Predictive Maintenance Procedure for Industrial Equipment Using Artificial Intelligence

This paper proposes a comprehensive, artificial intelligence (AI)-driven predictive-maintenance procedure for critical industrial assets—exemplified here on power transformers—that unifies best practices in time-series anomaly detection and supervised classification. Our pipeline begins with multi-sensor data collection and rigorous preprocessing (outlier removal, interpolation, normalization), then segments the stream into overlapping windows for feature engineering (time- and frequency-domain statistics plus cross-signal correlations). Two complementary models are trained on these windows: a Long Short-Term Memory (LSTM) autoencoder on “normal” sequences to produce an anomaly score (reconstruction error), and a Random Forest classifier on labeled windows to yield a “normality” probability. We fuse these heterogeneous outputs into a single Health Index via a convex combination whose weights are automatically optimized to maximize Receiver Operating Characteristic – Area Under the Curve (ROC-AUC). In production, the system ingests new windows in real time, computes the Health Index, and triggers alerts when the index crosses a threshold. A case study on transformer data demonstrates early fault detection (first alert 3 windows before failure), perfect discrimination (ROC-AUC=0.94), and practical lead-time for maintenance planning.