Energy-Efficient Neural Architecture Search for Joint Production Status and Predictive Maintenance Classification

Abstract

Unexpected machine breakdowns are a major contributor to productivity loss in modern manufacturing, accounting for nearly 20–30% of total downtime in smart industrial environments. Although Industry 4.0 enables continuous monitoring through interconnected sensors, transforming large volumes of operational data into actionable insights for fault detection remains challenging. Conventional fault analysis methods rely heavily on manual inspections and expert intervention, often resulting in delayed fault identification and inconsistent maintenance decisions. In this study, a comprehensive dataset is utilized, consisting of parameters such as timestamp, machine ID, temperature, vibration intensity, power consumption, pressure, material flow rate, cycle time, error rate, downtime, maintenance flag, efficiency score, and production status. The data undergoes systematic preprocessing, including noise reduction, normalization, missing-value handling, and feature alignment, followed by Exploratory Data Analysis to uncover operational patterns, feature relationships, and key fault indicators. Baseline models such as Adaptive Boosting (AdaBoost)- Classification and Regression Trees (CART), Extreme Gradient Boosting (XGBoost)-CART, and Passive-Aggressive (PA)-CART are implemented for performance comparison. To address the limitations of manual feature engineering, a hybrid framework combining Neural Architecture Search (NAS) with Greedy Rule Forest (GRF)-CART is proposed, where NAS automatically learns optimal feature representations from complex sensor interactions and GRF-CART enhances interpretability and decision robustness. The framework performs classification tasks to predict downtime occurrence, maintenance requirement, and production status, along with regressionbased estimation of efficiency score, and experimental results demonstrate improved prediction accuracy, reduced false maintenance alerts, and reliable efficiency assessment.