimization

摘要： 贮存可靠性是军事储备质量监测的重要环节,科学准确地预测贮存可靠度是现代化军事评估的必然要求。针对历史贮存数据,建立可靠度与年限的贮存可靠性预测模型,采用进化策略改进粒子群算法(PSO)优化BP神经网络进行贮存可靠性预测。通过数据扩充提高样本质量和数量,应用改进后的PSO算法优化BP神经网络的初始权值和阈值,提高网络的泛化能力。PSO算法较好的全局搜索能力与BP网络很强的局部搜索能力相结合,能够避免早熟现象,提高算法的收敛速度及预测精度。实验结果表明,改进的PSO-BP网络模型比PSO-BP和BP神经网络获得更好的预测性能。

关键词: 预测, 贮存可靠性, BP神经网络, 粒子群