Developing a Framework for Mixing Tube Tool Condition Monitoring of Abrasive Waterjet Systems

The abrasive waterjet machining process needs to be improved to increase its adoption in industry. The wear of the mixing tube affects cutting performance and must be considered for accurate machining. Frequent downtimes and subsequent costs can deter AWJ adoption. A wear monitoring system can help improve efficiency and lower costs by allowing for scheduled maintenance and better resource management. This thesis aimed to develop a framework for a real-time tool condition monitoring system that can predict mixing tube wear. Data collection challenges were identified to achieve this, and a data collection framework was designed. Wear time and a range of indirect sensor data was collected using the proposed methodology and analysed. The use of machine learning was explored to predict the mixing tube's exit diameter, a commonly used and direct measure of wear, and to classify the tool state using a 10\% exit diameter growth as a wear threshold. Machine learning was used as the problem involved analysing indirect sensor data, which presented challenges such as non-linearity and multivariate relationships that are better addressed by machine learning techniques than traditional analytical methods. The performance of machine learning algorithms using only the sensor data was compared with simpler linear algorithms trained on recorded wear time. The sensor-based machine learning approaches were outperformed by the wear-time-based linear models when evaluated under controlled experimental conditions where variations such as part changes were not considered. For exit diameter prediction, 0.023 and 0.01 root mean squared error scores were obtained for machine learning and linear approaches respectively. For tool state classification, 0.7 and 1.0 F1-scores, which represent the harmonic mean of precision and recall, were obtained for machine learning and linear approaches respectively. However, a hybrid approach using machine learning models trained on both sensor and wear time data was found to achieve the best performance under changing conditions. In conclusion, this thesis proposed a foundation for building a tool condition monitoring system for the abrasive waterjet mixing tube.