Fault Diagnosis Using Neural Networks for Parallel Shaft Gearboxes and Discussion on Its Generalization Ability

This paper presents a diagnosis method using back propagation neural networks (BPNNs) for parallel shaft gearboxes and discussion on its generalization ability. Feature vector, which is input of the BPNN, is extracted from the spectrum combining the data model. Normalization method is proposed based on the diagnosis mechanics. BPNN models are developed using the collecting data including two healthy and four faulty states of nine different speed and loading conditions. The performance of the trained models is evaluated using the traditional method by setting training and testing sets. At the same time, the generalization ability is evaluated using blind data set. The results reveal that the diagnosis accuracy is high in the training and testing sets. However, for the data of new working condition, the diagnosis accuracy will be obviously lower, which constrains the application in real practice. Some suggestions are made to improve the generalization ability of the data-driven diagnosis methods.