An application of neural network for selecting feature parameters in machinery diagnosis

When we want to monitor the machinery componentʹs condition and diagnose its failure, first of all, suitable feature parameters should be selected to sufficiently describe the vibration of machinery components. If we choose too many parameters, it will take wasteful time and cost to analyze the machine conditions, and if we choose too little ones (we will miss out key parameters), we will fail to diagnose the machine accurately. But up to now, selection methods of feature parameters depend heavily on empirical rules and experimental trials. In this paper, we propose a new selection method of feature parameters using a learned neural network. In our method, importance factor of feature parameters is measured by sensitivities from inputs to outputs in the learned neural network. We propose some useful algorithms of parameters selection by these sensitivities. And we also show the practicability of this method on sample data and real data. Using this method, we can eliminate dispensable parameters with accuracy and we can reduce number of sensors and lighten the load of real-time processing. So, we think that this method will also contribute to real-time observation of machinery.