In this article, a new operating performance assessment method based on adaptive fusion of multi-source heterogeneous information (AFMSHI) is proposed for magnesium melting process. First, the data pre-processing is performed for the multi-source heterogeneous information (MSHI) in the process of magnesium melting, and deep learning methods are used to formulate performance assessment sub-models based on different types of information. Secondly, to fully consider the impacts of MSHI on the assessment results under different melting states, the attention mechanism is used to establish an adaptive fusion network for the assessment results of each sub-model. Finally, the fused assessment results are input into a SoftMax classifier, and the magnesium melting process assessment model is formulated. The simulation results show that, comparing with the assessment model established by a single type of information or the existing deep learning MSHI assessment methods, the assessment accuracy of AFMSHI based on simulation platform data and actual production data reached 99. 5% and 98. 44% , respectively, which is higher than the compared methods by comprehensively considering the roles of MSHI. The effectiveness and the superiority of the proposed method are verified.