Optimal droplet transfer mode maintenance for wire + arc additive manufacturing (WAAM) based on deep learning

Abstract

In the last decade, wire + arc additive manufacturing (WAAM), which is one of the most promising metal additive manufacturing technologies, has been attracting high interest from both academia and industry. WAAM systems are increasingly employed in the industry and academia, but there are still several challenges and barriers to process stability control. The process stability is highly dependent on how the molten feed wire is added into the melt pool, which is known as the droplet transfer mode. To ensure a stable WAAM deposition process, it is essential to maintain the transfer mode in a suitable stable status. Without an effective transfer mode control method, the operators need to determine and control the transfer mode based on their experience using manual adjustment, which is difficult to achieve in a long period of production process. In this paper, a deep learning-based technology was proposed for the control of the droplet transfer mode based on the data collected from the WAAM process. A long short term memory neural network was applied as the core transfer mode classification model. A time-series data, arc voltage, was collected and statistical and frequency features were extracted, which included 11 relevant features, as the inputs of the classification model. Then, the distance between the melted wire and the melt pool was adjusted based on the determined transfer mode to keep a suitable stability of the process. A case study was used to evaluate the proposed approach and to show its merit. The proposed approach was compared to three commonly used machine learning algorithms, k-nearest neighbours, support vector machine, and decision tree. The proposed method obtained the highest accuracy in determining the transfer mode, which was over 91%. The performance of the proposed approach was also evaluated by the single-pass and oscillated wall building. The proposed deep learning based approach improved the process stability in real-time, which resulted in better deposition qualities, in terms of geometry size and processing cleanliness compared to without control. Furthermore, this data-driven method could be applied to other WAAM processes and materials.