Towards holistic machinability estimation of titanium alloy: An integrated approach with enhanced feature extraction and physics-guided deep multi-task learning

초록

Titanium alloy is one of the most widely adopted materials in various industries due to its desirable mechanical properties. However, its thermal properties require special considerations to ensure quality and productivity, particularly from the perspective of machinability estimation, including tool wear, surface roughness, and cutting force. Not only is each factor of machinability difficult to estimate, but there also exists a highly complex underlying dynamics. However, existing approaches are limited to providing inefficient and partial estimations, failing to utilize physical knowledge shared across machinability factors. Hence, this work proposes a novel data-driven method that effectively estimates machinability in various aspects. First, a deep multitask learning (MTL)-based approach is proposed to estimate multiple factors of machinability simultaneously with a single predictive model. Second, a physics-guided encoder (PGE) is developed to enable efficient representation learning suitable for each machinability factor while preserving physical constraints. Third, a temporal-aware cross-task attention (TACTA) mechanism is developed to facilitate knowledge transfer between different machinability factors. Comprehensive experiments are conducted using real-world datasets collected during the milling processes of a titanium alloy under various machining conditions. The experimental results consistently demonstrate the effectiveness of the proposed method in estimating various factors of machinability using vibration input signals. In particular, the proposed method exhibits superior performance compared to conventional data-driven approaches as well as state-of-the-art approaches, indicating real-world practical applicability.

키워드