Monitoring roundness profiles based on an unsupervised neural network algorithm

In modern manufacturing, approaches for profile monitoring can be adopted to detect unnatural behaviors of production processes, i.e. to signal when the relationship used to represent the profiles changes with time. Most of the literature concerned with profile monitoring deals with the problem of model identification and multivariate charting of parameters vector. In this paper, a different approach, which is based on an unsupervised neural network, is presented for profile monitoring. The neural network allows a computer to automatically learn from data the relationship to represent in-control profiles. Then, the algorithm may produce a signal when an input profile does not fit to the prototype learned from the in-control ones. The neural network does not require an analytical model for the statistical description of profiles faced (model-free approach). A comparison study is provided in this paper, in which the Phase II performance of the neural network is compared to that of approaches representative of the industrial practice. Performance is assessed by computer simulation, with reference to a case study related to profiles measured on machined items subject to geometrical specification (roundness). The results indicate that the neural network may outperform usual control charts in signaling out-of-control conditions, due to spindle-motion errors in several production scenarios. The proposed approach can be considered a valuable option for profile monitoring in industrial applications.