As the main component of the loss quantity of urban water supply network, the leakage quantity can be used as an important reference index for pipe network loss control. The existing calculation method of leakage is not accurate and the calculation process is complicated, whereas the blind source separation method needs less data and can achieve more accurate analysis results. However, the constrained independent component analysis (CICA) algorithm in blind source separation has the problem of misconvergence in the estimation of leakage quantity, which limits the further application of the blind source separation method. To solve this problem, an improved ICA-R algorithm was proposed to estimate the leakage of water supply network. The selection of objective function and signal processing were analyzed. An example pipe network was designed and the leakage experimental platform was built to verify the separation effect of the algorithm, and results were compared with the treatment effect of fast independent component analysis (FastICA). Pearson correlation coefficient was adopted to evaluate the effect of leakage separation, and average relative error was used to evaluate the separation accuracy of the improved ICA-R. Results show that the improved ICA-R algorithm could maintain the correlation coefficient of separation leakage at around 90% and the average relative error was within 15%, which was better than the FastICA algorithm. Instead of adding a large number of meters, the proposed method reduces the cost of calculation, and can provide reference for the calculation of leakage, so as to facilitate the water department to take targeted measures to control loss.