Autonomous supervision and optimization of product quality in a multi-stage manufacturing process based on self-adaptive prediction models

Abstract

In modern manufacturing facilities, there are basically two essential phases for assuring high production quality with low (or even zero) defects and waste in order to save costs for companies. The first phase concerns the early recognition of potentially arising problems in product quality, the second phase concerns proper reactions upon the recognition of such problems. In this paper, we address a holistic approach for handling both issues consecutively within a predictive maintenance framework at an on-line production system. Thereby, we address multi-stage functionality based on (i) data-driven forecast models for (measure-able) product quality criteria (QCs) at a latter stage, which are established and executed through process values (and their time series trends) recorded at an early stage of production (describing its progress), and (ii) process optimization cycles whose outputs are suggestions for proper reactions at an earlier stage in the case of forecasted downtrends or exceeds of allowed boundaries in product quality. The data-driven forecast models are established through a high-dimensional batch time-series modeling problem. In this, we employ a non-linear version of PLSR (partial least squares regression) by coupling PLS with generalized Takagi–Sugeno fuzzy systems (termed as PLS-fuzzy). The models are able to self-adapt over time based on recursive parameters adaptation and rule evolution functionalities. Two concepts for increased flexibility during model updates are proposed, (i) a dynamic outweighing strategy of older samples with an adaptive update of the forgetting factor (steering forgetting intensity) and (ii) an incremental update of the latent variable space spanned by the directions (loading vectors) achieved through PLS; the whole model update approach is termed as SAFM-IF (self-adaptive forecast models with increased flexibility). Process optimization is achieved through multi-objective optimization using evolutionary techniques, where the (trained and updated) forecast models serve as surrogate models to guide the optimization process to Pareto fronts (containing solution candidates) with high quality. A new influence analysis between process values and QCs is suggested based on the PLS-fuzzy forecast models in order to reduce the dimensionality of the optimization space and thus to guarantee high(er) quality of solutions within a reasonable amount of time (→ better usage in on-line mode). The methodologies have been comprehensively evaluated on real on-line process data from a (micro-fluidic) chip production system, where the early stage comprises the injection molding process and the latter stage the bonding process. The results show remarkable performance in terms of low prediction errors of the PLS-fuzzy forecast models (showing mostly lower errors than achieved by other model architectures) as well as in terms of Pareto fronts with individuals (solutions) whose fitness was close to the optimal values of three most important target QCs (being used for supervision): flatness, void events and RMSEs of the chips. Suggestions could thus be provided to experts/operators how to best change process values and associated machining parameters at the injection molding process in order to achieve significantly higher product quality for the final chips at the end of the bonding process.