Wear-Stage-Aware Multitask Learning and Optimization With a Causal Transformer for Milling Tool Wear Prediction Using Multisensor Data
Zengchao Su, Yue Wang
Shanxi Agricultural University
阅读操作
确认中在文库中上传 PDF 后可生成中文音频讲解。
摘要与影响
Accurate prediction of milling tool flank wear is of great significance for condition-based maintenance and quality assurance in intelligent manufacturing. Nevertheless, tool wear prediction faces several difficulties, including progressive and history-dependent wear evolution, the multistage wear mechanism in line with ISO 8688-1, and the generalization challenge across different tool instances with varying wear progression behaviors. To tackle these challenges, this article presents a hierarchical framework called StageFormer based on multitask learning and optimization. First, raw multisensor signals are split into fixed-length windows and processed by a segment-level 1-D convolutional neural network (1D-CNN) to extract local temporal features. Then, a learnable attention-based pooling mechanism is adopted to aggregate these features into compact cut-level representations. A causal Transformer encoder with explicit lower triangular masking is further employed to capture long-range temporal dependencies across sequential machining cuts while strictly avoiding future information leakage. In addition, a wear-stage-aware multitask prediction head is designed to simultaneously estimate the normalized flank wear value and its corresponding wear stage, where uncertainty-based loss weighting and adaptive optimization are used to realize adaptive task balancing. Extensive experiments are conducted on the PHM 2010 milling tool wear dataset under a leave-one-out evaluation protocol across the three labeled tool runs. The results illustrate that StageFormer outperforms 11 baseline methods, including both conventional machine learning models and state-of-the-art deep learning approaches. The proposed method achieves superior performance in all regression and classification metrics, with a remarkable reduction in root mean square error (RMSE) compared with the best baseline. Ablation studies verify the effectiveness of the key components, namely historical temporal modeling, causal masking, segment-level attention aggregation, and adaptive multitask optimization. Sensitivity analysis confirms that the proposed method achieves an optimal tradeoff between prediction accuracy and computational efficiency. The StageFormer framework provides a robust and effective solution for tool wear prediction across tool instances with diverse wear trajectories.
逐年被引趋势
暂无年度引用数据
关键指标
同类平均 = 1
同领域 · 同年份 · 同类型
Google Scholar 与 OpenAlex 的被引统计范围不同,数值存在差异属正常。
AI 辅助阅读
依据:摘要
可就本文提问;依据不足时会说明。
学术脉络
学科主题
工程Advanced machining processes and optimization
Machine Fault Diagnosis Techniques · Engineering Technology and Methodologies