为了解决大多数工程材料真实温度的测量问题,提出了基于BP神经网络的多光谱数据处理方法,分析了随机噪声对真实温度计算结果的影响。结果表明:在没有噪声的情况下,训练过的发射率样本真实温度的识别误差在±30K以内;未训练过的发射率样本真实温度的识别误差在±50K以内。随着随机噪声的增大,网络的识别误差也相应增大,但训练过的样本其网络的识别误差较小。说明加大发射率样本可以提高真实温度的识别精度。 In order to solve the problem of real temperature measurement of most engineering materials, a multispectral data processing method based on BP neural network is proposed, and the influence of random noise on real temperature calculation results is analyzed. The results show that in the absence of noise, the true temperature error of the trained emissivity samples is within ± 30K. The true errors of the emissivity samples are less than ± 50K. With the increase of random noise, the recognition error of the network also increases correspondingly, but the trained samples have less recognition error on the network. This shows that increasing the emissivity sample can improve the recognition accuracy of real temperature.