Production performance prognostics through model-based analytical method and recency-weighted stochastic approximation method

Abstract Distributed sensors are widely adopted in production systems to monitor system status and support daily operation of individual machine. However, the estimation of the potential problems in system performance using sensor information is still lagging behind. To predict production system performance in stochastic scenario, a data-driven model-based analytical method and recency-weighted stochastic approximation method are developed in this paper based on both production system physical properties and sensor information. The analytical method can be used to provide real-time prognostic information on the future behavior of two machine one buffer production lines with exponential machines. For general serial production systems, especially those with non-stationary reliability parameters, the stochastic approximation method utilizes the vast amounts of sensor data for learning system production performance in future time. The learning results will be updated and utilized in real-time production prognosis. The proposed prognostic methods and system real-time prognostic information reveal how current system status can influence the potential production performance in future time, which is critical for real-time production control and management. Numerical studies are presented to demonstrate the accuracy and effectiveness of the proposed methods.