IMPROVING PERFORMANCE OF AN ARTIFICIAL NEURAL NETWORK BASED GEARBOX FAULT DIAGNOSIS SYSTEM

The automatic vibration monitoring methods of gears and gearboxes due to their extensive applications in industry are improving. Hence, their vibration signal and its derived features, has been an interesting topic for researchers in this field. On the other hand, optimizing the number of vibration signal features used in the detection and diagnosis process is crucial for increasing the fault detection speed of automatic condition monitoring systems. In this paper, a system based on multiple layer perceptron artificial neural networks (MLP ANNs) is designed to diagnose different types of fault in a gearbox. Using a feature selection method, the system is optimized through eliminating unimportant features of vibration signals. This method is based on a simple and fast sensitivity evaluation process, which results in a considerable estimation, despite its simplicity. Consequently, the system's speed increases, while the classification error decreases or remains constant in some other cases. An experimental test rig data set is used to verify the effectiveness and accuracy of the mentioned method. Four different types of data which are generated through the test rig setup are: no fault condition, 5% fault (5% eroded tooth) gear, 20% eroded tooth gear and the broken tooth gear. The results verify that eliminating some input features of gear vibration signal, using this method, will increase the accuracy and detection speed of gear fault diagnosis methods. The improved systems with fewer input features and higher precision, facilitates the development of online automatic condition monitoring systems.