A heterogeneous-structure-driven time series forecasting model for tool wear prediction in Titanium alloy milling

Abstract

In practical titanium alloy milling processes, variations in machining conditions—such as cutting parameters, tool materials, and environmental factors—induce significant shifts in both the marginal and conditional distributions of tool wear data. This distribution drift poses substantial challenges for accurate tool condition monitoring and stable wear prediction. Achieving high-precision, multi-step, and cross-condition tool wear forecasting with limited data remains a critical issue in intelligent manufacturing and predictive maintenance. To address this, we propose a heterogeneous-structure-driven time series forecasting model, featuring a dual-stream architecture based on trend-seasonal decoupling and synergistic learning. The framework begins with preprocessing operations including data interpolation, windowing, slicing, and trend-seasonal decomposition, effectively enhancing data diversity and temporal representation. Experimental results demonstrate that the proposed model consistently achieves superior performance across diverse operating conditions and time horizons. Notably, in long-horizon cross-domain forecasting tasks, the model maintains high predictive accuracy, with the worst-case RMSE, MAE, and MAPE reaching 0.0057 mm, 0.021 mm, and 1.8%, respectively. This study presents an effective, robust, and generalizable solution for tool wear prediction under highly dynamic machining environments.