A typical manufacturing process consists of a machining (material removal) process followed by an inspection system for the quality checks. Usually these checks are performed at the end of the process and they may also involve removing the part to a dedicated inspection area. This paper presents an innovative perturbation signal based data generation and machine learning approach to build a robust process model with uncertainty quantification. The model is to map the in-process signal features collected during machining with the post-process quality results obtained upon inspection of the finished product. In particular, a probabilistic framework based on Gaussian Process Regression (GPR) is applied to build the process model that accurately and reliably predicts key process quality indicators. Raw data provided by multiple sensors including accelerometers, power transducers and acoustic emissions is first collected and then processed to extract a large number of signal features from both time and frequency domains. A strategy for the selection of most relevant features is also explored in this work in order to reduce the input space dimension and achieve faster training times. The proposed GPR model was tested on a multi-robot countersinking application for monitoring of the machined countersink depths in composite aircraft components. Experimental results showed that the model can be used as a tool to predict the part quality through in-process sensory information, which in turn, helps to reduce the total inspection time by identifying the parts that would require further investigation.
CORE (COnnecting REpositories)
CORE (COnnecting REpositories)