Condition-Adaptive 1DCNN-Transformer for Multi-Condition Tool Wear Prediction

Abstract Accurate real-time prediction of tool wear under multi-condition operations is critical for ensuring machining precision and industrial productivity. This study proposes an end-to-end tool wear prediction framework integrating a 1D Convolutional Neural Network (1DCNN) and Transformer architecture, enhanced by dynamic parameter adaptation and transfer learning strategies. The core innovation lies in two aspects: 1) A novel Conditional Layer Normalization (CLN) mechanism that dynamically adjusts normalization parameters based on operational conditions (cutting depth, feed rate, material), enabling effective feature alignment across diverse scenarios; 2) A hybrid 1DCNN-Transformer structure that synergizes local temporal feature extraction through dilated convolutions with global dependency modeling via multi-head attention. Transfer learning is systematically implemented by freezing backbone parameters while fine-tuning task-specific layers, ensuring efficient knowledge transfer from single-condition pre-trained models to multi-condition applications. The model was rigorously validated through two complementary approaches: NASA milling dataset analysis (16 operational conditions) and titanium alloy TC17 machining experiments. Quantitative evaluations demonstrated superior performance, achieving R² scores of 0.9069, 0.9058and 0.9105 in material-variation, cutting-depth variation and cutting-speed adaptation tasks, respectively, with RMSE consistently below 0.0936. These results confirm the framework's capability to establish robust signal-to-wear mapping relationships under heterogeneous conditions. The proposed condition-aware architecture provides a scalable solution for industrial tool health monitoring systems.