Wire Arc Additive Manufacturing (WAAM) is a process involving an electric arc heat source manipulated by either a numerically controlled machine tool or a robot [1]. The torch deposits a metallic bead onto a substrate, and the accumulation of these beads forms the final part [2]. Various heat sources are available to modulate both input energy and material feed [3]. One such solution is the combination of gas metal arc welding (GMAW) with cold metal transfer (CMT) technology. This approach facilitates high-rate material deposition and offers improved control over energy. WAAM is versatile, allowing the use of any weldable material.
However, this technology encounters several challenges [4], [5]. Various issues may arise during manufacturing, including porosity [6], [7], [8], residual stress [9], [10], microstructure concerns [11], and dimensional defects [12], [13]. Developing pertinent sensors for in-operando monitoring of these defects is the subject of many papers [14]. Several metrics can be monitored, such as temperature on a point [15] or welding current [14] but the melt pool seems a good indicator to predict defects [16]. The use of a near-infrared (NIR) camera (like a CMOS sensor) is identified as a suitable means to capture the melt pool shape [17]. From these images, it is possible to produce a thermal metric correlated to defects. Dellarre [18] shows that the image post-processing that produces the thermal metric is a key step in improving measurement accuracy. Moreover, WAAM can easily produce thin walls, and managing the thickness of the manufactured bead is a key challenge to produce near-net shape parts [19]. This article focuses on thin wall manufacturing and thickness as a quality indicator.
This article aims to propose a method to compare the accuracy of different post-processes of melt pool images from a CMOS NIR camera of a WAAM process to monitor deviation of wall thickness.
Aluminum alloys are particularly well adapted to WAAM manufacturing as they are dedicated to light parts, that is why this article focuses on aluminum alloy WAAM manufacturing. For steel manufacturing, it is easy to detect the melt pool from a CMOS image using canny or adaptive filters as demonstrated in Fig. 1. However, adapting these algorithms to aluminum alloys poses greater challenges. This is primarily because the thermal emission light from the aluminum melt pool is 1000–10,000 times lower than that of steel, and the boundaries of the aluminum melt pool are less distinct [17]. Figure 1 visually illustrates the contrast between two raw images captured during the manufacturing processes of aluminum alloy and steel.
The paper is organized as follows. First, the input data used to benchmark data are presented and analyzed. Then the method to evaluate the accuracy of a post-process is proposed and explained. The benchmarked post-processes are presented and classified into two categories: conventional image processing and neural network image processing. Results presenting the accuracy of the post-processes are discussed, followed by conclusions and perspectives.