A Hybrid Neural Networks based Machine Condition Forecaster and Classifier by using Multiple Vibration Parameters

The requirements of increasing the productivity and reducing the operation costs in modern factories has forced the managers to perform maintenance effectively. This may be indicated by the rapid adoption of predictive planned maintenance from conventional scheduled maintenance so that the time of shutting down the machines becomes minimum. In this paper, a hybrid neural networks approach based on the combination of Recurrent Back-Propagation (RBP) neural networks and fuzzy Adaptive Resonance Theory (ART) classifier is introduced. The main advantage of this approach is that the prediction is based on multiple parameters with on-line adaptation. To illustrate the effectiveness of this approach, the problem of predicting the operation conditions of a series of similar compressors has been studied. By extracting features from the vibration signals in both time and frequency domain, the RBP neural networks can be trained to forecast the future values of various vibration features. These future values will then input into the fuzzy ART neural networks in order to classify the operating conditions of the compressors and diagnose the existing of faults. Therefore, the operator will be acknowledged in advanced whether a maintenance is necessary. The results have proved that the proposed method can track the trend of features continuously and predict the conditions of machines in good accuracy.