Diagnosis of rotating machine unbalance using machine learning algorithms on vibration orbital features

Abstract

The diagnosis of failures in rotating machines has been subject to studies because of its benefits to maintenance improvement. Condition monitoring reduces maintenance costs, increases reliability and availability, and extends the useful life of critical rotating machinery in industry ambiance. Machine learning techniques have been evolving rapidly, and its applications are bringing better performance to many fields. This study presents a new strategy to improve the diagnosis performance of rotating machines using machine learning strategies on vibration orbital features. The advantage of using orbits in comparison to other vibration measurement systems is the simplicity of the instrumentation involved as well as the information multiplicity contained in the orbit. On the other hand, rolling element bearings are prevalent in industrial machinery. This type of bearing has less orbital oscillation and is noisier than sliding contact bearings. Therefore, it is more difficult to extract useful information. Practical results on an industry motor workbench with rolling element bearings are presented, and the algorithm robustness is evaluated by calculating diagnosis accuracy using inputs with different signal-to-noise ratios. For this kind of noisy scenario where signal analysis is naturally tough, the algorithm classifies approximately 85% of the time correctly. In a completely harsh environment, where the signal-to-noise ratio can be smaller than −25 dB, the accuracy achieved is close to 60%. These statistics show that the strategy proposed can be robust for rotating machine unbalance condition diagnosis even in the worst scenarios, which is required for industrial applications.