A Hybrid Waam Monitoring System to Correlate Electrical Signals with the Contact-Tip-Working-Distance

Graphical Abstract A Hybrid WAAM Monitoring System to Correlate Electrical Signals with the Contact-Tip-Working-Distance Paul D. Rosero-Montalvo, Martin Martinez-Baltar, Roi M´endez-Rial, F´elix Vidal-Vilari˜no This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5007393 Preprint not peer reviewed A Hybrid WAAM Monitoring System to Correlate Electrical Signals with the Contact-Tip-Working-Distance Paul D. Rosero-Montalvoa, Martin Martinez-Baltara, Roi M´endez-Riala, F´elix Vidal-Vilari˜noa aSmart Systems and Smart Manufacturing Group, AIMEN Technological Center, Pol´ıgono Industrial SUR-PPI-2, Porri˜no, 36418, Pontevedra, Spain Abstract Wire arc additive manufacturing utilizes an electric arc as a heat source to deposit metal wire layer by layer, enabling the production of large and complex metal components with reduced material waste and lead time. However, this process needs high voltage and current flow to keep the arc on melting the electrode, which is unstable since it injects noise into the electrical signal. Conversely, the wire arc spreads rays in several wavelengths, blinding most cameras. Therefore, this work aims, on one side, an electrical signal analysis must be done to detect outliers that could affect the weld pool at this height or width. On the other side, a welding camera could detect the contact-tip-working-distance of the wire to correlate with the electric pulses, validating outliers. As a main result, the hybrid monitoring systems could check where the welding machine fails when it is welding the material. Keywords: waam, monitoring system, additive manufacturing, sensor data, ML models 1. Introduction Personalization of products with changing customer requirements affects product lifecycles, forcing manufacturing companies to shorten product development ramp-up and industrialization periods. Moreover, Manufacturing companies have shifted to a circular manufacturing strategy to integrate products’ lifecycles, making them green and sustainable, supporting de-manufacturing operations and optimization, and creating a more susPreprint submitted to Additive Manufacturing Letters October 22, 2024 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5007393 Preprint not peer reviewed tainable product design to build a sustainable and responsive production environment [1, 2]. These factors pose continuous challenges to product developers and manufacturing companies in staying competitive in the global market, demanding product and production line design optimisation [3]. In this scenario, additive manufacturing (AM) is a well-known technique in the industrial sector. It is a customized part fabrication by adding layers with industrial materials [4]. It is time-efficient by reducing scrap and eliminating supply chain management and negligible tooling requirements. Furthermore, AM significantly reduces energy or fuel with less carbon print and greenhouse gases with a friendly user interaction between the robot that will do the task and the computer-aided design (CAD) to diminish surface errors [3]. Under AM standards, wire arc additive manufacturing (WAAM) utilizes an electric arc as a heat source to deposit metal wire layer by layer, enabling the production of large and complex metal components with reduced material waste and lead time [5]. In multi-layer wall geometries, the cold metal transfer (CMT) process might be the suitable fabrication configuration since it provides the best deposition control, leading to finer weld beads and better surface quality [4]. However, these benefits come with challenges. The lower heat input could result in insufficient fusion between layers, leading to potential defects such as lack of penetration or bonding issues [6]. Additionally, the reduced heat can exacerbate residual stresses and microstructural variations, particularly in thicker components, affecting the final product’s mechanical properties and dimensional accuracy [7]. Electrically, the CMT process tries to maintain stable voltage and current pulses to establish the arc between the metal wire and the melt pool. However, the success of this industrial process depends on internal factors, such as the wire speed, the layer-by-layer waiting time and the melted material [8]. Also, external parameters, such as melt pool temperature, humidity and temperature, could inject noise into the control system and degrade the quality of the metal piece built [9]. In a WAAM process with a CMT configuration, the control system usually runs into the welding machine, and the articulated robot executes the designed trajectories; both are in continuous communication where an operator monitors the process [4]. However, for the operator, the control system is a black box where it does its best to only see the internal parameters, and is unaware of the external conditions [10]. Therefore, external monitoring systems could help to provide the control system with extra information about the process to make better control decisions. Furthermore, a new con3 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5007393 Preprint not peer reviewed trol system built on the top of the robot’s trajectories could determine if the layer is done correctly or with defects to recreate new trajectories to patch the melting issues [7]. In this scenario, sensors could be added to the WAAM system to collect data from the CMT configuration (electric signals), and a camera could detect bias in the layers [6]. To match both signals and correlate their data, the camera could be focused on the contact-tip-work-distance (CTWD), which is the distance between the contact tip of the welding torch (where the wire exits) and the surface of the workpiece being welded or additively manufactured [11]. This critical