Performance monitoring and fault diagnosis for wastewater treatment processes was of great significance for safeguarding the normal operation of the treatment process and ensuring the standard quality of effluent water. Aiming at the problems of nonlinearity, uncertainty and susceptibility to random noises in wastewater treatment process, an improved Laplacian Eigenmap (LE) manifold learning algorithm based on commuting time distance (CTD) was proposed to realize the feature extraction of the complex process data. In this algorithm, CTD was used to measure the similarity between samples and construct the neighborhood graph. Both theoretical analysis and simulation test proved that the proposed algorithm could efficiently overcome the sensitivity problem caused by neighborhood parameter and improve the robustness of the normal LE algorithm. Then the CTD based LE algorithm was applied in fault detection modeling for actual wastewater treatment process, and the fault monitoring statistic was constructed in the low-dimensional feature subspace. Application results showed that CTD-LE based model can timely detect the faults with lower missing rate and false rate as compared with normal PCA based model and normal LE based model. Application results showed that this method could provide a feasible solution for fault monitoring of complex industrial processes such as wastewater treatment.