In this paper, a measurement system based on the FWHM (Full Width Half Maximum) method for junction temperature of LED is designed, which uses ordinary spectrometer. First, the relative spectral distribution of various color LEDs under different ambient temperatures and driving currents are measured by a general spectrometer. Since the spectral data collected by the spectrometer are discrete, in order to obtain a more accurate FWHM, it is necessary to fit the discrete spectral data into a continuous peak near the half of the strongest peak value Imax, i. e. 0.5Imax. So, the more accurate FWHM at different temperatures can be calculated and then the functional relationship between junction temperature Tj and FWHM can be obtained by fitting a certain function. Experiments show that the linear relationship between Tj-FWHM function of white and blue LED is higher than that of other color LED, and its linear index R2 is very close to 1. It shows that the two parameters of Tj and FWHM have strong linear function relationship. Using the functional relationship between Tj and FWHM, the junction temperature corresponding to any measured value FWHM can be calculated. Because this method uses the normal driving current, the self-heating effect can not be neglected. In order to reduce the junction temperature rise of LED devices caused by self-heating effect in fixed reaction time and the measurement error caused by the temperature deviation introduced by temperature control system, Tj and FWHM in a certain state are selected as the reference state, and the corresponding Tj and FWHM are obtained by point-by-point difference method, and then the corresponding ΔTj and ΔFWHM are obtained. ΔTj and ΔFWHM are fitted into corresponding linear functions to obtain the calibration function, which can greatly reduce the deviation caused by the self-heating effect and temperature control system. Finally, the results obtained by this method are compared with those obtained by T3Ster instrument of Mentor Graphics. It is found that the deviation is 2.5%, which is within the acceptable error range. The results show that the proposed method of measuring the junction temperature of LED by the FWHM method is feasible. This method overcomes the shortcomings of small peak wavelength drift of spectroscopy, which brings great errors to the test results, and has the advantages of not destroying the original packaging structure and not requiring expensive instruments.