Automatic feature constructing from vibration signals for machining state monitoring

Machining state monitoring is an important subject for intelligent manufacturing. Feature construction is accepted to be the most critical procedure for a signal-based monitoring system and has attracted a lot of research interest. The traditional manual constructing way is skill intensive and the performance cannot be guaranteed. This paper presented an automatic feature construction method which can reveal the inherent relationship between the input vibration signals and the output machining states, including idling moving, stable cutting and chatter, using a reasonable and mathematical way. Firstly a large signal set is carefully prepared by a series of machining experiments followed by some necessary preprocessing. And then, a deep belief network is trained on the signal set to automatically construct features using the two step training procedure, namely unsupervised greedily layer-wise pertaining and supervised fine-tuning. The automatically extracted features can exactly reveal the connection between the vibration signal and the machining states. Using the automatic extracted features, even a linear classifier can easily achieve nearly 100% modeling accuracy and wonderful generalization performance, besides good repeatability precision on a large well prepared signal set. For the actual online application, voting strategy is introduced to smooth the predicted states and make the final state identification to ensure the detection reliability by taking consideration of the machining history. Experiments proved the proposed method to be efficient in protecting the workpiece from serious chatter damage.