Development of an Autonomous, Self-Optimising Machine Learning Framework for use in Manufacturing Applications

Abstract

Traditionally manufacturing processes have used a wide variety of sensors to provide process control, quality and productivity metrics. Industry 4.0 introduces the concept of combining and analysing sensor derived production data to offer greater insights into all aspects of manufacturing operations. A key area of advancement is the ability to process and analyse large volumes of data, in real-time, to provide to process control. The objective of this thesis is to define a framework that enables adaptive, optimised production control via the use of automated machine learning algorithm selection.

The complete framework is capable of adaptively processing streamed production data directly from manufacturing processes along with other appropriate sources. To achieve this, a second longer term channel is used to autonomously evaluate and optimise competing algorithms and data strategies to select the most appropriate solution for near real-time control. A series of experiments demonstrates that the framework is suitable for production control applications which benefit from a focus on single or multiple accuracy metrics.

The framework can automatically switch algorithms via a supervisory mode. This allows it to take into account additional factors such as concept drift, production changes, computational complexity and algorithm stability. Algorithm switching is based on a flexible optimisation strategy for algorithm performance over variable time periods. Experimental results are provided for two optimisation methodologies.

Further development of the framework will involve full automation of the real-time algorithm selection method. This will remove the need for specialist data processing knowledge and hence allow the framework to be deployed in a wide range of existing manufacturing companies.