This paper first discusses the classification of ink experiment visual language and analyzes its basic features and functional value. Secondly, the residual network is utilized to represent the visual features of ink experimental images, and the BP neural network is introduced as the basis for the construction of a feature-level fusion network. Then the DS evidence theory is used for the decision-making of ink experimental visual feature images, and then the multi-source heterogeneous data fusion model is constructed. Finally, the training accuracy is tested and empirically analyzed for the multi-source heterogeneous data fusion model. The results show that the testing time of the multi-source heterogeneous data fusion model based on the BP neural network is 0.52s, which is 0.24s and 0.15s higher than the original VGG19 and VGG16 networks and the interpretation of the visual language of the ink experiment is mainly based on the static ink, dynamic ink and interactive ink. This shows that the interpretation of ink experimental visual language can be realized through multi-source heterogeneous data, which also provides a new direction for the innovative development path of ink art.