College of Control Science and Engineering, Zhejiang University
NSFC-Zhejiang Joint Fund for the Integration of Industrialization and Informatization(No. U1709211),Zhejiang Key Research and Development Project (2019C03100), Alibaba-Zhejiang University Joint Institute of Frontier Technologies
在工业领域，数据缺失十分普遍，对解决下游任务如软测量、异常检测造成阻碍，这些任务大多依赖完 整而高质量的数据集构造模型。现有缺失数据填补方法很少考虑数据填补后的具体下游任务——本文中指软测 量。如何根据下游任务针对性地进行数据填补是当前研究中的挑战之一。为此，本文提出了一种加入临时软测 量模块的对抗生成数据填补模型 (Imputation Generative Adversarial Network with Soft Sensor, SSIGAN)。与生成 对抗数据填补模型 (Imputation Generative Adversarial Network, GAIN) 相比，SSIGAN 模型显式地考虑了软测量 损失对数据填补模型的影响，通过临时软测量模块指导对质量相关变量的修复，实现数据填补的“定制化”，用于 更精准的工业软测量建模。本文通过某工业炼钢过程中的终点成分软测量实验验证了所提方法对软测量质量相 关变量缺失数据填补效果以及最终软测量效果的提升。
Missing data is quite common in the industrial field, resulting in problems in downstream applications such as soft sensing and anomaly detection, as most data driven methods used in these applications rely on complete and high-quality dataset to build a high-quality model. Current data imputation methods hardly take its following applications like soft sensing into consideration. A considerable challenge is how to refine missing data repair according to its downstream application. In this paper, we propose an Imputation Generative Adversarial Network with Soft Sensor (SSIGAN) which considers the loss of soft sensors as data is imputed. Compared with the Imputation Generative Adversarial Network(GAIN), the proposed SSIGAN model introduces the influence of data imputation on the soft sensor. The temporary soft sensor model gives guidance for better repair of quality-related variables. Thus, “customized” data imputation can be achieved for building a more accurate industrial soft sensor. An experiment of soft sensing the end-point composition in a steel-making process is conducted and verifies the improvement of data imputation of quality-related variables and that of the soft sensor with the proposed data imputation model.