Intelligent tool wear prediction based on Informer encoder and stacked bidirectional gated recurrent unit

Abstract

Tool wear prediction is critical to the safety of the machining environment and the reliability of machining quality. To achieve accurate prediction, we propose a novel deep learning-based model with the backbone of the Informer encoder and a stacked bidirectional gated recurrent unit (IE-SBiGRU) network. A long temporal feature sequence with abundant information is initially derived from multichannel sensor signals. Compared to convolutional and recurrent neural network based approaches, the proposed model realizes a global inceptive field of long input sequence without multi-layer stacks and parallel computing of long-range dependencies, respectively. In comparison with the Transformer encoder, the Informer encoder reduces the attention matrix sparsity of the global temporal features, and the dominant feature map is selected with increased computing efficiency. A stacked bidirectional gated recurrent unit network (SBiGRU) is then employed to strengthen the local correlation of the halving cascading features sequence. Finally, the global max pooling (GMP) layer is used to substantially lower the fully connected layer's dimension, and the predicted tool wear values are generated. The PHM2010 benchmark dataset and local experimental dataset are used to demonstrate the high accuracy of the proposed model. Ablation studies are also conducted to indicate the effectiveness of model modules and recommended hyperparameters.