Search

• Tool structure parameters, process parameters, etc. are introduced to the model structure design. • Frequency and spatial components can be identified and weighted adaptively by designed attention modules. • The contribution of each attention module of the model is analyzed in detail through ablation experiments. • By visualizing and comparing the attention weights and features, the interpretability is explored. Real-time and accurate monitoring of tool wear status is critical for optimizing product quality and production costs. Deep learning has received much attention as an emerging data-driven technique in the field of tool wear monitoring because of its powerful feature extraction and nonlinear mapping capabilities. However, the model structure and features' poor interpretability limits the application of deep learning in practical machining. Herein, a frequency-spatial hybrid attention network is proposed, driven by tool structure and process parameters (FSHAN-SPD). The structure of FSHAN-SPD considers the recognition of critical frequencies and spatial positions simultaneously, and the receptive field of its spatial attention module can effectively distinguish the cutting stage of each tooth. Attention modules-weighted features are used for interpretability analysis. The effectiveness and flexibility of this method are verified in the PHM2010 dataset. A milling experiment illustrates a more detailed procedure. The experimental results are well interpretable in both frequency and spatial. [ABSTRACT FROM AUTHOR]

• An autoencoder model based on a multi-domain adversarial network was proposed. • The multi-domain adversarial network is developed to improve feature transfer. • A bearing life prediction network based on multi-channel features is proposed. • A degradation occurrence time identification method was proposed. Predicting the remaining useful life (RUL) of rolling bearings under different working conditions improved significantly by transfer learning. However, existing methods have not studied the following problems thoroughly: (1) The influence of the discrepancy between features of different dimensions on the feature transfer process; (2) The feature transfer process in the degradation stage with apparent discrepancy has a significant influence on the transfer prediction of remaining useful life. In this study, a degradation occurrence time identification method based on the distribution differences in reconstructing degradation indicators has been proposed to obtain samples of degradation stages. A stack convolutional autoencoder model based on a multi-domain adversarial network is also proposed to reduce the impact of discrepancies among extracted degradation features on the feature transfer process. As per the experimental results, it was found that the proposed method can effectively improve the RUL prediction accuracy. [ABSTRACT FROM AUTHOR]