Low-latency network-based fast melt pool state recognition method

Real-time melt pool state recognition is key to achieving high-quality metal melting deposition, making faster and more accurate identification highly significant. However, existing deep learning-based melt pool state recognition methods suffer from relatively high latency, which is not conducive to real-time monitoring of melt pool states. In this paper, we propose a low-latency, lightweight convolution network for melt pool state recognition, which has a faster recognition speed. Specifically, a high-speed fundamental network unit is designed to extract spatial features from melt pool images more effectively, based on a specially designed simple yet fast and effective selective convolution operator that can reduce redundant computations during feature extraction. Then, we establish the low-latency melt pool state recognition network relying on the designed high-speed fundamental network unit. Finally, we built a laser melting deposition additive manufacturing system to collect an image dataset about the melt pool states. Experimental results show that the proposed method achieves a melt pool state recognition accuracy as high as 98.50%, with an execution time as low as 5.26 ms, demonstrating its superiority. Code is available at https://github.com/zhangtaosia/Low-Latency-Melt-Pool-States-Recognition.