Abstract: To improve the performance of using fine-tune to construct deep neural network surrogate models(named also by DSMs), an active closed-loop Monte Carlo method for design of experiments was presented. The design point was associated with the model gradient through Fisher information matrix, and solved by multiplication algorithm. A random-discretization based Monte Carlo algorithm was then given for closed-loop sampling, so that the design points had the statistical characteristics covering the entire design space. Based on this method, a DSM of action characteristics of a telescopic forklift was constructed with multilayer perception, and combined with minimizing the predictor and expected improvement to realize the design optimization of the hydraulic control systems. Experimental results show that compared with the benchmarks, the simulation data required by the proposed method is reduced by 64.3%. The pressure fluctuation of luffing cylinders for the optimized forklift booms is more stable, and the maximal value is decreased by 46%.

Key words: forklift boom, deep surrogate model(DSM), transfer learning, design optimization

摘要： 为了提升利用微调构造的深度神经网络代理模型（又名深度代理模型）的性能，提出了一种主动闭环蒙特卡罗试验设计方法，通过费雪尔信息矩阵将设计点与模型梯度关联，并利用乘法算法求解，然后引入随机离散蒙特卡罗算法进行闭环采样，使得设计点具有覆盖整个设计空间的统计学特性。基于该方法，利用多层感知器建立了某伸缩臂叉车臂架动作特性深度代理模型，并结合最小预测和预期改善，实现了液压控制系统的设计优化。实验结果显示：与当前基准相比，提出的方法所需仿真数据减少了64.3%；优化后叉车臂架变幅缸压力波动更加平稳，且最大值减小了46%。

关键词: 叉车臂架, 深度代理模型, 迁移学习, 设计优化