From Prototype to Production: Common ML Pitfalls and Solutions for Monitoring Infrastructure Assets
Year: 2025
Type: preprint
Abstract: Machine learning is rapidly moving from lab prototypes and synthetic environments to field-deployed systems, yet several of these initiatives stall when confronted with real-world data and operating conditions. We synthesize six recurring pitfalls that derail model performance in engineering contexts-(1) severe class imbalance, (2) noisy or incomplete sensor streams, (3) domain shift between laboratory and field, (4) high-dimensional inputs, (5) overreliance on synthetic or augmented data, and (6) temporal nonstationarity and concept drift-and pair each with validated remedies. Anchored by the latest published research, we outline why these failures arise through symptoms and root causes, and present evidence-backed solutions. Topics range from targeted preprocessing steps to post-training model evaluation strategies, providing a diverse and holistic approach to improving model performance and applicability to a production-ready level.
Author: Aynan Kareem
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4413515725