The absence of excitation characteristic parameters significantly hinders the efficacy and precision of underwater noise predictions, especially under multiple excitations. Addressing this issue, a novel noise prediction reconstructed source method has been proposed for the noise analysis of a stiffened cylindrical shell. This method diverts attention from the actual sources and reconstructs the source utilizing the shell's vibrational response and transfer functions. The structural vibration response induced by the reconstructed source remains congruent with the actual behavior. The issue of non-unique solutions inherent in the reconstructed source is resolved through the application of the least squares principle. With the reconstructed sources and the stability of the transfer functions, underwater noise prediction becomes readily attainable. The method has been substantiated via an experiment from a stiffened cylindrical shell. The results indicate that the noise predictions derived from the reconstructed source method exhibit excellent agreement with experimental data, the overall level error of noise prediction is up to 0.7 dB. Furthermore, in comparison to the BEM, this method significantly enhances the efficiency under identical computational settings, with the prediction duration constituting a mere 1% of that required by the BEM.