Application of new training method combined with feedforward artificial neural network for rolling bearing fault diagnosis

A new technique for the training of ANNs is presented. The time-domain vibration signals of rolling bearings with different fault conditions are preprocessed using differential evolution method, then further being trained by Levenberg Marquardt method. The processed data are applied as input vectors to artificial neural networks (ANNs) for rolling bearing fault classification. The hybrid training method overcomes the defects of network training, for example lower convergence speed of evolutionary artificial neural network and easiness of falling into local minimum. And it also has the advantages of quick convergence speed and good global continuous optimization ability. In addition, probabilistic adaptive strategy which could save computation time in various situations is adopted. The proposed method is applied to the rolling bearings faults diagnosis, and compared with other training methods. The results for both real and simulated bearing vibration data show that, high correct classification rate were obtained through LM, and the presented method demonstrated rapid convergence and good stability than traditional method such as LM and other methods. The probabilistic adaptive strategy improved the convergence rate and obtained higher correct rate.

I. Introduction

The training methods of feedforward network include gradient decent method (GD), conjugate gradient method (CG), Levenberg-Marquardt method (LM) etc. And some improved methods based on traditional methods are put forward to avoid slow convergence and being trapped in local minima. Recently, the evolutionary artificial neural network (EANN) with heuristic training method attracts the interest of researchers. However, in experiments of solving complicated problems, the convergence of the new method is not as good as some of the traditional methods [1].