Autoencoder-based fault diagnosis for grinding system

Abstract

At present, most fault diagnosis for grinding system is based on artificial judgments, which is inefficient, low accurate, high cost and easy to cause casualties. The traditional neural network has an unsatisfying performance to predict on high dimensional dataset, and is hard to extract crucial features, which brings about terrible classification results. To solve the above problems, the paper present a deep learning based on autoencoder to realize the intelligent diagnosis for grinding system. The algorithm applies autoencoder to extract features from fault dataset, and transit the non-linearized features to Softmax classification to recognize the fault category. This paper compares autoencoder-based deep learning networks and the traditional BP neural networks in experiments, and it is concluded that the autoencoder-based deep learning outperforms BP networks in the unbalanced classification. The classification precision is up to 92.4% by using the proposed method.