Online porosity prediction in laser welding of aluminum alloys based on a multi-fidelity deep learning framework

Pore is one kind of the typical defects in aluminum alloys laser welding. Porosity is an important indicator for evaluating welding quality, and porosity assessment has attracted increasing attention. This paper presents a multi-fidelity deep learning framework (MFDLF) that enables online porosity prediction without post-weld destructive inspection or radioactive detection. In the proposed approach, the maximum temperature on the bottom wall of the keyhole acquired by numerical simulation is used as the data of fidelity 1 (F1), and the coherent optical measurement technology is used to acquire the keyhole depth as the data of fidelity 2 (F2). After extracting the respective four fluctuation characteristics of the multi-fidelity data separately, a sparse auto-encoder (SAE) is used to fuse the four characteristics into an overall feature. Based on the obvious correspondence between porosity and multi-fidelity fusion features, the MFDLF is constructed with tandem two deep belief network (DBN) models, where the former DBN utilizes process parameters to predict the overall feature of F1 data (Feature 1) that is difficult to obtain in real time. Feature 1 is combined with the overall feature of F2 data (Feature 2) that can be obtained online to predict porosity through the latter DBN. The results show that the MFDLF can predict porosity with significantly higher accuracy than the models constructed using only single-fidelity data.