基于人工神经网络的舰船灵敏度分析

Parametric sensitivity analysis of ships based on neural network

舰船参数的选择对于评价舰船综合性能至关重要,但不同参数在评价过程中所占的比重不一,量化的舰船参数灵敏度分析是后续舰船性能评价、类型识别等的重要前提。本文利用24-10-10-1结构的BP神经网络对舰船的抗沉性、最大航速、适航性、载重量4个重要参数进行灵敏度分析,并建立起各参数灵敏度与舰船综合性能的对应关系。在用BP神经网络对样本进行训练的同时,利用Skeletonization灵敏度剪枝方法计算输入节点和隐节点的灵敏度。测试结果表明,本文的灵敏度分析算法不仅优化神经网络结构,而且学习过程收敛后,可获得各输入节点稳定的灵敏度值。

Ship parameters selection is vital for comprehensive performance evaluation of ships,but different parameters have different weight in the evaluation,the quantitative parameter sensitivity analysis is base for the following ship performance evaluation,type recognition procedure. Using BP neural network with 24-10-10-1 structure, this paper gives sensitivity analysis of four important parameters： heavy resistance,maximum speed of the ship,seaworthiness and deadweight,the corresponding relations were made for ships comprehensive performances and the parameter sensitivity. When training samples with BP neural network,Skeletonization sensitivity pruning method was used to calculate the sensitivity of input and hidden nodes. Results show that the algorithm not only optimizes the neural network structure,but also obtain stable sensitivity for input node after the convergence of learning process.