MACHINE LEARNING-BASED THROUGHPUT OPTIMIZATION FOR SERIAL PRODUCTION LINES

Abstract

Production systems require efficient planning and design to meet output objectives or client requirements. Performance assessment in these systems is particularly challenging. Understanding average performance enables a more effective system design and management, and helps in part with distribution and meeting daily targets. Performance is assessed through a discrete-time system consisting of k machines, each with an independent processing time, and a finite buffer capacity that controls the output. Probabilistic rules govern the machines' state transitions, with the number of machines and buffer capacity determining the total system states. However, large-scale systems face dimensionality challenges, making precise computation impossible, while machine learning-based methods provide an efficient alternative. The study aims to efficiently determine the average throughput of serial production lines using machine learning predictive models, exploring various models, and developing a model capable of predicting throughput for varying lengths of production lines. Synthetically created serial production lines modeling machine parameters and buffer capacities derived from machine efficiencies with two distinct ranges, a broader range (50%–97%) for validation against proven cases, and a narrower range (90%–97%) reflecting operational conditions were used in a simulation to obtain the throughput. Then, predictive machine learning modeling of several families, including linear models, tree-based ensemble methods, instance learning models, and multiple deep neural network architectures was conducted. LightGBM proved it was the best model for predicting the throughput with R^2 of 99.60% when number of machines was two, to a R^2of 83.93% when it was 20 followed by the XGBoost model with R^2 of 99.59% and R^2of 77.23% respectively suggesting boosting models are the best choice for modeling SPL without deteriorating performance as the number of machines increase. While linear models, Decision Trees, and KNN formed the lowest-performing group, their performance degraded as the length of the production line increased. Deep learning models demonstrated a great ability in recognizing these configuration sequences. These results highlight that applying machine learning in the design and manufacturing industry has great potential for predicting the throughput.