Fault detection in rotatory machines is an important necessity to provide system operational reliability and lower unexpected downtime in industrial systems. Conventional feature extraction methods, although good in controlled situations, tend to fall short in noisy situations that mirror real-world operation situations. To fulfill this gap, this work explores machine learning classifiers-Random Forest (RF), Gradient Boosting (GB), XGBoost (XGB), and LightGBM (LGBM)-with two different feature sets: statistical features and Recurrence Quantification Analysis (RQA) features. First, we systematically validate the proposed framework by benchmarking recurrence-based dynamical features against conventional statistical features on a synthetic nonlinear benchmark composed of four well-established dynamical systems: the Lorenz system, the Rössler system, the Hénon map, and the Duffing oscillator. This controlled evaluation enables a methodological comparison across distinct nonlinear regimes before applying the framework to industrial fault data. Then, using experimental rotary machines data, we further examine the robustness of the classifiers under two different operation situations: noise-free and noise-included situations. In a noise-free situation, all classifiers resulted near perfect classification with all statistical features obtaining 99%, whereas RQA features resulted similarly around 93%. Yet, when a Gaussian noise is added (15%-75% of signal magnitude), statistical features marked a considerable decline to 85% accuracy, whereas RQA features remained highly efficient above 94%; consistent results were obtained when other types of noise were added, namely, Brownian noise and impulsive noise. These results present a demonstration of RQA-based feature superiority in long-term dynamics capture and cite its potential to provide a more reliable fault detection in industrial practice where the use of noise and variability of environments is an inescapable consequence.
Zaitouny et al. (Wed,) studied this question.