Semi-Supervised Data Distillation for Deep Soft Sensing in Industrial Cyber-Physical Systems

Abstract

While the performances of current industrial data-driven soft sensors are mainly improved through the model aspect, little attention has been paid to the data aspect which has been considered as an even more important element from the recent viewpoint of data-centric engineering and artificial intelligence. In this paper, a new exponential weighted moving average data distillation method is proposed for the design of a semi-supervised soft sensor for industrial cyber-physical systems. Through the introduction of a data distillation method, additional unlabeled data samples can be naturally incorporated for semi-supervised soft sensor modeling, without increasing the number of training data samples, thus the computational complexity of the model can be well controlled. Particularly, a new condensed labeled dataset is formulated, which contains information from both original labeled and unlabeled datasets. Another benefit of data distillation lies in its potential solution for the problem of timeline alignment between input and output variables in soft sensor modeling, which has been ignored in most current works. Two industrial case studies are provided for performance evaluation of the developed method, based on which both feasibility and effectiveness are confirmed.