Analysis of feature extracting ability for cutting state monitoring using deep belief networks
Information extracting method from numerous measured signals is a critical technique for intelligent manufacturing application to further reduce the manpower cost and improve the productivity and workpiece quality. Manually defining signal features, as the common way, unfortunately will lose most of the information and the performance can't be guaranteed. In the past few years, machine learning method with deep structure has been the most promising automatic feature extracting method which has made great breakthrough in computer vision and automatic speech recognition. In this paper, deep belief networks are employed using vibration signal obtained from end milling to build feature space for cutting states monitoring. Greedy layer-wise strategy is adopted to pre-train the network and standard samples are used for fine-tuning by applying back-propagation method. Comparisons are made with several manually defined features both in time and frequency domain, like MFCC and wavelet method. Different modeling methods are also employed in the research for comparisons. Results show that the deep learning method has similar ability to characterize the signal for cutting states monitoring compared to those manually defined features. And the modeling accuracy is much better than other traditional modeling methods. Furthermore, benefitting from the potential capability in information fusion, deep learning method would be a promising solution for more complex applications, like tool wear monitoring, machining surface prediction et al.