Performance comparison of various Decision-Based Neural Networks in gearbox fault diagnosis

Condition monitoring and fault diagnosis using intelligent systems has been of interest to researchers in recent decades. This paper studies the discrimination of faulty and healthy gears using artificial intelligence. Worn and chipped gears are considered for faulty data. Diagnosis process can be divided into three main phases: Empirical data collection, feature extraction and classification. Vibration signal obtained from the three accelerometers embedded in certain parts of the gearbox test setup, are inputs of the model. The experiments are done on a Samand 5-speed manual gearbox test setup in Acoustics Research Lab of Mechanical Engineering Department, Amirkabir University of Technology (Tehran Polytechnic). In feature extraction phase, statistical properties of wavelet coefficients are the most prominent characteristic being used. Performance of different artificial neural network configurations such as Linear Perceptron and hierarchical Decision Based Neural Networks (DBNN) is compared in the classification phase. The effect of Linear Basis Function (LBF) and Radial Basis Function (RBF) in different structures of neural network is also considered. The judgment is based on different point of views, such as rate of convergence, correct classification percentage and so on. It’s been shown that Subcluster structure has better classification capability than Linear Perceptron Networks.