针对现代流程工业过程变量存在的多变量、非线性、时滞及测量异常值等问题,提出一种基于门控循环单元(gate recurrent unit,GRU)的鲁棒软测量算法.首先,利用历史数据训练一个初始GRU模型.其次,提出一种改进的最小绝对Lp惩罚解(improved least absolute Lp penalized solution,ILAPPS)的算法,将中值绝对偏差融入截断平均绝对误差损失,使其对异常值有更强的抵抗力,并将基于随机森林的置换重要性排序设计自适应算子嵌入Lp正则化项,提高输入变量选择的准确度.再次,将ILAPPS算法嵌入GRU输入层和循环层,以实现GRU的稀疏化.最后,将所提出的算法应用于人工数据集和实际工业铜矿浮选过程,并与其它先进算法进行性能比较,实验结果表明所提算法具有较高的预测精度和鲁棒性.
A soft sensor algorithm for gated recurrent unit based on improved least absolute Lp penalized solution
A robust soft sensor algorithm based on gate recurrent unit(GRU)is proposed to address the issues of multi-variability,nonlinearity,time delay,and measurement outliers in modern process industrial process variables.Firstly,an initial model of GRU is trained by historical data.Secondly,an improved least absolute Lp penalized solution(ILAPPS)algorithm is proposed,which incorporates median absolute deviation into the capped mean absolute error loss,making it more resistant to outliers;and embeds an adaptive operator based on permutation importance ranking design of random forest into Lp-regularization term to improve the accuracy of input variable selection.In addition,ILAPPS is embedded in the GRU input and recurrent layers to achieve sparse optimization of the network.Finally,the proposed algorithm is applied to artificial datasets and actual industrial copper ore flotation processes,and its performance is compared with other advanced algorithms.The experimental results show that the proposed algorithm has higher prediction accuracy and robustness.
gated recurrent unitrobust estimationsparse optimizationsoft sensorcapped loss
NETL
NSTL
万方数据
维普
