In deep foundation excavation projects, using reasonable soil mechanical parameters to calculate the lateral deformation of diaphragm wall is essential to optimize the support design and reduce engineering risks. However, the soil parameters are generally affected by the uneven distribution and geotechnical testing errors, which often show obvious uncertainties and reduce the credibility of the lateral deformation calculated of diaphragm walls. In view of the considerations above, this paper proposes a back analysis method of soil parameters based on Bayesian parameter updating framework and site monitoring data. This method uses GA-BP neural network to establish the implicit function relationship between the soil parameters and the diaphragm wall lateral displacement in the numerical model, and combines the site monitoring data to establish the Bayesian back analysis model of the soil parameters. This method was used to analyze a deep excavation project, and the feasibility of the method was verified. The maximum lateral displacement and multi-point displacement value of the diaphragm wall were used as indicators to invert the soil mechanical parameters, and the updated soil parameters were used to predict the final lateral displacement. The results show that compared with the non-updating soil parameters, the variation coefficient of soil parameters decreases after updating, and the obtained results fit with the monitoring results better in the subsequent construction steps; the prediction effect of using multi-point observations for soil parameter updating is significantly better than that when only the maximum displacement value is used.