AI-aided thermal imaging with multispectral camera for Direct Energy Deposition

Measuring the temperature of red-heat or molten metals is relevant for Additive Manufacturing (AM) process monitoring and control. Sensors commonly integrated in AM machines are pyrometers and thermal cameras: the first can provide accurate but localized measurements while the latter can visualize a temperature distribution over a larger area but are very sensitive to the emissivity parameter, usually unknown. This paper presents a novel method for camera calibration and temperature measurement based on a multispectral camera, combining accuracy and spatial resolution. In the calibration phase optical properties of the camera are estimated by comparing the observed spectra of an AISI 316L substrate heated by an oven between 800 and 1100 °C with simulated ones. Then, a synthetic dataset is generated, extending the simulation range from 750 °C to 3950 °C, well above the temperature reachable by black body sources, and used to train a neural network that retrieves temperature from live images. The method reaches a few Celsius degrees accuracy on calibration images and generalizes well to other materials like Inconel 718. Preliminary tests of the calibrated multispectral camera, installed in a Direct Energy Deposition machine, show promising results estimating the temperature of the melt pool compatible with the melting point of the material.