Noise-driven challenges in feature selection strategies for PdM systems

Paweł Knap, Urszula Jachymczyk

Abstract: In industrial environments, vibration signals and other measured quantities are often affected by complex, overlapping sources of noise and disturbance. Such conditions can degrade the reliability of machine learning (ML) models that are based on these signals for predictive maintenance (PdM) or fault detection. In this study, we investigate the robustness of time-and frequency-domain scalar indicators when exposed to varying noise levels. We apply feature selection algorithms, including Random Forest-based methods and L1 regularization, to identify, whether there is a subset of features that remain informative despite significant noise contamination. To simulate industrial conditions, we generate multiple noisy datasets by systematically introducing controlled disturbances. The aim of this work is to compare effectiveness of PdM systems in noisy enviroment, when operating on contition indicators versus raw time series data. The effectiveness of ML models, is analyzed under both clean and noisy conditions, using accuracy score and statistical Friedman Test. This rigorous evaluation framework allows us to objectively assess whether observed performance differences are statistically meaningful or merely due to chance. By highlighting the challenges of feature selection in noisy environments, our findings aim to guide the design of more robust, noise-tolerant strategies for PdM systems in real-world industrial environments.

External IDs:doi:10.1109/iccc65605.2025.11022802