An intelligent grinding burn detection system based on two-stage feature selection and stacked sparse autoencoder

Abstract

Grinding burn is a typical quality defect that negatively affects microstructures and mechanical properties of parts. In this article, an intelligent grinding burn detection system that consists of processing of grinding signals, extraction of signal features, selection of optimal feature subset, and classification of grinding burn is presented. During the grinding of maraging steel 3J33, physical signals are collected by a dynamometer, accelerometer, and acoustic emission sensor. A large amount of preliminary features in time domain and frequency domain are extracted after wavelet packet decomposition and ensemble empirical mode decomposition. To reduce the dimensionality of feature space and increase relevancy of feature to grinding burn, a two-stage feature selection strategy combining filter and wrapper method is proposed. The filter ranks individual features by ReliefF, while the wrapper evaluates different feature subsets according to the model performance. A deep learning network, stacked sparse autoencoder (SSAE), is adopted to develop the classification model of grinding burn and is compared with other machine learning models. The results demonstrate that the two-stage feature selection method is able to provide the optimal feature subset for the model and SSAE model obtains the best training and testing classification rate of grinding burn with least features among all comparative models.