Multi-Category Classification of Tool Conditions Using Wavelet Packets and ART2 Network

Abstract

This paper proposes a new approach for multi-category identification of turning tool conditions. It uses the time-frequency feature information of the AE signal obtained from best-basis wavelet packet analysis. By applying the philosophy of divide-and-conquer and a local wavelet packet extraction technique, acoustic emission (AE) signals from turning process have been separated into transient and continuous components. The transient and continuous AE components are used respectively for transient tool conditions and tool wear identification. For transient tool condition identification, a 16-element feature vector derived from the frequency band value of wavelet packet coefficients in the time-frequency phase plane is used to identify tool fracture, chipping and chip breakage through an ART2 network. To identify tool wear status, spectral and statistical analysis techniques have been employed to extract three primary features: the frequency band power at 300 kHz–600 kHz, the skew and kurtosis. The mean and standard deviation within a moving window of the primary features are then computed to give three secondary features. The six features form the inputs to an ART2 neural network to identify fresh and worn state of the tool. Cutting experimental results have shown that this approach is highly successful in identifying both the transient and progressive tool wear states over a wide range of turning conditions.