Skip to Content

Semi-supervised learning for soft sensors in advanced manufacturing

Semi-supervised learning for soft sensors in advanced manufacturing

PhD project title and outline, including interdisciplinary dimension:
Semi-supervised learning for soft sensors in advanced manufacturing

Primary Supervisor: Prof. Seán McLoone, Centre for Intelligent Autonomous Manufacturing Systems (i-AMS), QUB
With the rapid development in sensing, communication and storage technologies companies are now collecting and storing large quantities of data on their manufacturing processes – temperatures, pressures, flow rates, etc. At the same time, measuring the final product quality is generally only done through infrequent sampling, due to the cost and time involved (e.g. testing product quality may require destructive testing, or take several hours). To achieve better control of manufacturing processes and improved efficiency in terms of waste and energy consumption real-time measurements of product quality are desirable. One approach to solving this problem that is an active area of research in advanced manufacturing, and in the pharmaceutical and semiconductors sectors, in particular, is to develop so called soft sensing models that can predict product quality from the available process measurements.

Soft sensing models are generally quite complex and challenging to develop and require the use of sophisticated machine learning and system identification techniques. One of the major challenges with building them is that datasets are often ill-conditioned with a large number of candidate process variables available as model inputs but only a small number of output training samples, by virtue of the restricted product quality sampling regimes normally employed. The current practice is to discard the large volume of data collected which does not have corresponding product quality information. Discarding this data, referred to as unlabelled in the machine learning community, represents a huge waste of resource and a missed opportunity. As such the objectives of this PhD project is to explore techniques for utilizing unlabelled data to enhance model building and to develop methodologies that can make the best use of all available data in the development of robust soft sensors for advanced manufacturing processes.

Primary supervisor: Professor Seán McLoone (EEECS)
Secondary supervisor: Dr Hannah Mitchell (Maths and Physics)
External Partner/Organisation: Irish Manufacturing Research