Abstract:
This research was conducted in multinational doll manufacturing company, to
address high defect rates in blister accessory packaging. The main problem
identified was detached blisters, which accounted for nearly 40% of packaging
defects during drop tests. Root cause analysis using Pareto charts and fishbone
diagrams confirmed that unstandardized ultrasonic welding parameters were the
dominant cause. To address this, a Design of Experiments with a full factorial
design was applied to evaluate the effects of welding time, cooling time, and
amplitude. Random Forest machine learning was then used to enhance prediction
accuracy and validate parameter importance. This study measures weld quality as
the primary response variable, defined as the proportion of blister accessories that
remain attached during standardized drop testing. A defect occurs when blister
accessories detach from the cardboard insert, compromising product integrity and
customer satisfaction. Results showed that welding time and cooling time were the
most influential factors, with amplitude playing a smaller but supporting role. The
optimum parameter range was identified as welding time of 1.0–1.2 seconds,
cooling time of 1.4-1.5 seconds, and amplitude of 50–64%, which reduced defects
from 40% to 13%, or reduced by 22%, equivalent to avoiding 545 defectives in a
2,400-unit run. This research demonstrates that combining DOE with machine
learning provides both statistical evidence and predictive capability, offering a
practical framework to improve packaging quality and efficiency in manufacturing.