parameter influences the welding current, arc stability, and overall heat input. A shorter CTWD can increase the welding current and heat input, while a longer CTWD can reduce them, potentially leading to issues such as poor penetration or unstable arc conditions. Consequently, this work aims to develop a hybrid monitoring system that correlates the electrical signal of the welding machine with the torch behaviour (CTWD). Therefore, a two-channel oscilloscope is selected to gather electrical data (i.e., voltage and current pulses) and a welding camera to detect and measure the CTWD [12]. Then, an unsupervised machine-learning technique is applied to recognize the current pulse that is classified as an outlier or not afterwards. On the other hand, the welding camera took frames from the WAAM process, where an object detection model recognizes the CTWD and converts it into a wire measurement to correlate with the electrical signal and double-check the outlier [13]. As a main result, the unsupervised algorithm could detect outliers, and they were confirmed by the CTWD detected by YoloV9 nano, demonstrating a strong relation for further control system deployment. The rest of the manuscript is structured as follows: Section II shows the background and the literature. The methodology is presented in section III. The Results are shown in section IV, and their discussion is given in section V. Finally, conclusions and future works are described in section VI. 2. Background and literature review Given this scenario, several works have been presented data acquisition systems to improve the WAAM process. Consequently, Xiong et al [14]., presented an early work about adding a passive vision-sensing system to detect the bead height and width through the deposition of thin walls. Then, Li et al. demonstrated a thermal behaviour analysis based on finite ele4 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5007393 Preprint not peer reviewed ments analysis to define a suitable layer underpass time. Conversely, Yan et al [15]. designed one of the first approaches to using neural networks, especially in the additive manufacturing process. Following this trend, works such as [16, 9, 7, 12, 17] presented monitoring systems to collect data from the WAAM process and extract knowledge by using ML methods and improve the system performances. Furthermore, [6, 18] also added a flexible control system that enhances its performance by sensor information. Nowadays, works such as Holscher et al [19]., presented a novel CTWD detection in a WAAM process with a closed-loop layer height control. For its part, Novelino et al [11]. show how the CTWD influences the wall geometry to analyze the influence of the parameters in bead and multi-layer wall geometries fabricated by the CMT process to select the configurations that result in the best deposition control. Finally, Franke et al [20]., presented a vision-based process monitoring system, and image segmentation of the welding wire is used to monitor the working distance as well as the horizontal position of the wire during welding, and classic image processing techniques are applied to capture spatter formation. However, even when those works already detect and apply the CTWD in different control strategies, correlating with the electric signals that represent the welding machine configuration is still an open challenge. This information could provide insights into new control techniques and a new sensor configuration to improve the WAAM process. 3. Materials and Methods This section shows the test bench setup with all hardware specifications of each component. The following section details the welding machinery setup. Next, the current pulse detection is presented with its steps, starting from the raw data to detect outliers. Lastly, the CWTD detection is described, especially the image transformation into a distance measurement. 3.1. Test bench setup The WAAM cell comprises an ABB robot and a welding machine to make weld beads. As the monitoring system, one oscilloscope is placed near the welding machine to collect electrical parameters, such as the voltage and the current. Therefore, a voltmeter and amperometric clamp are needed. Furthermore, a welding camera is placed in front of the welding torch to capture frames of the CTWD. The Edge server synchronises the Robot’s trajectories with their axis rotations, the welding parameters configured, the 5 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5007393 Preprint not peer reviewed voltage and current from the oscilloscope, and takes frames from the welding camera. Then, raw data is stored in .csv files and frames in different folders. Also, the monitoring system merges all these .csv files with the frames’ names into a single HDF5 file for further analysis. All hardware components are described as follows: •ABB robot: It has a 6-axis movement where the welding torch is placed. The ABB robot publishes its trajectories and welding machine data from a TCP port as an array. •Fronius TP400i: It is a high-performance welding power source designed for demanding industrial applications. The TPS 400i supports MIG/MAG, Pulse, and TIG welding processes. It has a user-friendly interface and adaptive controls that optimize welding parameters in real-time, enhancing weld quality and reducing spatter. •Oscilloscope Pico2206b: This oscilloscope is a compact, high-performance oscilloscope for various electronic testing and diagnostic applications. This device offers an 8-bit resolution and a bandwidth of 50 MHz, making it suitable for capturing detailed waveforms in real-time. It features two input channels and a 1 GS/s sampling rate, ensuring accurate signal acquisition and analysis. •Cavitar Camera C400: The Cavitar C400 is a high-performance laser illumination system designed for challenging imaging applications in industrial environments. It delivers a monochromatic laser light, making it ideal for high-speed imaging and visualizing fast processes such as welding, combustion, and fluid dynamics. •Edge server: The Dell Precision 5860, equipped with an NVIDIA T1000 graphics card, is a robust workstation tailored for data monitoring systems and real-time control. It has a powerful Intel Xeon W-Series with 8 cores and 32 GB of RAM: 2 x 16 GB, DDR5. The NVIDIA T1000 GPU, with 4 GB of GDDR6 memory, enhances visual performance and supports multiple displays via four Mini DisplayPort connections. Figure 5 summarizes the test bench setup with all the hardware connections. 