The increased power of the motor directly results in a higher temperature rise effect on the rotor. High temperatures can cause turn-to-turn short circuits or permanent demagnetization of the motor rotor, which can seriously affect the reliability of the generator operation and the stability of the combat system. Therefore, the research of online rotor temperature measurement technology is of great significance. Compared to electronic sensors, Fiber Bragg Gratings (FBGs) are resistant to electromagnetic interference, small in size, require no power supply, and can be used for quasi-distributed measurements. This is a great advantage for monitoring the motor rotor temperature. However, research on FBG rotor temperature monitoring systems is still relatively rare. The state of motion of the motor rotor is an important factor affecting the accuracy of the system. The effect of rotor whirling on measurement accuracy due to unbalanced mass is even more difficult to ignore and has not been studied.In this paper, a model of FBG scanning spectra under rotor whrling conditions is developed by combining the transmission matrix theory of FBG and the coupled transmission theory of self-focusing lenses. The scanning error of the center wavelength and its influencing factors have been investigated in conjunction with relevant experiments. The results show that the whirling of the rotor leads to aberrations in the scanning spectra of the FBG, which are mainly manifested in the offset of the reflection peaks and the reduction of the 3 dB bandwidth. The main factors affecting the temperature measurement error of the system are the rotor whirling frequency, the demodulator scanning frequency, the radial displacement of the rotor at the end face, the axial ratio of the axial trajectory and the deflection angle of the axis. As the ratio of the coupling loss period to the spectral scan time (q-value) increases, the maximum center-wavelength scan error and the peak-seeking error both show a rapid decrease, followed by a slow decrease to stability. The key to ensuring a low level of temperature measurement error in the system is to ensure that q>10. The peak-finding error of the Gaussian curve fitting method is reduced to the level of the centroid method when q>40. When the rotor radial vibration amplitude is 200 μm, the axis ratio of the axial trajectory is 3, and the axis deflection angle is 0.1°, for a typical demodulator operating bandwidth of 40 nm and FBG bandwidth of 0.3 nm, if it is hoped that the measurement error of the polyimide-coated FBG to be less than 0.5 ℃, it should be ensured that the q value reaches 115 or more. The corresponding scanning frequency must be approximately 1.74 times higher than the whirling frequency. When using the centroid method or the Gaussian curve fitting method for peak finding, it is only necessary to make the scanning frequency about 0.21 and 0.44 times higher than the whirling frequency, respectively. The centroid method is more advantageous than the Gaussian curve fitting method for rotors with strong whirling.In addition, a study was conducted to investigate the effect of the intensity of the rotor whirling on the maximum scanning error at the center wavelength of the FBG and the system temperature measurement error. The results show that the maximum scanning error at the center wavelength increases slowly and then linearly as the rotor radial displacement amplitude or axis declination amplitude increases. The peak-finding error of the centroid and Gaussian curve fitting methods increases slowly and then dramatically. As the axial ratio of the rotor axis trajectory increases, the maximum scanning error and the peak detection error at the center wavelength both increase dramatically and then slowly increase to a steady state. When the amplitude of radial displacement is less than 200 μm and the axial deflection angle is less than 0.167°, the maximum temperature measurement error caused by whirling motion is 2.9 ℃, which is reflected by 15 sampling spectra under the condition of q=10 and the axial ratio of n=3. When peak detection is performed using the centroid method or the Gaussian curve fitting method, the temperature measurement error of the system is reduced to 0.6 ℃ and 1.1 ℃, respectively.