ARIMA-AdaBoost hybrid approach for product quality prediction in advanced transformer manufacturing

Abstract

End product quality prediction is one of the key issues in smart manufacturing. Reliable evaluation and parameter optimization is needed to ensure their high-quality production outputs. This study develops a novel approach that integrates adaptive machine learning and nonlinear regression to accurately predict highly customized end product quality based on small sets of supply-chain data through digital transformation, especially for complex industrial machinery manufacturing. This study was conducted in collaboration with a major power transformer manufacturer and its supply chain partners. Using the supply chain’s key component real dataset, the qualities of end products are predicted using the adaptive model trained and validated with reliable accuracy. The power transformer key component parameter, i.e., core (or iron) loss, is incorporated as an input dataset for prediction model training and testing. The novel model integrates autoregressive integrated moving average (ARIMA) and adaptive boosting (AdaBoost) machine learning, called ARIMA-AdaBoost. Compared with the previous research, the experiment results have shown that the ARIMA-AdaBoost outperforms the simple AdaBoost and Long Short-Term Memory (LSTM)-AdaBoost for transformer quality predictions. The ARIMA-AdaBoost model outperforms the existing methodologies in terms of mean absolute percentage error (MAE) and root mean square error (RMSE) in real-data verification. The proposed approach benefits the manufacturers in overall production costs for predicting complex, expensive, and highly customized industrial product qualities and is adaptable to various industrial sectors.