Predicting the Distribution of Product Completion Time in Multi-Product Manufacturing Systems

Product completion time is a random variable resulting from the random disturbances in production systems that delay the processing of products unexpectedly. Existing methods for product completion time prediction mostly predict its mean value. However, mean value only accounts for the first moment of a probability distribution, and is not sufficient for depicting the full spread of the product completion time. In this paper, we propose a novel method for predicting the probability distribution of production completion time by combining system model and deep learning. The original data collected from the plant floor are boosted through a model-based oversampling process. The location family of Tweedie distribution is discovered to fit the distribution of product competition time well. A hybrid framework is established to predict distribution parameters given system state as input, so as to predict the completion time distributions in a real-time fashion. The location parameter is analytically evaluated with system model. Other parameters are predicted or determined with data-driven methods, including a long-short term memory network and classic Tweedie prediction techniques.