A Multimodal Information Fusion Network Combining CNN and Adaptive Mamba for Fault Monitoring of Selective Laser Melting Ceramic Printing Instrument

Selective laser melting (SLM) technology based on additive manufacturing has been widely used in ceramic printing. Whether or not the SLM ceramic printing instrument is processed normally will directly determine the quality of the final product. Therefore, the monitoring of ceramic printing instrument is of particular importance. Existing monitoring methods lack contextual modelling and multi-source information fusion capabilities, resulting in a significant degradation of monitoring performance. In this research, a multi-modal information fusion network (ResMambaNet) combining convolutional neural network (CNN) and adaptive Mamba is proposed for instrument monitoring in SLM ceramic printing. The method achieves global-local modelling of visual, infrared thermal imaging (ITI) and acoustic emission (AE) signal features through three independent feature extraction branches. In three branches, residual convolution is used to extract local detailed features in each modal signal. Adaptive Mamba is used to dynamically capture and model the global dependencies between different modalities. In addition, a multi-directional interactive learning module (MILM) is designed to achieve effective fusion of multi-modal signals. This module facilitates feature sharing and information complementarity between multi-modal signals through a multi-directional interaction mechanism, which effectively improves the monitoring capability of the model. When tested on a collected multi-modal ceramic printing dataset, our proposed ResMambaNet outperforms twelve state-of-the-art single-modal and multi-modal methods with a monitoring accuracy of 97.31%. Experiments validate the utility of ResMambaNet. In addition, high monitoring accuracy is also achieved on the publicly available laser directed energy deposition (LDED) dataset.