Abstract: Due to the fact that both of the temperature and stress have effect on MBN signals, in the detection process of the rail temperature stress, so how to get the accurate stress value by temperature compensating for the test stress was a key point of this study. A temperature stress detection system by Barkhausen noise, based on the BP neural network, was built by taking the temperature of the ferromagnetic specimens, the frequency and voltage of the excitation signal, and the mean value, RMS value, ringing counts and peak-wide ratio of the processed Barkhausen signal as the main impact factor, and taking the corresponding compressive stress of specimen as the output data. After being trained by many training samples, the network was tested by some test samples. It showed that the required accuracy was reached by comparing the test results with the actual stress value, which indicated that the stress detection system by Barkhausen noise based on BP neural network obtained the perfect temperature compensation, and was very efficient and accurate.