Cascaded Machine Learning and Deep Learning Models for Reliable Predictive Maintenance in Industry 4.0

AI-based Predictive Maintenance (PdM) often shows strong results in laboratory settings but struggles when applied to real industrial environments. Many studies rely on benchmark datasets or simplified failure scenarios, which do not reflect the complexity of shop-floor conditions, where sensor data are heterogeneous, labels are sparse or noisy, and maintenance logs are unstructured. This work aims to address these gaps by developing and validating an end-to-end PdM pipeline trained exclusively on raw operational data from an industrial plant. The study provides three main contributions. First, it compares traditional machine learning models (XGBoost) with deep learning sequential models (LSTM), evaluating not only accuracy and F1 scores but also operational metrics such as lead time and alarm load. Second, it introduces a cascading pipeline: a lightweight first-stage model detects imminent failures, while subsequent models classify severity and fault type, reducing false alarms and focusing predictive power where needed. Third, it frames the modeling within a transparent and reproducible data-processing workflow, including cleaning, synchronization, target definition, and time-aware validation. Preliminary results show that the cascade approach achieves high performance in early fault detection (≈98% accuracy) and improves severity prediction, while component classification remains challenging due to class imbalance and noisy labels.