A0866
Title: Anomaly detection in profile monitoring using functional conformal prediction
Authors: Teresa Bortolotti - Politecnico di Milano (Italy)
Egon Prioglio - Politecnico di Milano (Italy)
Bianca Maria Colosimo - Politecnico di Milano (Italy)
Simone Vantini - Politecnico di Milano (Italy) [presenting]
Abstract: A novel methodology is proposed for detecting anomalies in the monitoring of profiles in industrial processes. The approach is grounded in functional data analysis (i.e., by modeling profiles as functions) and integrates conformal prediction and copula-based methods to detect unusual patterns. Theoretically, the control of the probability of having one or more false anomalies (i.e. type I errors) along the functional domain is also guaranteed in the case of small sample sizes and non-Gaussian data, which are common, for instance, in 3D printing applications. The functional control limits are obtained by inverting simultaneous functional conformal prediction bands, which have been proposed in the literature recently. Furthermore, to enhance interpretability and increase the anomaly detection power of the proposed procedure, the methodology is extended by employing copulas to simultaneously monitor functions and their higher-order derivatives. An extensive simulation study showcases the potential of the proposed approach and proves the effectiveness of functional conformal prediction and copula adjustment in detecting anomalies while controlling the probability of false anomalies. The applicability of the methodology is finally illustrated across various industrial applications in the field of 3D printing.
