Predictive Maintenance in SMT Machines Using Electrical Multiparameter Sensors and Hybrid Machine Learning Models

Surface-Mount Technology (SMT) manufacturing demands high operational reliability, as unexpected failures lead to costly production downtime. Predictive maintenance based on continuous sensor monitoring has emerged as a promising approach to anticipate failures; however, its effectiveness depends on robust and stable anomaly detection systems. In this industrial context, anomaly detection faces specific challenges, including the absence of labeled data for supervised training, the presence of multiple operational regimes with distinct electrical characteristics, and the need for temporal stability in alarm generation. Traditional global methods such as Isolation Forest, One-Class SVM (OCSVM), and Local Outlier Factor (LOF) assume a single operational context, failing to adapt detection behavior across different regimes and often producing unstable alarms that undermine system reliability.

This work proposes a hybrid, context-aware approach that combines K-Means clustering for the automatic segmentation of operational regimes with cluster-specific Isolation Forest models for anomaly detection. The method was validated using 47,493 samples of electrical sensor data (15 variables) collected over three months from an SMT insertion machine operating in a real production environment. The proposed approach was compared against three baselines: global Isolation Forest, K-Means with OCSVM, and K-Means with LOF. Performance was evaluated in terms of regime separability (Silhouette score), temporal stability (coefficient of variation), and anomaly score consistency (interquartile range).

The hybrid method identified three distinct operational contexts, achieving 53% higher separability than the global approach (Silhouette 0.63 vs. 0.41) and 37% greater temporal stability compared to global Isolation Forest (CV 0.96 vs. 1.52), effectively reducing erratic alarm peaks. Score consistency was substantially higher than OCSVM-based clustering (IQR 0.10 vs. 5.31), while maintaining an equivalent detection rate of 1.0%, confirming that performance gains arise from contextual adaptation rather than sensitivity increase.

These results demonstrate that adapting anomaly detection to operational regimes provides a methodological advantage for industrial predictive maintenance, balancing temporal stability and sensitivity in continuous monitoring.