An accurate prediction method of multiple deterioration forms of tool based on multitask learning with low rank tensor constraint. (January 2021)

• Record Type: Journal Article
• Authors: Liu, Changqing; Ni, Jincheng; Wan, Peng
• Is Part Of: Journal of manufacturing systems. Volume 58(2021)Part A
• Page Start: 193
• Page End: 204
• Publication Date: 2021-01
• Subjects: Tool condition prediction -- Multiple deterioration forms -- Multitask learning -- Low rank tensor constraint; Manufacturing processes -- Periodicals; Production engineering -- Data processing -- Periodicals; Robots, Industrial -- Periodicals; Production, Technique de la -- Informatique -- Périodiques; Robots industriels -- Périodiques; Electronic journals; 670.42
• DOI: 10.1016/j.jmsy.2020.11.018
• Languages: English
• ISSNs: 0278-6125

Abstract

Highlights: Prediction of multiform tool deterioration based on multitask learning is proposed. The strategy of low rank tensor constraint is used for multitask learning model. A 3-order tensor is constructed by integrating the weights of base neural networks. The hidden relationship of the multiple tasks is learned by the multitask learning. Both the prediction accuracy of tool wear and chipping is significant improved. Abstract: Tool deterioration is a common issue in Numerical Control (NC) machining, which directly affects part quality, production efficiency and manufacturing cost. Due to the complexity of machining, multiple deterioration forms of tool are involved during the tool deterioration process, which imposes a significant challenge for tool condition prediction because of the coupling effects among different deterioration forms. In order to address this issue, an accurate prediction method of multiple deterioration forms of tool based on multitask learning with low rank tensor constraint is proposed in this paper. A base model for prediction of each deterioration form is firstly constructed by using neural network, and then a 3-order tensor is constructed by stacking the weight matrices of the hidden layers of the base neural networks. The hidden relationship among the multiple related tasks is expected to be learned by means of a mathematical approach, i.e., constraining the 3-order tensor with low rank, which is realized by joint training the base prediction models. The experimental results show that the proposed method can reduce the prediction error by about 13 % for wear and 25 % for chipping respectively compared with the corresponding single task models.