A method for predicting spindle rotation accuracy using vibration

Authors:

Abstract
When investigating motorized spindle reliability, it is very difficult to monitor spindle rotation accuracy in real-time. To address this, a method for predicting spindle rotation accuracy using the vibration signal from the front bearing is proposed. Waveform prediction based on a back propagation (BP) neural network is used for boundary extension to suppress local mean decomposition (LMD) end effects; the vibration signal from the front bearing is decomposed using the modified LMD, and sample entropy extracted from the PFs is used to build the training-prediction model, with spindle rotation accuracy based on long short-term memory (LSTM). An experiment based on the spindle’s load spectrum was performed to acquire data and validate the prediction model. The results show that waveform prediction based on a BP neural network can effectively suppress end effects. The mean error of the extracted sample entropy was just 2.28%, and the accuracy of the spindle rotation prediction was 93.15%. Results from training the LSTM show that the rotation accuracy of the spindle can be accurately predicted based on the vibration signal from the spindle front bearing.