Online porosity monitoring in laser metal deposition via MFST-Net: A multimodal feature-level spatiotemporal fusion approach

In recent years, data-driven approaches for online defect monitoring in metal laser additive manufacturing (LAM) have achieved remarkable progress. However, most existing studies primarily rely on spatial features extracted from single-modal transient images, which are insufficient to capture the temporal evolution characteristics of the melt pool and the associated variations in local thermal history during the laser metal deposition (LMD) process. Moreover, the complementary information provided by multi-sensor data has often been overlooked. To address these limitations, this study proposes a multimodal feature-level spatiotemporal network (MFST-Net), which enables joint modeling and deep fusion of melt pool image sequences and in-situ process temperature signals. Specifically, a spatiotemporal feature fusion neural network (STFNN) is constructed to extract spatial distribution patterns from melt pool images while capturing multi-scale temporal dependencies at both the intra-layer and inter-layer levels. In parallel, a self-attention convolutional long short-term memory (SAConvLSTM) network is employed to model the dynamic evolution of thermal signals. Finally, cross-modal feature fusion is performed at the feature level to characterize the relationship between thermal–morphological evolution and pore formation mechanisms. Experimental results demonstrate the effectiveness and superiority of MFST-Net in online monitoring of local porosity, achieving an accuracy of 95.8%. These findings provide a promising reference for the integration of multimodal spatiotemporal feature fusion in complex manufacturing process monitoring.