Small faults with insignificant abnormal symptoms were usually submerged in large faults or noise.Most fault diagnosis methods were invalid in the case when small faults occurred.We present a multi-level space decomposition mechanism and a small fault diagnosis algorithm.Implement designated component analysis (DCA) to the observation data for significant variation patterns;Remove the effect of significant designated patterns to get the residual which will increase the signal-to-noise rate of small fault signal;Determining whether small faults have occurred in the system using the projection significance index in the residual space;Repeat this process until all possible small faults are diagnosed.Simulation for observation data involved 4 faults shows its efficiency of this algorithm.