基于BP神经网络响应面的BLISS研究及应用

Research and application of BLISS based on BP neural network response surface

在标准BLISS方法基础上,依据神经网络的全局映射性,提出了基于BP神经网络建立学科层优化目标函数与系统变量之间的响应面近似模型,并以绝对和相对形式描述的误差函数替代了传统BP算法中单一的绝对形式描述的误差函数。选用样本点在参数空间分布均匀程度更高的试验设计方法——CVT(Centroidal Voronoi Tessellations)试验设计方法来产生训练样本和测试样本,从理论上保证了近似模型的精度。最后利用多属性决策法从算法实施的难易度、优化结果准确性、系统级计算量、算法鲁棒性及收敛性5个方面来评估多学科可行方法(MDF)、改进二级系统合成一体化优化方法(BLISS)的综合性能,定量说明改进BLISS方法更加适合YM160锚杆钻机动力头优化设计。

Based on standard Bi-Level Integrated System Synthesis（BLISS）,according to global mapping of neural networks,the response surface approximation model between subject layer optimization objective function and system variables was proposed based on BP neural network,and the error function described in single absolute form in traditional BP algorithm was replaced by that described in absolute and relative forms.The Design Of Experiment（DOE）-Centroidal Voronoi Tessellations（CVT） is selected to generate training samples and test samples of neural network,in which uniform distribution of sample points in parameter space is high,ensuring the accuracy of approximation model in theory.Finally,the multi attribute decision making algorithm was utilized to evaluate comprehensive performance of Multi Disciplinary Feasible（MDF） method and BLISS from five aspects： difficulty of algorithm implementation,the accuracy of optimization results,system level computation,algorithm robustness and the convergence.The result quantitatively indicates that the improved BLISS method is much more suitable for the optimization design of YM160 roofbolter power head.