基于DNN的活性炭吸附预测及性能评价

Construction of deep neural network aided adsorption prediction and performance evaluation

传统机器学习在处理有限样本数据时,容易出现过拟合和梯度消失等问题。为解决该类问题,文中以生物质活性炭亚甲基蓝吸附作为研究对象,构建了一种基于深度学习的DNN预测模型。使用Adam算法动态调整学习率、加速网络收敛;采用Dropout函数缓解过拟合;使用ReLU函数作为激活函数解决梯度消失问题。所搭建的DNN模型预测精度和稳健性均显著高于传统的人工神经网络模型,在面对单一来源数据时,预测平均准确率达到99.9%,面对来自不同实验室的多重来源数据时,依然拥有99.8%的平均预测准确率。搭建好的DNN模型以较强的鲁棒性保证了自身的安全性和容错能力,符合数据庞杂且复杂多变的实际应用情况,同时可进行关键影响因子的非线性定量关系预测,从而辅助制备工艺的决策优化。

Traditional machine learning processes are usually restricted by over fitting and gradient disappearance, especially dealing with limited data. In order to solve this kind of problem, this work presented deep neural network approach to rapidly predict methylene blue(MB) adsorption on biomass derived AC. Adam algorithm was used to dynamically adjust the learning rate and accelerate the convergence of the network, and Dropout function was used to alleviate over fitting. Meanwhile, ReLU function was used as the activation function to solve the gradient disappearance problem. The prediction accuracy and robustness of the DNN model are significantly higher than the traditional artificial neural network method. The average prediction accuracy reaches 99.9% when faced with data from a single source, and still has an average prediction accuracy of 99.8% when faced with data from multiple sources in different labs. The constructed DNN model ensures its safety and fault tolerance with strong robustness, which is in line with the practical application of complex and changeable data. Especially, DNN can help to predict the nonlinear quantitative relationship of key influencing factors, so as to assist in the following decision-making and process optimization.