Multispectral-band radiation thermometry method based on double-stage emissivity neural network

Uunknown emissivity of objects is an unignorable obstacle in radiation thermometry, especially for multispectral-band pyrometry. Assumed wavelength-emissivity modules are wildly used for solving the underdefined system of multispectral-band pyrometry while the accuracy is strongly dependent on the similarity between real object emissivity characteristic and the assumed one. Data processing methods based on optimization theory have the capability to solving the underdefined system without the emissivity knowledge, however, the calculation time could not satisfy the need for real-time use. Here, we propose a double-stage neural network for multispectral-band pyrometry in the NIR range independent of advanced emissivity knowledge. The first stage is trained by dataset generated by blackbody source and Plank’s law with ideal spectral-band radiation energy output, while the second stage is trained by dataset generated by ideal blackbody spectral-band radiation energy added different type wavelength-emissivity module (graybody, linear, sine, mixed graybody-linear-sine), and network final output is the true temperature. Multispectral-band temperature measurement experiment results of Ti-6A1-4V alloy shows that the relative error of predicted temperature and the true temperature is smaller than 2.5% for the neural network trained by single type wavelength-emissivity module, and the relative error is less than 3.0% for the neural network trained by mixed-type wavelength-emissivity module.