Fast fuzzy neural network for fault diagnosis of rotational machine parts using general parameter learning and adaptation

We compare empirically the performance of nonlinear radial basis function neural networks (RBFN) and time delay neural networks (TDNN) in accuracy and speed for fault detection in rotational machine parts. We use the advantageous general parameter (GP) approach for initializing the weights of the RBFN model in the beginning of the offline system identification phase, as well as for fine-tuning the modeling accuracy of RBFN. The GP-RBFN scheme is adaptive but still computationally efficient due to the single adaptive parameter and its simple learning rule. The fault measure is the moving average of a general parameter. In order to verify the performance of the proposed schemes, they are applied to fault detection of automobile transmission gears. As the acoustic time series is slightly nonlinear, the RBFN gives high-speed fault detection, but detection accuracy is not so high. To overcome this problem a TDNN is developed that achieves more accurate fault detection although it needs more computational time. A fault is detected through regression lines. Both methods are empirically compared in speed and accuracy for fault detection of automobile transmission gears.