Predictive maintenance to avoid fatigue and failure enhances the reliability of mechanics, herewith, this paper explores vibrational time-domain data in advancing fault diagnosis of predictive maintenance. This study leveraged a belt-drive system with the properties: operating rotational speeds of 500–2000 RPM, belt pretensions at 70 and 150 N, and three operational cases of healthy, faulty and unbalanced, which leads to 12 studied cases. In this analysis, two one-axis piezoelectric accelerometers were utilized to capture vibration signals near the driver and pulley. Five advanced statistics were calculated during signal processing, namely Variance, Mean Absolute Deviation (MAD), Zero Crossing Rate (ZCR), Autocorrelation Coefficient, and the signal's Energy. The Taguchi method was used to test the five selected features on the basis of Signal-to-Noise (S/N) ratio. For classifications, an expert system was used based on artificial intelligence where a Random Forest (RF) model was trained on untraditional parameters for optimizing the accuracy. The resulted 0.990 and 0.999, accuracy and AUC, demonstrate the RF model's high dependability. Evidently, the methodology highlights the features potential when progressed into expert systems, which advances predictive maintenance strategies for belt-drive systems.
Read full abstract