A novel method based on Bayesian regularized artificial neural networks for measurement uncertainty evaluation

Coordinate measuring machines (CMMs) are complex measuring systems that are widely used in manufacturing industry for form, size, position, and orientation assessment. In essence, these systems collect a set of individual data points that in practice is often a relatively small sample of an object. Their software then processes these points in order to produce a geometric result or to establish a local coordinate system from datum features. The subject of CMM evaluation is a broad and multifaceted one. This paper is concerned with the uncertainty in the coordinates of each point within the measuring volume of the CMM. Therefore, a novel method for measurement uncertainty evaluation using limited-size data sets is conceived and developed. The proposed method is based on a Bayesian regularized artificial neural network (BRANN) model consisting of three inputs and one output. The inputs are: the nominal coordinates; the ambient temperature; and the temperature of the workpiece. The output is the measured (actual) coordinates. An algorithm is developed and implemented before training the BRANN in order to map each nominal coordinate associated with the other inputs to the target coordinate. For validation the model is trained using a relatively small sample size of ten data sets to predict the variability of a larger sample size of ninety data sets. The calculated uncertainty is improved by more than 80% using the predicted variability compared to the uncertainty from the limited sample data set.