A novel feature extraction method based on discriminative graph regularized autoencoder for fault diagnosis

Abstract Autoencoder has been popularly used as an effective feature extraction method in fault diagnosis. However, the autoencoder algorithms neglect local structure and class information that is available in the training set. To address this problem, a novel feature extraction approach based on discriminative graph regularized autoencoder is proposed for fault diagnosis task. A single-layer autoencoder with nonlinear layers is adopted to extract nonlinear features automatically from input signals. Locality relationship of original data is propagated to the feature extraction stage via a graph to learn internal representations that go beyond reconstruction and on to locality preservation. To better exploit the discriminative information, the label information of training samples is embedded to the graph to improve the fault diagnosis performance. A real industrial process are used to comparing the performance with commonly used diagnosis method, the promising experimental results validate the superiority of the proposed method.