Search

Thin-walled parts with low stiffness and poor machinability are prone to chatter during milling. Chatter can adversely affect machined surface quality and workpiece performance. Most existing chatter monitoring models require training and test data from workpieces with the same stiffness as well as large amounts of training data. In this paper, force signals from workpieces with different stiffnesses are obtained through a series of experiments as source and target domain data. Autoencoders are utilized for feature extraction and support vector machines for chatter recognition. Model-agnostic meta-learning (MAML) optimizes the global model parameters, and transfer learning reduces the demand for target domain samples. The model can monitor the chatter of two types of workpieces at the same time through one training. Compared with common transfer learning models visual geometry group network-16 (VGG-16), this paper decomposes force signals using ensemble empirical mode decomposition (EEMD) and extracts intrinsic mode function (IMF) components as features. VGG-16 model is optimized by adding fully-connected layers, dropout, L2 regularization for chatter recognition. The proposed method achieves 99% accuracy on source domain and 98% on target domain. The target domain accuracy is 6% higher than without using MAML optimization, and 8% higher than the VGG-16 method. [ABSTRACT FROM AUTHOR]