A two-step machining and active learning approach for right-first-time robotic countersinking through in-process error compensation and prediction of depth of cuts

Robotic machining processes are characterised by errors arising from the limitations of the industrial robots. These robot-related errors can compromise the overall manufacturing process performance, resulting in final products with dimensions different from the nominal specifications. To avoid accumulation of errors through several manufacturing stages, a quality inspection step is usually performed after the cutting operation. This work presents an innovative two-step manufacturing method for achieving right-first-time characteristics in robotic machining operations through in-process inspection and compensation of the systematic errors, whilst collecting suitable training data for building predictive models. The key idea behind the proposed method is based on the observation that under certain conditions, the robotic machining errors remain largely consistent, and therefore by splitting the process into two similar steps and having an inspection step in between, a prediction and then compensation of the systematic errors would be possible. A Gaussian Process Regression (GPR) framework is applied for the creation of robust process models that predict the post-process inspection result from in-process signal features, with the associated confidence intervals. An active learning algorithm that makes online decisions on the inspection task based on the current confidence of the models, is also proposed. The two-step machining method and the active learning approach were both tested on a robotic countersinking process experiment. The results showed that the in-process inspection and error compensation of the proposed two-step machining method was able to achieve final countersink depths very close to the desired target, confirming the potential for right-first-time robotic machining. In addition, the active learning results highlighted the ability of the algorithm to reduce the number of required post-process inspections, thus saving both time and costs, whilst also identifying novel data relevant for the model training.