A Hybrid Attention-Based Paralleled Deep Learning model for tool wear prediction

In modern manufacturing process, tool condition significantly affects work efficiency, machinery downtime and operating profit. Convolutional neural network (CNN), recurrent neural network (RNN) or other deep learning models are widely adopted to learn sensitive features individually or sequentially from enormous samples for tool status monitoring. However, these models only learn partial features due to their inherent structures. And features extraction performance of the model with simple sequential combination is also restricted by inner mutual block interference. In this paper, a novel deep learning network named Hybrid Attention-Based Parallel Deep Learning (HABPDL) model is proposed to address these problems. Specifically, ResNet and BiLSTM blocks individually learn features. Their corresponding attention layers, namely convolutional block attention module (CBAM) and general attention unit in BiLSTM, are stacked in sequence to highlight extracted features. And global average pooling (GAP) is applied to reduce superfluous spatial features and increase model interpretability after CBAM layer. Finally, these features maps from CNN and RNN parts are concatenated to predict tool wear value more accurately. Life cycle milling experiments are conducted, and vibration signals are acquired for model training and validation. After model hyperparameters optimization, comparison experiment results validate that the proposed model can learn more complete features without any inner interference, and own brilliant prediction performance due to well-designed parallel structure and block-attention units. Proposed HABPDL model achieves the best prediction results, and MAPE, MAE, RMSE and R2 reach 10.8%, 6.072, 7.955, and 0.933, respectively. The model also outperforms other models even under noisy environment.