HYBRID CLASSIFICATION AND REGRESSION MODELS FOR PREDICTIVE MAINTENANCE IN SMART FACTORIES

Abstract

Recent studies report that unplanned machine failures account for nearly 20–30% of total production downtime in smart manufacturing environments. With Industry 4.0 enabling continuous sensor-based monitoring, industries now generate massive volumes of operational data, yet effective fault analysis and decision-making remain a major challenge. Conventional manual fault analysis systems suffer from critical limitations, including heavy dependence on human expertise and periodic inspections, which often result in delayed fault detection and inconsistent decision-making. In this study, a comprehensive data set is utilized containing parameters such as timestamp, machineid, temperature, vibrationlevel, powerconsumption, pressure, materialflowrate, cycletime, errorrate, downtime, maintenanceflag, efficiencyscore, and productionstatus. The data undergoes systematic Preprocessing, including noise handling, normalization, missing-value treatment, and feature alignment, followed by Exploratory Data Analysis (EDA) to understand operational patterns, fault correlations, and feature importance. Existing machine learning models such as AdaBoost-CART, XGBoost-CART, and Passive-Aggressive (PA)-CART are implemented as baseline methods. To overcome the limitations of fixed-feature learning and manual feature engineering, a proposed Neural Architecture Search (NAS)-based feature extraction framework integrated with a Greedy Rule Forest (GRF)-CART model is introduced. The NAS component automatically learns optimal feature representations from complex sensor interactions, while the GRF-CART enhances interpretability and decision robustness. The proposed framework performs classification tasks to predict downtime occurrence, maintenance requirement (maintenance flag), and production status, along with a regression task to accurately estimate the efficiency score. Experimental results demonstrate that the proposed NAS–GRF–CART approach significantly improves fault prediction accuracy, reduces false maintenance alerts, and provides reliable efficiency assessment, making it well-suited for intelligent, data-driven maintenance strategies in Industry 4.0 environments.