Sensor fusion and adaptive learning for optimization and control of Wire-Arc Additive Manufacturing – with applications to functionally graded materials

A niche research trend in advanced materials is the creation of materials that are non-isotropic or graded in single or multiple directions of a component, which is referred to as functionally graded materials (FGM). Directional material properties are seen in nature in various forms and this can be utilized in engineering applications where a gradation in thermal, mechanical or tribological properties are required. FGMs can be created with wire-arc additive manufacturing (WAAM). WAAM is an additive manufacturing (AM) method that used the heat created from an arc to deposit and build material (usually metallic). Twin-WAAM (TWAM) upgrades the original WAAM process by using two wire feeders. Varying the feed rates of these wires independently enables us to produce FGMs strategically in one or more directions. The production of FGMs using TWAM is a novel yet is a complex challenge. The process needs to be optimized and, thus, requires a robust monitoring system. WAAM has a scarce literature on monitoring but monitoring strategies on other similar processes such as Direct Energy Deposition (DED) and Arc Welding can be referenced due to their similarity with WAAM [1]. The typical sensors that are used in literature are vision, acoustic, spectral, and thermal sensing. Vision and spectral sensors and video-based thermal sensors generate video feeds, while acoustic sensor signals require high sampling rates to detect frequencies at the MHz range. The variety, velocity, and volume of these sensor signals will be costly when implemented in an online in-situ monitoring system and should be managed. A potential approach is by using a low level representation of the process state which is achieved by using Deep Convolutional Neural Networks (DCNN) [2]. With a correct design and implementation, properly trained DCNNs could detect the low-level process features, predict defects and monitor the process quality. The natural continuation is then to use these monitored signals to control the process, by creating corrective actions that can reduce or eliminate the defects. Classical control algorithms such as PID or fuzzy controllers exist, but machine learning – reinforcement learning, can potentially improve the current control systems. WAAM process is very dynamic in a way that the buildup layer can affect the signals significantly. Complex parts also make the process challenging to predict. A typical controller may be able to preserve geometry and microstructure details given when they are designed to simple components, though these controller designs may not be able to adapt to more complex geometries. Reinforcement learning can be used to adapt to unexpected changes by learning from experience and improving the controller’s performance through active observation and interaction to the process[3]. This PhD project aims on the following: 1. Review the existing sensing methodologies for WAAM or similar processes, e.g. Direct Energy Deposition (DED), Fused deposition method (FDM), Arc Welding, etc. in the context of FGMs 2. A study on the potential features that are generated during the WAAM process; down-select the sensors that can measure these features; design and implement monitoring strategy 3. Data collection, analysis, and interpretation; develop ML models to predict defects and final material composition from the WAAM process 4. Demonstration of in-situ monitoring using the developed models and validation of these models with actual components 5. Development of a control system (using ML/RL methods) and comparing with benchmark control methods [1] C. Xia et al., “A review on wire arc additive manufacturing: Monitoring, control and a framework of automated system,” Journal of Manufacturing Systems, vol. 57. Elsevier B.V., pp. 31–45, Oct. 01, 2020, doi: 10.1016/j.jmsy.2020.08.008. [2] J. Günther, P. M. Pilarski, G. Helfrich, H. Shen, and K. Diepold, “Intelligent laser welding through representation, prediction, and control learning: An architecture with deep neural networks and reinforcement learning,” Mechatronics, vol. 34, pp. 1–11, Mar. 2016, doi: 10.1016/j.mechatronics.2015.09.004. [3] R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction. 2018.