APFC: Adaptive Particle Filter for Change Point Detection of Profile Data in Manufacturing Systems

Change point detection is critical in quality inspection and assessment in manufacturing systems. As one of the most popular Bayesian inference techniques, particle filter algorithm has been successfully applied to estimate the change points of profile data in various manufacturing processes. However, particle filter is computationally expensive, which hinders its wide application for online change point detection. To overcome this challenge, we propose an adaptive particle filter algorithm (APFC) for online change point detection in this paper. With the full consideration of change mechanism, the particle sizes are adaptively selected for parameter estimation as time evolves. The proposed method is validated through extensive simulation studies and two real cases of pipe tightening process and nano manufacturing process.
Note to Practitioners
— This article is motivated by the problem that the particle filter algorithms have large computational costs when applying for the change point detection of profile data in manufacturing processes. Existing implementations of the particle filter algorithms for change point detection problems usually use a fixed and large particle size for estimations, which results in a large computational cost. The fact is that the computational costs can be reduced with smaller particle sizes. However, how to reduce the particle sizes yet keep the accuracy of change detection is challenging. To address this challenge, we develop an adaptive scheme to select the particle sizes to reduce the computation load. More particle sizes are used near the change point to keep the detection accuracy, while particle sizes are reduced to speed up the estimation process when the state of manufacturing processes is stable. We also provide mathematical support for setting the hyperparameters in the APFC framework. To better apply our method in the general change point detection problems, domain knowledge of the corresponding manufacturing processes needs to be considered for the hyperparameter settings.