MS-DenseNet-GRU tool wear prediction method based on attention mechanism

Abstract

Tool wear was an inevitable physical phenomenon in the cutting procedure. Serious tool wear has a direct effect on the level of processing quality and the effectiveness of production, and it even leads to abnormal cutting processes and a series of safety problems. Effective tool wear prediction can provide a basis for the rational use and replacement of tools to improve tool efficiency and ensure the stable operation of the machining process. Therefore, a tool wear prediction method combining multiple deep learning modules was proposed. To begin, the vibration signal was broken up using the complete ensemble empirical mode decomposition with adaptive noise algorithm. Then, the intrinsic mode functions with a strong correlation with the original signal were screened out according to the Pearson correlation coefficient for signal reconstruction. Additionally, the DenseNet module, the gate recurrent unit (GRU) module and the efficient channel attention module were deeply integrated to build a multi-scale DenseNet-GRU tool wear prediction model with attention mechanisms by learning the relationship of mapping between signal features and tool wear. Finally, the model was trained and tested using milling experimental data. The experiments’ outcomes demonstrated that the suggested method can accurately and reliably estimate the tool wear value. Compared with the DenseNet model, convolutional neural network–long short-term memory model, and DenseNet-GRU model, it further shows that it had superior performance in prediction accuracy and generalization ability. The research results can provide certain technical support for the prediction of tool wear intelligently, which is vital to raising the quality of processing, reducing production costs, and promoting the manufacturing industry’s intelligent development.