Intelligent monitoring system for production lines in smart factories: A hybrid method integrating Transformer and Kalman filter

Abstract

Intelligent monitoring systems for production lines in smart factories are used to ensure production efficiency, quality control and fault warning, promoting the optimization of production processes and resource allocation. Tool wear monitoring (TWM) bridges the gap between the perception of machining state information and accurate health management of tools. However, signal features undergo significant and abnormal changes during the late-stage of tool wear, posing substantial challenges to the development of an accurate tool wear intelligent monitoring model. In this paper, a hybrid TWM model integrating the Transformer and Kalman filter is proposed, with a state-space model of tool wear constructed and dynamically updated to address the gap in late-stage prediction accuracy of existing TWM methods. Specifically, the Transformer model is designed to describe the observation model of early tool wear states based on monitoring data. A system model is developed based on the actual tool wear mechanism to describe the relationship between the tool wear rate and tool-workpiece contact load over time. The Kalman filter is used to estimate the parameters of the mechanism model and track the evolution of wear. Within the Bayesian inference framework, measurement noise in the monitoring data is accounted for to optimize and update state estimation deviations and mechanism model parameters, enabling late-stage wear prediction through posterior estimation. The effectiveness and generalization of the proposed method are validated through milling experiments on both thin-walled and rectangular block parts. The experimental results indicate that the average RMSE error of tool wear prediction for thin-walled parts is 6.02, while for rectangular block parts, it is 4.70. The average RMSE errors of the proposed method are reduced by 16.34 % and 11.31 %, respectively, with respect to the single model. More importantly, the proposed method for TWM demonstrates strong predictive performance in the late-stage of tool wear while quantifying the uncertainty of wear prediction.