An intelligent monitoring system of grinding wheel wear based on two-stage feature selection and Long Short-Term Memory network

Abstract

Grinding wheel condition is considered as the key factor affecting grinding performance, and therefore, accurate monitoring of wheel wear is necessary to prevent the deterioration of part quality. An intelligent wheel wear monitoring system is introduced in this article to realize processing of grinding signal, extraction of signal features, selection of optimal feature subset, and prediction of wheel wear. Physical information generated during the grinding of C-250 maraging steel is collected by a dynamometer, accelerometer, and acoustic emission sensor, and a large quantity of features in time domain and frequency domain are extracted from the processed grinding signals. To reduce feature redundancy and increase relevancy of feature to wheel wear, a two-stage feature selection approach combining filter and wrapper framework is proposed. The filter preselects individual features by minimum Redundancy Maximum Relevance method, while the wrapper evaluates different feature subsets by the model performance. A deep learning network structure named Long Short-Term Memory network is adopted to develop the wheel wear monitoring model and is compared with a conventional machine learning algorithm, Random Forest. The results have shown that the two-stage feature selection method is able to provide the globally optimal feature subset for the model. Long Short-Term Memory model achieves an R2 of 0.994 and a root-mean-square error of 0.240 with four features, while Random Forest model obtains an R2 of 0.980 and a root-mean-square error of 0.463 with seven features, which indicates that Long Short-Term Memory model is capable of predicting wheel wear more accurately even with less features.