A decision support system for interactive planning and scheduling in smart manufacturing

• Main Author: Chua, Ping Chong
• Other Authors: Moon Seung Ki
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Industrial engineering
• Online Access: https://hdl.handle.net/10356/166319
• Institution: Nanyang Technological University

High uncertainty in terms of incoming production orders, supply chain disruption and dynamic shop-floor condition are critical issues faced by manufacturing firms in daily production. These uncertainties pose extreme challenges for companies to handle production planning and scheduling in disconnected production environments, especially in production systems with legacy equipment. Understanding the interactions between various components of a production system is imperative to enable better coordination of production planning and scheduling to increase productivity and minimize operating cost. Before a production schedule is generated to process the incoming production orders, the production planning stage should take place. Production planning involves determining the input plan of the required resources and production quantity. Given the multiple input parameters required in the production planning and scheduling, the difficulty arises in seeking improvements in production performance due to decision makers being unsure of key parameters to improve performance. Especially during the iteration phase in optimization process, selecting and optimizing the wrong combination of input parameters can lead to a greater deviation between planning and scheduling. With the possibility of one or more input parameters affecting performance measures such as tardiness, the importance lies in getting the right combination of an input parameter plan to feedback to the original plan for greater synchronization between the planning and scheduling. Current studies deal with modification of the production plan at the iteration phase via ad-hoc or preselection of production parameters to be modified based on an expert’s opinion. By making use of historical production data retrieved within smart manufacturing setting, this research aims to develop an interactive planning and scheduling decision support system (IPS-DSS) for the generation of production plan in iterative production planning and scheduling performed offline to improve production performance in a systematic and rigorous manner. The proposed IPS-DSS consists of two parts. The first part aims at developing a surrogate based adaptive annealing genetic algorithm (AAGA) to generate the production plan during the revision stage of iterative production planning and scheduling. The surrogate based AAGA consists of two stages, i.e., construction of surrogate model and AAGA. Based on three categories of production data collected, namely current production system load, product-based and machine-based parameters, the surrogate model is constructed using multivariate adaptive regression spline (MARS). The surrogate model is used to minimize the number of expensive evaluation of the production performance from the production schedule. The surrogate model is then used as fitness function for the AAGA stage which is used to generate the different combination of the production plan based on the input variables from the surrogate model. The application of the proposed surrogate based AAGA is demonstrated using a simulation model based on an industrial case study involving a wafer fabrication production line. Based on a random sampling of varying number of training data sets, surrogate models of varying highest degree constructed using MARS outperforms that of linear/nonlinear polynomial regression by showing a high correlation coefficient of more than 0.94 and a large reduction in the number of input parameters of more than 96% for both linear and nonlinear cases with relation to three performance measures, i.e., flowtime, tardiness, and machine utilization. In addition, using the selected surrogate model as an instance of the surrogate based AAGA, most instances show a greater improvement in production performance for approximately half or more of the fifteen test cases after three iterations as compared to the modification of production parameters through trial and error. Traditional dispatching rules may not achieve the desired performance target. With dispatching rule forming part of the machine-based parameters, the second part of the IPS-DSS involves the application of deep reinforcement learning for the selection of customized dispatching rules for each machine to increase the number of combinations of machine-based parameters within the existing production plan for further improvement of production performance. Each dispatching rule is formed through the importance selection and stochastic weighted combinations of production parameters obtained within the production system. In addition, the selection of customized dispatching rule to determine the next production order to be processed is based on the current state of the single machine. Using fifty instances of a single-machine production scheduling scenario, customized dispatching rules obtained from the deep reinforcement learning approach demonstrates better performance in the minimization of total tardiness as compared to well-known existing dispatching rules. With the implementation of IPS-DSS, the system is able to produce the appropriate production plan using updated offline data retrieved within smart manufacturing setting for production planners and/or schedulers in a more computationally efficient and informed manner.