To precisely evaluate the noise induced by magnetic field and magnetic field gradient fluctuations acting on the test masses in space gravitational wave detection missions, a Multi-stage Bias Correction Model (MSBCM) is proposed for the accurate reconstruction of the magnetic fields at the test mass. Based on the ensemble learning method, the standard fully connected neural network modules and residual fully connected neural network modules are constructed as weak predictors for the MSBCM. Each weak predictor sequentially corrects the prediction biases from the preceding model, cumulatively forming a robust predictive model to realize precise magnetic field reconstruction at test mass locations. Magnetic field reconstruction experiments conducts on the LISA Pathfinder, eLISA, and Taiji-2 space gravitational wave detection spacecraft, and the proposed MSBCM method demonstrates the lowest mean relative errors along sensitive axes in comparison with other interpolation or estimation methods. In simulating on-orbit experiments, the MSBCM method achieves the root mean square error of magnetic field fluctuations and gradient fluctuations in acceleration noise on the sensitive axis of test mass 1 of 1.68×10^-17 (m/s²/Hz^1/2) and 4.00×10^-17 (m/s²/Hz^1/2), respectively. Additionally, MSBCM closely only to the distance weighted method in minimizing the root mean square error for magnetic field fluctuations and gradient acceleration noise on the sensitive axis of test mass 2, records at 1.72×10^-16 (m/s²/Hz^1/2) and 2.93×10^-16 (m/s²/Hz^1/2), further validating the advantages of the proposed method in assessing magnetic fields around test masses in space-based gravitational wave detection missions.