Magnetic Barkhausen Noise （MBN） technique can be used to quantitatively evaluate the surface stress of ferromagnetic materials. The current MBN stress assessment technology has the disadvantages of difficult feature selection， complex quantitative prediction model and low fitting accuracy of the calibration data set. A data-driven nonlinear mapping algorithm is proposed to fit the relationship between MBN noise and stress. The time-frequency feature based on wavelet packet transform coefficients is used to replace the statistical feature， which reduces the amount of sample data calculation. The wavelet packet transform coefficients of MBN noise in the wavelet packet transform time-frequency domain are used as eigenvectors. The dimensionality reduction algorithm based on singular value decomposition is used to reduce the dimension of the eigenvectors， and the eigenvectors after data dimension reduction are input into the Back Pagation （BP） neural network. Model training is performed to build predictive models. The results show that the data dimensionality reduction algorithm based on singular value decomposition can reduce the complexity of the model， and the BP neural network can be trained by using the eigenvectors of the wavelet packet transform coefficients after dimensionality reduction to achieve high-precision prediction of surface stress of ferromagnetic materials. The characterization method proposed can effectively solve the problem of stress distribution imaging of ferromagnetic components， and has great potential in application in stress corrosion prevention， fatigue strength improvement， and other damage early warning.