Local-feature and global-dependency based tool wear prediction using deep learning

Evaluation of tool wear is vital in manufacturing system, since early detections on worn-out condition can ensure workpiece quality, improve machining efficiency. With the development of intelligent manufacturing, tool wear prediction technology plays an increasingly important role. However, traditional tool wear prediction methods rely on experience and knowledge of experts and are labor-extensive. Deep learning provides an effective way to extract features of raw data and establish the mapping relationship between features and targets automatically. In this paper, a new local-feature and global-dependency based tool wear prediction method is proposed. It is a hybrid approach combining manual features with automatic features. Firstly, an enhanced CNN network is designed and applied on the transformed wavelet scalogram to learn the local single-scale specific features and multi-scale correlation features automatically. Secondly, sequence of local feature vectors combining manual features with automatic features are fed into multi-layer LSTM step by step for the global dependency. A fully connected layer is then trained to predict tool wear. Finally, two statistics are proposed to illustrate the overall prediction performance and generalization ability of the model. An experiment illustrates the effectiveness of our proposed method under multiple working conditions.