Abstract: Milling tool wear condition affects real production. Thus, study on tool condition monitoring has great importance in engineering. Deep learning models, for example convolutional neural network, have been applied in tool condition monitoring during milling process. Yet the model interpretability is poor, and the prediction results vary a lot. As a novel variant of convolutional neural network, principal component analysis network (PCANet) model is well-explained. However, the self-supervised features extraction capacity still requires improvement, and few industrial cases have been studied. In order to address these problems, original PCANet model structure is optimized, then activated PCANet with max pooling and support vector regression (APCANet-MP-SVR) model is proposed to extract sensitive features in unsupervised way and predict tool wear accurately. In detail, tanh activation function is applied to improve model generalization capacity, and then max pooling layer is introduced for features selection to replace complex Hash encoding and spectrum process. In the end, support vector regression is utilized to predict current tool wear. Case has been further studied to validate the brilliant performance and industrial suitability of the proposed model.

Key words: tool wear, deep learning, PCANet, activation function, pooling layer