Feature Engineering and End-to-End Deep Learning in Tool Wear Monitoring
Machinery health monitoring is an important tool for improving the productivity of today's manufacturers. Two data-driven techniques are prominent in machinery health monitoring: feature engineering and end-to-end deep learning. The research presented here explores these two data-driven techniques in the context of tool wear monitoring during metal machining. Feature engineering was used as a practical approach to predict when CNC machine tools were worn. A robust data-pipeline was developed, and multiple features and models were tested. The best model achieved a precision-recall area-under-curve (PR-AUC) of 0.349, an acceptable result for the imbalanced data set. For the first known time in tool wear monitoring, and as a demonstration of end-to-end deep learning, self-supervised learning with a disentangled-variational-autoencoder was used. An anomaly detection technique was employed to make predictions in both the input and latent spaces. The method achieved a top PR-AUC score of 0.45 on a common milling data set. However, the approach generalized poorly across the combination of parts on the industrial CNC data set, but a top score of 0.41 was achieved for certain parts. Further data collection will enhance the results for all approaches. Overall, the work highlights the abundance of opportunities at the intersection between machine learning and machinery health monitoring.
URI for this recordhttp://hdl.handle.net/1974/28150
Request an alternative formatIf you require this document in an alternate, accessible format, please contact the Queen's Adaptive Technology Centre
The following license files are associated with this item: