Our Research

Investigating Machine Learning

This research aims to develop toolkits for robust, resilient and transferable Cyber Manufacturing Systems (CMS).

Background

Digital Manufacturing relies on pervasive and ubiquitous use of Information and Communication Technology (ICT), sensors, intelligent robots to deliver the next generation of intelligent, co-operating and interconnected manufacturing systems. The research is aimed to improve techniques that can be used to develop digitalised manufacturing systems to reduce existing inefficiencies in production processes that impact on production costs, unplanned downtime, quality and yields. This is not only detrimental to manufacturing businesses but has a negative impact on the UK Economy. The current productivity levels of UK manufacturers and suppliers is lagging behind global competitors and prevents the UK from successfully competing with other countries in the manufacturing domain - which is vital to keep businesses and jobs in the UK rather than relocate production abroad.

The UK Government wants to increase the strength of the UK Manufacturing Sector. A key means of doing this is the widespread adoption of industrial digital technologies (IDT). Cyber Manufacturing Systems (CMS) are the building blocks of digitalised manufacturing and generate vast amount of data that can be used for real time decision making to achieve optimised performance through predictive and prescriptive analytics. The latter are techniques that use, combine and analyse available data to develop computational models that can predict future outcomes and determine the best course of action. The research, under the fellowship, solves some of the existing problems in this area (CMS), developing new techniques and resources for predictive and prescriptive analytics with the potential to increase efficiency, accuracy and productivity of manufacturing processes. Businesses are therefore more likely to adopt IDTs and improve profitability and sustainability and provide high-quality jobs in a thriving part of the economy.

Aims & Objectives

This project studies novel and robust data analytics methods that enable us to build predictive models that take into account uncertainty, complexity and dynamic behaviour of productions systems.

The project involves:

Objective 1 - developing algorithms that can reuse previously acquired data/knowledge to build more accurate predictive models that work well in the presence of noise (i.e. 'robust'), are able to adapt to changes over time (i.e. 'resilient') and can be scaled up across multiple factories (i.e. 'transferable').
Objective 2 - developing and testing novel non-parametric methods for estimation of uncertainty and risks associated to a decision to enable real time mission and safety critical decision making (both automated and human driven) based on predictions.
Objective 3 - iteratively developing, deploying and testing predictive and prescriptive models in real and simulated industrial scenarios to obtain acceptable level of performance, usability and robustness.

Our industrial collaborators are Crown Holdings, TATA Steel, Sony and Vortex IoT - they are providing labelled and aggregated datasets for testing the methods through computer simulations and enabling feasibility studies to be conducted in factory environments.

The outcomes of the research are ultimately to improve the quality of products, achieving less wastage and unnecessary costs. Through increased adoption of IDTs, the production of goods will, importantly, be more efficient, reliable and profitable. This will support the regeneration of the Manufacturing Sector and boost the global competitiveness of the UK.

The project is funded through Engineering and Physical Sciences Research Council.

Three interacting work packages have the following objectives:

Research to develop novel, self-adaptive and robust Machine Learning (ML) methods for CMS based on Transfer Learning (TL) for predictive operations of CMS:

Study Transfer Learning approaches including domain adaptation and multi-tasking learning and their applications for timeseries prediction, prediction from noisy variables and to improve scalability of predictive models
Develop and test novel application of TL to the manufacturing domain using real world data sets provided by industrial partners
Demonstrate suitability of TL to build predictive models that can capture systems dynamics of CMS and are robust to noise
Publication and dissemination of results
This objective will require industrial engagement to identify requirement and assumptions and will give a steer towards the most promising TL methods and provide suitable data for testing.

Develop new methods to improve decision-making capabilities of CMS in dynamic and uncertain environments for prescriptive operations of CMS:

Develop and test novel methods to embed uncertainty estimations in predictions in methods developed in Objective 1
Develop metrics and methodology to evaluate risks associated to uncertainty of predictive models
Develop a framework for decision making under uncertainty for CMS
Publication and dissemination of results
Industrial partners will engage in discussion and identification of requirements for prescriptive operations of CMS.

Develop a Machine Learning toolkit and recommendations for CMS:

Integrate methods developed in Objectives 1 and 2 to test applicability to real factoring environment
Develop proof of concepts (PoC) to demonstrate the applicability in real or simulated scenarios in diverse factory environments (assembly, packaging, and heavy process manufacturing)
Further refine ML methods based on performance, usability and robustness
Develop a generic toolkit and recommendations for building prescriptive and predictive models for CMS
Following successful PoC write further grant applications for technical demonstrators
Involvement of industrial partners will enable identification of use cases and facilitate testing.