基于可解释物理引导空间注意力改进的跨工艺参数立铣刀磨损辨识

Improved Interpretable-based Physically Guided Spatial Attention for Cross-process Parameters End Milling Cutter Wear Identification

深度学习的可解释性不足已成为制约其工业应用的重要问题。刀具磨损状态智能评估方法有着较高的速度、自动化和智能化程度,然而其端到端的模式及提取的特征难以被理解。尤其在跨工艺参数时,其可解释性差、可靠性不足。为此提出一种基于可解释物理引导空间注意力机制改进的跨工艺参数立铣刀磨损辨识方法,首先根据信号的周期不连续特性构建了一种物理引导的空间注意力模块,实现了在跨工艺参数(进给速度、切深)下关键信号片段的自适应捕捉;其次结合最大均值差异及方差约束对特征进行约束,缩小了不同工艺参数下磨损特征的分布差异,同时提升了同类磨损特征的聚集程度。通过实际的刀具切削试验验证,所提方法有效提升了跨工艺参数下的辨识精度;空间注意力权重分布的可视化分析结果表明,所提方法提取出的特征具有明显的物理可解释性。所提方法可为刀具磨损监测的深度辨识模型进行可解释性优化提供参考。

The insufficient interpretability of deep learning has become a critical issue restraining its industrial applications.Intelligent assessment methods for tool wear state exhibit high levels of speed,automation,and intelligence;however,the end-to-end patterns and extracted features are challenging to be understood.Especially when dealing with cross-process parameters,their interpretability is poor,and their reliability is insufficient.Cross-process parameters end milling cutter wear state identification model is proposed to address this based on an improved interpretable-based physically guided spatial attention mechanism.Firstly,a physically guided spatial attention module is constructed based on the periodic discontinuity characteristics of the signals,enabling the adaptive capture of key signal fragments under cross-process parameters such as feed rate and cutting depth.Secondly,constraints are applied to the features using maximum mean discrepancy and variance,reducing the distribution discrepancies of wear features under different process parameters while enhancing the aggregation of similar wear features.Through practical cutting experiments on the tool,the effectiveness of the proposed method in improving the identification accuracy under cross-process parameters is validated.Visual analysis of the spatial attention weight distribution indicates that the proposed method's extracted features possess obvious physical interpretability.The proposed method can provide a reference for the interpretability optimization of the deep learning-based identification model of tool wear monitoring.