Statistical Aware Transformer for Tool Wear Monitoring

Authors:
Open Access
Publisher website
Google Scholar
Abstract
Tool wear state identification under complex cutting conditions remains a challenging task regarding accuracy and robustness. Existing deep learning approaches often fail to fully exploit the complementary information between time-frequency images and one-dimensional statistical indicators. Furthermore, simple feature concatenation strategies are insufficient for achieving deep cross-modal fusion. In response to these challenges, this study presents, this paper proposes a novel multimodal monitoring framework named the Statistical Aware Transformer (SAT). Crucially, a statistical adaptive normalization mechanism is introduced. This module utilizes hand-crafted physical statistical features as conditional vectors to dynamically generate affine transformation parameters (scale and shift), thereby modulating the high-level feature extraction process of time-frequency images. By integrating statistical priors, this mechanism enables the model to adaptively attend to frequency bands critical for wear progression. Validated on the PHM2010 milling dataset, the proposed method demonstrates superior identification accuracy compared to traditional Convolutional Neural Networks (CNNs) and standard Vision Transformers (ViT) in cross-condition testing, confirming the efficacy of SAT in facilitating effective multimodal feature fusion.

This publication has 7 references indexed in Scilit: