AI 기반 크롬 도금탱크 농도 예측을 위한 시계열 모델 연구

초록

In the chromium plating process, analyzing the CrO₃ concentration in the bath inherently requires a long lead time of more than 24 hours. Consequently, field operations have largely relied on the subjective experience and intuition of operators to determine the chemical dosage, which frequently leads to process variations and quality instability. Therefore, this paper proposes a weekly time-series forecasting framework to transition from intuition-based decision-making to a proactive, data-driven system. Concentration data accumulated from 2022 to 2024 were used for the training phase, while 26 weeks of actual data from the first half of 2025 served as the evaluation set. This study compared the performance of traditional models, including ARIMA, Holt-Winters, and Prophet, with emerging time-series foundation models, specifically TimesFM and Chronos. The experimental results showed that while ARIMA and Holt-Winters achieved superior accuracy in the stable Tank 3, TimesFM exhibited the most robust predictive performance and operational stability in the volatile Tank 5 and across the integrated evaluations. Furthermore, the proposed forecasting technique could complement operators' judgments and significantly enhance the overall reliability of the manufacturing process by quantitatively analyzing the error characteristics near the quality control limits (234-255 g/L).

키워드