Multi-Factor Prediction of Cumulative Deposited Height in CMT Wire Arc Additive Manufacturing Using an Improved Temporal Convolutional Network

Cold Metal Transfer wire arc additive manufacturing (CMT-WAAM) efficiently fabricates medium and large metallic components, yet deposited wall cumulative height varies spatially along the build direction and deposition path due to heat accumulation, inter-layer cooling, travel direction and inherited layer geometry. This study establishes a multi-factor prediction framework for single-bead multi-layer walls and evaluates CAFi-TCN, a temporal convolutional network enhanced with feature-wise linear modulation and causal attention. Height profiles were extracted from registered point clouds under 2–4 mininter-layerr cooling; the model uses deposition position, layer number, cooling time, travel direction, prior height increment and cumulative height to predict current-layer cumulative height. On the tenth-layer test set, CAFi-TCN achieved the lowest mean absolute error (MAE = 0.1836 mm) among the evaluated direct-height models, reducing MAE by 74.3%, 40.4% and 53.9% versus standard TCN, polynomial ridge regression and MLP, respectively. A Random Forest model trained on the height-increment target (RF-Δ) produced slightly higher MAE but lower RMSE and maximum absolute error, showing a trade-off between average-error control and extreme-error suppression. Additional no-PreDH, simple increment-baseline, path-block bootstrap and rolling-layer analyses show that prediction performance depends on both process-state variables and inherited geometry rather than simple copying of the previous layer. The results support bounded, layer-wise height forecasting for single-bead WAAM walls and provide a basis for pre-adjustment error identification.