As an important part of dam seepage system, the quality of concrete cut-off wall is of great importance. In order to obtain effective feedback on the quality of the wall with limited observation data, this paper introduces the idea of multi-scale parameter inversion. Meanwhile, considering the instability of this ill-posed inversion problem, a regularized coefficient is derived to control the reliability of the final solution. Afterwards, we deduce the analytic sensitivity expression of the seepage head, and simplify the expression based on the adjoint state theory, which greatly improves the computational efficiency of the ill-posed multi-scale parameter inversion. Then, EM clustering algorithm, multi-scale data edit and semi-supervised fuzzy particle swarm optimization algorithm (SFPSO) are proposed to optimize and classify the parameter samples. Finally, we can make an effective identification of the damage inside the cut-off wall through clustering ensemble and progressive learning of the previous sample classification. To verify the accuracy and reliability of this method, we compare the calculation results with actual experimental data. The results show that the proposed multi-scale progressive inversion method is in good agreement with the experimental results and can be used to identify the damage of the actual cut-off wall.