Centrifugal pumps often suffer from low operating efficiency and present challenges in accurately diagnosing impeller blade fracture faults. Therefore, this study develops a method for improving centrifugal pump efficiency and diagnosing faults based on digital twins. For efficiency improvement, a digital twin optimization framework for the centrifugal pump impeller is established. Initial samples are generated through Latin hypercube sampling, and a surrogate model is constructed by combining computational fluid dynamics numerical simulations with an approximate model. A PSO algorithm is applied to solve a multi-objective programming problem with the objectives of minimizing entropy production and maximizing efficiency. For fault diagnosis, the generated pressure and velocity contour map datasets are input into a target detection network for training, and subsequently, a stacked ensemble strategy is introduced to fuse the two types of detection results for decision-making. The findings reveal that the optimized impeller’s hydraulic efficiency is improved by 4.6%, while entropy production is reduced by 18.6%. The proposed method can enhance centrifugal pumps’ performance and achieve accurate diagnosis of blade fracture faults, thereby providing a new technical path for intelligent efficiency improvement and accurate fault diagnosis of centrifugal pumps.