Aiming at the fact that there is no preimage in nonlinear mapping and fault variable cannot be identified which result in that it is difficult for engineering application of kernel principle analysis, an improved KPCA residual direction gradient algorithm is proposed to overcome the above drawbacks in this paper. By use of the correlation between the partial differential of principle statistic and residual statistic, the gram matrix partial differential intermediate computation process is simplified and the KPCA residual direction gradient index is obtained, combined with residual statistic a new fault detection method is proposed. Nonlinear system simulation computation shows that improved KPCA residual direction gradient method has excellent capability of fault variable identification while computational complexity is greatly decreased and the calculation time is shortened. Furthermore, largescale thermodynamic system application shows that the proposed method has better capability in fault detection whenever in case of single fault or multiple faults and there is no residual contamination while it is very suitable for engineering realization.