Abstract:

Aiming at the problem that the key parameters of complex chemical process are difficult to be accurately soft-measured due to the non-stationarity and event-driven characteristics, an event-driven deep belief network (EDDBN) soft-sensing model is proposed. First, the operating data of chemical process is obtained and a deep belief network (DBN) model is built. The DBN model is trained in a data-driven way to obtain a soft sensor model based on DBN. Second, some events are defined based on the training-error characteristics of the DBN model. The learning step of parameters in DBN model will be accelerated when the positive events occur, and skip the current data sample and directly go to the next data sample. This event-driven selective learning strategy not only efficiently optimizes the training process of soft-sensing model, but also reduces the computational complexity. Meanwhile, this paper analyzes the boundedness of difference between performance potentials from two consecutive events by construct Markov chain-based dynamicl earning process, which gives convergence analysis of EDDBN training process. Finally, the EDDBN-based soft-sensing model is used to predict the concentration of SO₂ in wet-flue-gas desulfurization system. The results show that it can efficiently and accurately predict the concentration of SO₂ under such non-stationary operating conditions, and the computational complexities of data set ① and data set ② are nearly reduced by 63.83% and 63.33%, respectively.

Key words: event-driven learning, deep belief network, soft-sensing, chemical process, wet-flue-gas desulfurization system

摘要：

针对复杂化工过程非平稳性、事件驱动性导致的关键指标参数难以精确软测量的问题，提出了一种事件驱动的深度信念网络（event-driven deep belief network，EDDBN）软测量模型设计方法。首先，获取化工过程运行数据并搭建深度信念网络（driven deep belief network，DBN）模型，以数据驱动的方式对DBN模型进行训练，获得基于DBN的软测量模型。其次，根据DBN模型的训练误差变化特性定义事件，当积极事件发生时会加速当前模型参数的学习步长，当消极事件发生时会跳过当前数据样本并直接进入下一时刻的数据样本学习。这种事件驱动的选择性学习策略不仅能够有效地优化软测量模型训练过程，而且还能降低计算复杂度。同时，通过构造基于马尔可夫链的动态学习过程，分析任意连续两次事件对应输出性能势之差的有界性，给出了EDDBN训练过程的收敛性分析。最后，将EDDBN软测量模型用于湿法烟气脱硫系统二氧化硫（SO₂）浓度软测量实验，结果表明所提出的EDDBN软测量模型能够在非平稳运行工况下实现对SO₂浓度快速、精确地预测分析，并且计算复杂度在数据集①和数据集②上分别降低约63.83%和63.33%。

关键词: 事件驱动的学习, 深度信念网络, 软测量, 化工过程, 湿法烟气脱硫系统