An evolving neuro-fuzzy classifier for fault diagnosis of gear systems

Gear systems (or gearboxes) are widely used in rotating machinery. Reliable gear fault diagnostic techniques and systems are critically needed to provide early warning of a possible defect so as to prevent machinery operation degradation and to reduce costs related to predictive maintenance. In this work, a new evolving neuro-fuzzy (eNF) classifier is proposed for real-time gear system fault diagnostics. In evolving process, the constraints related to gear health states are used to guide the partition of the output space, to prevent possible misleading clusters. A new training algorithm based on the normalized Adadelta function is suggested to improve eNF training convergence and accuracy. The effectiveness of the proposed eNF classifier is tested by simulation and experiment tests.