The first step towards intelligent wire arc additive manufacturing: An automatic bead modelling system using machine learning through industrial information integration

Abstract

Wire Arc Additive Manufacturing (WAAM) has revolutionized the manufacturing paradigm for fabricating medium to large scale metallic parts featuring high buy-to-fly ratios such as aerospace components. As a promising technology for the manufacturing industry, it is necessary to develop an automated WAAM system with high efficiency and low labour cost. Generally, to achieve a fully intelligent WAAM system, the first step is to develop an intelligent weld bead modelling system which is able to provide users with appropriate welding parameters in terms of producing components with high accuracy. Knowledge from many disciplines, such as computer science, material engineering, mechanical engineering, and industrial system engineering, is advantageous to develop such an automated system. Thus, an intelligent bead modelling system was developed by integrating a number of industrial sectors in this study. The bead modelling system includes three critical modules, including data generation module, model creation module, and welding parameter generation module. It is worth mentioning that a novel algorithm using Support Vector Machines (SVM) was proposed for creating the model with a high level of accuracy. Optimal combinations of wire feed rate and travel speed under various temperatures were generated accordingly. The experiment results demonstrated that the system can significantly improve product quality and reduce manufacturing costs, including raw material usage and manual labour.

Introduction

Additive Manufacturing (AM) has gained popularity worldwide over the past decades. It is often interchangeable when referring to the AM process with terms of 3D printing, layered manufacturing, rapid prototyping, and rapid manufacturing [1]. AM technology covers many processes in which consumable material is deposited or jointed under the control of certain Computer Aid Manufacturing (CAM), depositing three-dimensional object layer by layer [2]. Recently, the focus of AM has changed to fabricate expensive metal components such as titanium [3] and nickel alloys [4] in the aerospace industry where such components often suffer an extremely high buy-to-fly ratio [5]. Commonly, AM technologies for metallic parts include: i) wire arc additive manufacturing (WAAM); ii) Electron beam freeform fabrication (EBF3); iii) Electron beam melting (EBM); iv) Selective laser sintering (SLS); and Selective laser melting (SLM) [6]. amongst these, WAAM, which combines electric arc as the power source with welding wire as a feedstock, is able to fabricate large scale parts with high efficiency and low cost.

As shown in Fig. 1, the WAAM process consists of several major processing steps: (i) bead modelling, the relationship between the welding parameters and the bead geometries; (ii) 3D slicing, a 3D model is sliced into a set of 2.5D layers with predetermined thickness; (iii) 2D path planning, the optimal deposition path is designed to fill the 2.5D cross-sections; (iv) deposition process, the welding process parameters are selected according to the weld bead modelling and the molten material is deposited along the path layer upon layer; and (v) post-process machining, the part is machined for removing extra support structure and/or to meet a desired surface quality [7]. In the WAAM process, each layer is deposited with a large number of single weld beads side by side. The offset distance of adjacent beads has a great impact on the geometric accuracy and surface quality [5]. Besides, the height of the weld bead, to some extent, is also required to be the same as the layer thickness for 3D slicing. To realize a practical WAAM process, the relationship between weld bead geometry and welding process parameters is critical to select optimal welding process variables. However, for current WAAM systems, the welding variables are still determined manually or partially automated based on a database and operator experience that may affect the accuracy and the efficiency of the WAAM process [8]. For example, in the study by Ding, et al. [7], the authors proposed a fully automated WAAM system, which only requires the 3D profile as system input and creates the final part without human intervention. However, the bead modelling module is only capable of predicting bead geometry with the given welding parameters. Similarly, in the WAAM system proposed by Prado-Cerqueira, et al. [9], the collected experimental results were used as a database to provide information for weld settings. Therefore, it is imperative to develop an automated bead modelling system to achieve an intelligent WAAM process.

Currently, many modelling methods have been proposed for creating a forward model for weld bead formation, such as traditional regression model [10], [11], [12], Taguchi approach [13], and Artificial neural networks (ANNs) [14], [15], [16]. The forward model is referring to the predictive model that is able to predict the bead geometry based on the given welding parameters. However, it is difficult to produce accurate weld bead geometry with the limited welding parameter sets. This is considered a drawback of the forward model. In practical industrial production, parts may have complex geometries, therefore the weld bead geometries are always required to be varied slightly when fabricating parts with complex structures. Ideally, the weld bead model should be used in reverse, which means the welding parameters can be obtained based on the weld bead geometry generated in the path planning process. Xiong, et al. [8] used ANNs method to develop both forward and backward model for predicting welding parameters. The backward model is referring to the predictive model that is able to predict the welding parameters based on designated bead geometry. The authors conclude that the accuracy of the backward model is unsatisfactory, and the forward model was used to examine the accuracy of the backward model. If the error between the predicted and desired bead geometry is unacceptable, the welding variables are slightly adjusted for the next iteration until the error becomes smaller than a threshold. Note that the parameters are adjusted based on the influence of welding variables on the bead geometry according to the reference [10].

Compared to conventional ANNs based model, Support vector machines (SVMs), an alternative machine learning algorithm, has the potential to provide a more accurate and efficient solution for the bead modelling process in WAAM. In machine learning, SVMs are supervised learning methods with corresponding learning algorithms that process data used for regression analysis [17]. In many studies, the SVMs are identified to be superior to ANNs, as they avoid some major weaknesses of ANNs: (i) SVMs often converge on global minimum rather than local minimum, which means ANNs model may miss the optimal result [18]. (ii) ANNs has the limitations on generalization giving rise to models that may lead to overfit the data. (iii) The training time of SVMs is substantially less than that of ANNs [19]. In the current literature, SVMs have been demonstrated a good capability of solving regression issues for various applications [20], [21], [22]. For the modelling problem in welding, Chen, et al. [23] proposed a modelling method to predict weld penetration using SVMs in gas tungsten arc welding. The experimental results indicate that the SVM-based model shows a higher level of accuracy in predicting welding penetration by comparing with the ANN method. To date, little has been done to create a highly accurate weld bead model for WAAM using SVMs.

This study aims to develop an automated bead modelling system from weld bead deposition, data collection, and processing to welding parameters prediction in WAAM process. Conventionally, the system would be developed by manufacturing engineer. However, knowledge on a single discipline can not meet the industrial requirements of high level of accuracy, product quality, and system automation, the integration of technologies from a number of disciplines should be considered. Industrial information integration engineering is carried out to solve complex industrial problems by combining methods [24]. As comprehensively reviewed in [25,26], industrial information integration engineering is an emerging subject attracting much attention in academia. Moreover, Chen [25] also pointed out that research on the manufacturing category was the second biggest category in this industrial information integration. Thus, the bead modelling system was proposed through integrating information and knowledge from computer science, material engineering, mechanical engineering, and industrial system engineering. As the first step of the WAAM process, the fully automated WAAM system can improve product quality and reduce the time required for the design and manufacturing cycle. In addition, Machine learning is firstly used for weld bead modelling in WAAM process. A novel algorithm using SVM was proposed for the system which provides users with a set of accurate welding parameters for the deposition of weld bead with designate geometry.

In the following sections, the workflow of the proposed system is described in Section 2. In Section 3, the data collection and processing system is presented and the SVMs based modelling algorithm is detailed in Section 4. The effectiveness of the proposed system is examined through the experimental validation and a case study in Section 5, and followed by a conclusion in Section 6.