In recent years, in order to reduce the dependence on fossil fuels and the strong demand for energy conservation and environmental protection, biomass boilers with rice husk, straw and other renewable resources as fuel have attracted more and more attention in the industry. In order to further improve the combustion efficiency of the biomass boiler and optimize the furnace structure, it is necessary to monitor the flame temperature real-time. The traditional thermocouple temperature measurement method is not conducive to long-term high temperature measurement, and the CCD temperature measurement method is difficult to measure the real temperature, while the multispectral radiation temperature measurement method has the advantages of fast response, no upper limit of measurement and can obtain the real temperature, which is one of the most powerful tools for measuring the flame temperature of a biomass boiler. Multispectral radiation thermometry is to obtain the real temperature by measuring the spectral radiation intensity information of a certain point of the object to be measured and inversing it with Planck formula. However, the unknown spectral emissivity is the biggest obstacle in the inversion process of multi-spectral radiation thermometry. At present, a group of emissivity models (emissivity wavelength or emissivity temperature models) are usually assumed in advance. If the assumed models are consistent with the actual situation, the inversion results can meet the requirements. If the assumed models are inconsistent with the actual situation, the inversion results have a large error. Whether the direct inversion of true and moderate spectral emissivity can be realized without any spectral emissivity hypothesis model is always a hot and difficult topic in the theoretical research of multispectral radiation thermometry. For this reason, a constrained optimization algorithm of penalty function is proposed, which transforms the inversion problem of multi spectral radiation temperature measurement into a constrained optimization problem. Because the outer point method is adopted, it is not affected by the initial value of emissivity, which further improves the adaptability of the algorithm to the emissivity of different materials. Based on the flame-s model optical fiber spectrometer produced by American ocean optics, a flame temperature measuring device of rice husk biomass boiler furnace was built. After the calibration of blackbody furnace in the laboratory, the temperature of the furnace flame of rice husk biomass boiler from initial combustion to stable combustion was measured, and the results were compared with those of thermocouple. The results show that the maximum absolute error is 35.7 K, and the maximum relative error is 3.2% compared with the thermocouple. The results show that the measurement device and the inversion algorithm can realize the measurement of the combustion temperature in the furnace of the biomass boiler, which provides the preliminary research basis for the subsequent combustion diagnosis and boiler design optimization of the biomass boiler.