Abstract: Dough mixing is one of the indispensable steps in the preparation of wheat noodles. Conventional processing depends mainly on empirical judgement to identify the end point of mixing. However, an automated evaluation is still lacking using artificial intelligence analysis. In this study, an automatically mixing was realized to prepare the wheat noodle. The mixing stages were divided to evaluate the macroscopic state of the dough crumbs during mixing. A big data analysis was performed on the dough crumbs images, then to determine the physicochemical properties of wheat noodles, including the texture properties, moisture distribution, protein molecular properties, and microstructure features. The results showed that the dough mixing was divided into five stages, including original flour (stage 1), flour wetting and adhesion (stage 2), crosslinking and formation (stage 3), breaking and dispersing (stage 4), and stable equilibrium (stage 5). Among them, the time of the stable equilibrium stage accounted for about 50% of the total dough mixing time. The hardness, springiness, and chewiness of the dough sheets gradually increased and then reached the highest from stage 2 to the beginning of the stable equilibrium stage (stage 5-1). The texture parameters decreased in the middle and late stages of the stable equilibrium stage(stages 5-2 and 5-3), such as the hardness of the dough sheets. The low field nuclear magnetic resonance (LF-NMR) and NMR imaging showed that the moisture degree of freedom was high during the wetting and adhesion stage. The reason was that the mixing process accelerated the exchange of moisture between the solid, liquid, and gas phases in the dough system, where the stability reached stage 5. The intensity of hydrogen bonds and ionic bonds in all mixing stages was significantly lower than that of hydrophobic interactions. In addition, the content of free sulfhydryl groups decreased to be stable at stages 5-2 and 5-3. The disulfide bonds content gradually increased and then tended to be stable in these stages, where the proteins were fully cross-linked. The fully cross-linked proteins were determined the microstructure and macroscopic state of the noodle. The Transfer-ResNet50 network model was established using deep learning to realize the automatic recognition of dough crumbs images at the different mixing stages. The recognition accuracy of the model was up to 98.5%, indicating excellent accuracy and reliability. The automatic judgment can be accurately implemented at the end point of the mixing process. In summary, the macroscopic particle state of the dough crumbs can be used for the division of the mixing process, and deep learning can provide new ideas for the automation of the dough process.