According to the non-linear and non-stationary characteristics of thermodynamic parameter parameters, this paper proposes a new fault diagnosis method for steam turbine flow passage (STFP) based on improved empirical mode decomposition (EMD) and probability neural network (PNN). In view of the end effect in the conventional EMD, an improved EMD is proposed to get more reliable results of intrinsic mode functions (IMF) by using mirror extension based on waveform similarity. Then, it is applied to decompose the thermal parameter signals to obtain a series of stationary IMF and a residual, through which the feature extraction of STFP fault is realized effectively. Finally, the feature vectors are inputted into the PNN to recognize the fault patterns. Simulation experiments are carried out based on the actual operation data of a 600MW thermal power plant unit. The results verify that: the proposed fault diagnosis method can quickly and accurately identify the fault patterns of STFP, and it has better performance than STFP fault diagnosis method based on conventional EMD-PNN.