Abstract: Due to limitations inherent in offline tool wear measurement methods, the availability of wear samples was restricted, and measurement noise was often unavoidable, which complicated the reliability of tool wear monitoring. To address these challenges, a soft sensor modeling and uncertainty analysis approach of tool wear was proposed based on semi-supervised Bayesian Transformer by integrating a semi-supervised Transformer model, Dropout and Monte Carlo(MC) simulation methods. Firstly, a soft sensor model was constructed based on the semi-supervised Transformer network architecture, the network training methods of unsupervised feature extraction and supervised fine-tuning were used to guide the construction of the tool wear soft sensor model under small samples. Then, in order to quantify the impacts of noise on tool wear, a noise network channel was designed for uncertainty analysis. Finally, using approximate Bayesian computation based MC-Dropout, the random uncertainty caused by noise and the cognitive uncertainty resulting from model modeling errors were quantified, aiming at providing more comprehensive information for tool wear assessment. The results show that the proposed soft sensor model and the uncertainty analysis framework may provide a powerful tool for tool health management.

Key words: tool wear, deep learning, wear prediction, Monte Carlo（MC） simulation

摘要： 受刀具磨损离线测量方式的限制，磨损样本少，且测量噪声难以避免，导致难以实现可靠的刀具磨损监测。针对该问题，融合半监督Transformer模型、Dropout、蒙特卡罗(MC)模拟方法，提出了基于半监督贝叶斯Transformer的刀具磨损软测量及其不确定性分析方法。首先，构建基于半监督Transformer网络架构的软测量模型，利用无监督特征提取和有监督微调的网络训练方式，指导小样本下的刀具磨损软测量模型构建；然后，为量化噪声对刀具磨损的影响，设计面向不确定性分析的噪声网络通道；最后，结合MC-Dropout近似贝叶斯过程，对噪声引起的随机不确定性和建模误差引起的认知不确定性进行量化，为刀具磨损评估提供更加全面的信息。研究结果表明，所提出的刀具磨损软测量模型及其不确定性分析框架能够为刀具健康管理提供有力工具。

关键词: 刀具磨损, 深度学习, 磨损预测, 蒙特卡罗模拟