Predictive Tool Condition Monitoring in Computer Numerically Controlled Milling

Abstract

The surface integrity and dimensional accuracy of a computer numerically controlled (CNC) milled part is greatly affected by the wear condition of the cutting tool. An effective tool condition monitoring scheme is thus necessary to allow for a timely tool change to safeguard the aforementioned. Even though visual-based techniques, such as optical imaging, capable of providing an outright measure of a tool’s in-process wear condition are available, the impracticalities associated with the measurements in an uncontrolled practical environment inhibits their usage. As such, data-based modeling of monitoring sensory data offers the more viable option, with artificial in- telligence (AI) based techniques of processing this data showing significant promise. However, several challenges associated with the modeling approach inhibit their suc- cessful development and deployment. These are; redundancy and noise in captured data, varying wear distributions of different cutters even on same cutting conditions, data scarcity, complex associations caused by presence of several contributing fac- tors, and uncertainty in predictions from slight variations in model weights. This study thus sought alternative approaches to address these challenges while devel- oping data-based models under both constant and varying machining conditions in case scenarios of ample and insufficient model training data availability. In a high data regime of sufficient model training data, machine learning based deep model- ing was utilized to develop a transformer based wear model architecture capable of processing complex associations from combined sensory data and cutting parameters while automatically eliminating the negative influences of redundancy and noise in captured data. A gated residual network in a parallel processing structure was devel- oped for this task with the consequent results showing its ability in processing raw data automatically irrespective of scale. This resulted in a data model that does not require initial data pre-processing unlike those reported in literature. This model structure was re-used for a case point of insufficient training data availability. In or- der to facilitate model development for this case point, a helper model was developed for generating synthetic data that is statistically similar to collected experimental data. The synthetic data was then utilized in model pre-training, with the generative model used as an alternate inexpensive tool to tackle the data scarcity problem in data based modeling. The end result has been the development of end-to-end re- gressive and classifier wear models capable of processing raw data directly to provide an accurate prediction of a tool’s flank wear and state. Analysis of the performance of the developed models on experimental CNC milling data sourced from public do- main, coupled with comparison of results with other models reported in literature on the same data sets was also carried out. For an experiment under constant cut- ting conditions in a high data regime, the developed wear model attained an MAE of 5.7, 7.3 and 8.5 µm for three cutters under consideration, with a resultant over- all prediction accuracy of 93% which was above the minimum acceptable accuracy threshold of 90%, all without data pre-processing. Under varying cutting conditions in a low training data scenario and with 15 cutters under consideration, an averaged

cross-validated MAPE of 12.5% was attained, with an overall accuracy enhancement of over 25% on a base comparison model only trained on few experimental data samples. The knowledge learnt from the synthetic data enabled this performance enhancement qualifying the adoption of this approach in model development. These results were well within an acceptable wear boundary of ±20% error variation as those of other reported models on same data, qualifying the adopted development approaches.