Automated Testing of Advanced Cutting Tool Materials

Tool condition monitoring is an area of research where the main focus is predicting time to tool failure. Ultimately, the goal in the manufacturing industry is to establish a production line with a predictive maintenance strategy in order to reduce waste. For the industrial partner and specialised tool materials developer Element Six Ltd., the aim is to predict tool wear in order to increase the speed at which new tool material grades are tested. Owing to the complex nature of machining where numerous parameters affect the characteristics of tool wear, prediction has been a challenging pursuit. For instance, cutting conditions, workpiece and tool material compositions, machining vibrations, forces and temperatures inside the machine are some of the many factors that can influence a tool's behaviour. This makes tool wear predictions extremely difficult. Another crucial challenge in tool condition monitoring is the availability of descriptive labels of tool wear states. It is not possible to directly acquire damage labels whilst the tool is in use. Consequently, collecting machining datasets is expensive and time consuming as the process must be interrupted to obtain damage labels. It is the aim of this thesis to address these challenges in four steps. Firstly, the experimental procedure used to collect acoustic emission data and tool wear scans from a set of turning tests is detailed. By replicating the procedure already in place at Element Six Ltd., it is possible to ensure that the collected data is representative of the actual machining process, and therefore any prediction algorithms presented later in the thesis are feasible to implement. The second step focuses on finding novel labels that may provide a better description of tool damage than ones previously available. By adopting a new semi-automated technique of feature extraction from 3D models of worn tools, it is possible to study the wear scar progression across the tool for the first time. Using a pragmatic ranking approach features that best represent the machining process and therefore the generated acoustic emissions can be found. It is important to extract features that correlate well with damage labels in order to make accurate predictions later on. Consequently, acoustic emissions generated during turning is explored in the time and frequency domain in the following step. Analysis of chips formed during machining is used to aid in further understanding of these signals. A supervised learning approach using linear response surfaces is then adopted as a prognosis tool to predict damage for previously unseen tools. This method, along with Gaussian process regression, is implemented due to the availability of multiple inputs to predict a single output. For Element Six Ltd., this method can be applied when testing multiple tools of the same material composition following an extensive training period. Finally, the thesis incorporates the idea of online unsupervised learning method to warn the operator of imminent tool failure. Dirichlet process mixture models adopt the use of incomplete label sets and alleviates the need for pre-labelled training data, allowing the testing of entirely unseen tools. The costs associated with data collection can be reduced significantly and interruptions can be kept to a minimum. The Dirichlet process mixture model allows online clustering as the data is collected without the need to set the number of possible clusters a-priori. This will enable operators to visualise when the characteristics of the cutting process change during machining, avoiding the need for exhaustive measures for tracking tool wear, or the early disposal of tools, depending on the context.