Physics-informed multi-source data fusion for inferencing residual stress field of rotary parts

Residual stress is the main reason for machining deformation, so the obtaining of residual stress is essential. Existing residual stress field construction methods that depend on a single data source still suffer from accuracy and stability issues. The use of multi-source data is a potentially effective method to improve construction accuracy. However, due to the differences in data forms, it is often difficult to use multi-source data comprehensively. In this paper, the residual stress field is modeled as a neural process that models not only the correlation between input and output but also the correlation of samples, i.e., the residual stress value of each specific position, estimating and inferring residual stress by monitored deformation forces data and local residual stress measurement data, which are taken as the input-context pair for the neural process. In the simulation verification, the inference accuracy is 49% higher than that of the Conditional Variational Autoencoder. In the physical machining verification, the average error of (Formula presented.) is 16.62MPa compared with the contour method which is taken as ground truth.