New current analysis method for the diagnosis of gearbox faults under variable load and speed
Authors: Chaima Ben Abdallah; Mahfoud Bouzouidja; Abdenour Soualhi; Hubert Razik; Noureddine Zerhouni
Extended abstract:
Rotating machines like gear motors play a pivotal role in various industrial applications, ranging from manufacturing and transportation to energy generation. However, the reliable operation of these machines can be compromised by the occurrence of faults. The timely detection and diagnosis of these faults are essential for ensuring operational efficiency, preventing costly breakdowns, and ensuring workplace safety. The diagnosis of bearing and gear faults in these systems solely through non-intrusive current signal analysis is an interesting approach. This paper’s primary contribution lies in the extraction of a novel health indicator for identifying bearing outer race, inner and ball bearing faults, as well as broken and gear surface damage faults irrespective of speed and load variations. In fact, comprehensive monitoring systems encompassing variable speed and load regimes alongside combined gearbox faults solely using electrical signals are scarce in the existing literature. To address this gap, we proposed a pioneering method. The proposed method uses a regime normalization technique with different current sensors, allowing the grouping of various regimes under the same health state. By minimizing dispersion among class observations and distinguishing between different health states, including variations in speed and load, this method promises heightened diagnostic accuracy. This paper proposes also to improve the diagnosis of faults by using the Naïve Bayes classifier with the introduction of a criterion called the threshold in order to address the uncertainty. Addressing uncertainty is crucial to prevent false alarms during diagnosis. To verify the effectiveness of the proposed method, current data collected from a test bench composed of a gearbox system operating under time-dependent regimes is tested and validated. Also, the proposed method is compared with diverse machine learning classifiers to test the effectiveness of introducing the uncertainty to improve the diagnosis. The proposed model is performed based on accuracy of 97.55%, precision of 99.8% and F1-score of 98.73%.
