Lead Institution: University of Liverpool
Industrial Partner: Renishaw
Project Team: Dr Peter Green (PI, Liverpool), Dr Kate Black (Liverpool), Dr Chris Sutcliffe (Liverpool and Renishaw)
Project Duration: 8 months (August 2017 – March 2018)
Additive Manufacturing (AM) has the potential to bring transformative change to the UK’s manufacturing industry. Innovate UK identified AM as one of the UK’s 22 priority process technologies and described it as a ‘key potential growth area’ for the UK in the 2020s. Unfortunately, large scale adoption of AM methods is severely hindered by uncertainties associated with the quality of printed components. This is particularly true in the healthcare and aerospace sectors, where requirements on part quality are typically the most stringent.
The field of machine learning offers potential solutions to this problem. Machine learning has been successfully applied to several manufacturing challenges and can, potentially, be used to automatically identify faulty components using measurements that are obtained, in real-time, during the AM process (a time series of powder bed temperatures taken during SLM, for example). This typically involves presenting an algorithm with process data which is ‘labelled’ – data where it has been established that a particular manufacturing process has led to a component that is either ‘faulty’ or ‘not faulty’ (this approach to machine learning is commonly called ‘supervised learning’). Once successfully trained, such an algorithm will then be able to classify future components, based solely on their individual process measurements (and without costly manual inspections). Unfortunately, the costs associated with obtaining the necessary quantity of ‘labelled data’ prevent the application of supervised learning to the identification of faulty, additively manufactured components.
This feasibility study aims to deliver an algorithm which, using process measurements obtained during SLM, will automatically certify additively manufactured components without needing large amounts of labelled data. Its premise is that, during the AM printing process, large sets of measurement data can be generated relatively easily but that the establishment of the resulting component’s quality is more time consuming and expensive. The proposed methodology is based on the hypothesis that, by combining large amounts of ‘unlabelled’ data with a small amount of ‘labelled’ data, it will be possible to detect future faulty components with a far greater accuracy than if only a small quantity of labelled data were utilised. This will exploit a field of machine learning known as ‘semi-supervised learning’. The project will utilise novel reflective light technology, recently developed by Renishaw, which makes it possible to generate real time measurements of a component’s temperature distribution and position, during SLM. In the long term, the authors aim to utilise the results of this feasibility study to develop machine-learnt process control for AM.