Introduction
Laser cladding (LC) is an additive manufacturing process in which a laser beam melts feedstock material into a substrate, producing a clad bead. Disturbances on the resulting clad bead geometry usually lead to deviances on the part’s final shape, loss of surface quality, and even structural defects, leading to the disposal of the final part. It is, thus, crucial to closely monitor this geometry. An alternative for its monitoring is through the acquisition of the molten pool image.
Through optical monitoring of the molten pool, one can acquire important information about the process. The molten pool size reveals information of the final geometry of that clad bead, while its brightness is closely related to molten pool temperature.
The first approaches aiming at molten pool image acquisition for the laser cladding process consisted of the use of multiple cameras with different specifications directed to the molten pool, successfully measuring the clad bead height and width. The very first approaches, by Meriaudeau et al, consisted of two different cameras directed at the molten pool, on different angles \cite{No1168,No1164,No1163,Meriaudeau}. It was then possible to acquire the molten pool image for the first time, even measuring clad bead height, width, and molten pool temperature. A later approach, by Hu et al, developed a first closed-loop controller where the laser power was regulated accordingly to the area of the molten pool \cite{No1176,No1169}.
Following, Toyserkani et al developed a closed-loop PID-based control system based on clad bead height measurement \cite{No1162}. Xing \cite{No1160} developed an optical monitoring system based on colorimetry, achieving both molten pool temperature and clad bead height measurement through laser triangulation. Later, Hofman et al gathered molten pool infrared (IR) wavelengths with a laser coaxial IR camera. After some image processing, the molten pool boundary could be segmented and its area, width, length, and rotation angle were measured. Molten pool width was then used for a PID-based closed-loop control, which regulated molten pool temperature by adjusting laser power \cite{No315}. Lei et al achieved a molten pool image acquisition on a CO2 laser-based cladding process, measuring molten pool temperature through the image’s brightness.
Arias et al \cite{No1156} developed an FPGA-based system which gathered mostly IR wavelengths detecting molten pool based on blob detection. With the molten pool width as the biggest detected blob’s width, a closed-loop control was implemented by adjusting laser power. From the molten pool image also came the work of Ocylok et al \cite{No1172}, where image thresholding successfully segmented the molten pool and relations between its geometry and process parameters were found.
Moralejo et al \cite{No1157} developed a PI-based control-loop for molten pool geometry in real time. Molten pool border was previously detected by laser cladding experiments without the addition of powder. Its width was chosen as the control variable due to its overall stability.
Iravani-Tabrizipour et al \cite{No318,No317} continued the work from Toyserkani et al \cite{No1162} implementing a trinocular system for measuring clad bead height from three radially spaced cameras. After fitting the molten pools into ellipses, their main features were fed into a recurrent neural network (RNN) which calculated clad bead height. This became the first works that made use of neural networks for geometry estimation on laser cladding.
Mondal et al \cite{No320} aimed at finding a relationship between bead geometry and process input parameters. Through cross sectioning, bead geometry was acquired. This bead geometry, along with the process input parameters (laser power, travel speed, and mass flow) were fed into an artificial neural network (ANN), yielding a coefficient of determination (R²) of 0.981 for the best fit line.
Aggarwal et al \cite{No312}, targeting at bead geometry optimization, took three approaches – one experimental, one based on predictive models and the last on ANN, all of them based on cross sections measurements. The ANN approach surpassed the remaining ones, with 96.3% of confidence level on its results.
Caiazzo et al \cite{No313} acquired data from cross sections to develop an ANN capable of estimating process parameters. At first, the ANN was used to estimate the process parameters from bead geometries that were already deposited, based on their cross-sections. Later, it was reversed, using the bead geometry to estimate process parameters, achieving errors below 6% for travel speed and powder feed rate, and of 2% for laser power.
Finally, Huaming et al \cite{No316} predicted geometric characteristics from input process parameters by a genetic algorithm and backpropagation neural network-based approach (GA-BPNN). Again, based on cross-section measurements, the ANN were trained, but their architectures were optimized by the use of genetic algorithms (GA). The resulting networks trained much faster than the initial ones, achieving superior results. It was also observed the better performance of the networks which output only a single parameter at a time.
Other approaches involving either the molten pool or the clad bead measurement consist on low-cost alternatives \cite{No1166}, laser triangulation over newly formed clad bead \cite{No1155}, and even a methodology based on previous experiments calibration \cite{No1150}. An approach to temperature acquisition through black body calibration could also be found \cite{No1173}.
An attentive reader may observe mainly two types of research. While most of them use the molten pool image in order to acquire molten pool related dimensions, others use solely bead geometry measurements acquired from cross sections for bead geometry estimation with ANN techniques. An approach where both molten pool image and process parameters are combined and fed to an appropriate ANN has not been found on this review.
This work presents a novel approach where the molten pool image is directly processed by a convolutional neural network (CNN). As input, the network takes both molten pool image and input process parameters, estimating clad bead width and height in return. Six different CNN architectures were tested. To make the network training feasible, measurements from the clad beads were acquired my active photogrammetry means. Different height and width values were measured to match each corresponding image frame. After training, the CNN estimated width and height values in great agreement with the experimental values.