6 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5007393 Preprint not peer reviewed Figure 1: Test bench setup description of the hybrid anomaly detection in a WAAM cell 3.2. Welding Parameter Configurations Fronius TP400I is configured as follows: •Wire-speed: 3.5 m/min •Current: 131 Amps - real: 113 •Voltage: 13 V - real: 15.2 •Layer high: 1.8 mm •layer width: 4.4 mm •Configuration: CMT •Shielding Gas: Argon, 20 L/min •CTWD: 1.5 cm Conversely, the designed test involves welding single-meld beads to configure the camera and synchronize sensors. Then, a 16-layer wall with a waiting time between layers of 2 seconds is welded to validate that the outlier detection system is working properly. 3.3. Current pulse detection •Raw data: The data was gathered from the oscilloscope, which has a sample rate 10KHz for both the voltage and current signals. Since, in the CMT configuration, the voltage signal remains stable/constant; the current signal is the target of detecting outliers. 7 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5007393 Preprint not peer reviewed •Split down in layers: After electrical data was collected, it was broken down into layers to process each sample individually since the number of samples was vast. Also, the dataset contains dead zones where the torch was not working, and the oscilloscope saves empty values that might affect the following steps. Figure 2 (a) shows an example of the 16 layers-wall. •Take the first pulse out: The first electrical pulse is indeed anomalous since the wire creates the first arc with the substrate plate or melt bead. Therefore, it can make a unique cluster or confuse the clustering algorithm in the following step. Figure 2 (b) shows the first current pulse compared with the rest. •Clustering pulses: When each layer is individually stored and the first pulse is removed, the DBSCAN algorithm is applied. This nonparametric clustering algorithm groups points that are closely packed and marks as outliers points that lie alone in low-density regions. Consequently, 4 clusters were made; cluster 1 groups the valley points, middle point (cluster 2), peak points (cluster 3), and noise (points between cluster 1 and 2 ). Figure 2 (c) shows the clusters applied to the current signal. •Detect peaks and valleys: With cluster 3, the maximum point is detected on each current pulse and the valley points could be removed to determine outliers only on the current pulse without noise. •Select pulses: With the maximum-current-pulse-point, a moving leftto-right algorithm is developed to gather only the pulse without valleys. It means only using the points between the last valley point of the left of the current pulse and the first current point of the right valley points. •Detect outliers: When all pulses were taken solely, they were merged into single columns of the same size. If one pulse does not contain the same sample points, a moving average algorithm is applied to fill empty spaces. Then, the Isolation Forest algorithm is applied to the dataset to recognize outliers. Then, outlier current pulses return to their initial form attached to the valleys to get the exact length they were initially. In addition, pulses that have been assigned as normal ones are set to a zero in this array to match the original samples. Figure 2 (e) shows the outliers at the same time of the original samples. 8 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5007393 Preprint not peer reviewed (a) Current pulses of the 16 layers-wall (b) First current pulse compared with the rest (c) Clustering electric pulses (d) Outlier samples Figure 2: Current pulses in a WAAM process 3.4. CWTD detection •Raw frames: The Cavitar camera took 1440x1080 frames at 60 frames per second (FPS) from the WAAM process. The video was stored in a .wav file; then, the video was split into frames in .png format. Figure 3 (a) shows a frame example. •Annotate frames: The dataset was annotated with bounding boxes to highlight the CTWD location. This step is crucial to train object detection models. In the end, 150 images were annotated. •Pre-process data: Since the annotated image number could be a few samples to train ML models, pre-processing data is the following step; annotated images could be rotated, saturated, resized, flipped, and rescaled. As a result, the dataset increased in number and got extra information to detect the CTWD. 300 stretched images into 604x640 is the dataset where 70% is for the training set, 10% for the validation and 20% for the test. •Train object detection model: The object detection model selected was YOLOv9 since it presents a high accuracy with pre-trained weights. 9 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5007393 Preprint not peer reviewed [4] A. Waqas, Q. Xiansheng, X. Jiangtao, W. Hongbo, M. Muzamil, A. Majeed, Directional tensile properties of steel structure manufactured by robotic assisted GMAW additive manufacturing, in: 2019 16th International Bhurban Conference on Applied Sciences and Technology (IBCAST), pp. 239–242, ISSN: 2151-1411. doi:10.1109/IBCAST.2019. 8667262. URL https://ieeexplore.ieee.org/document/8667262 [5] Y. Li, H. Mu, J. Polden, H. Li, L. Wang, C. Xia, Z. Pan, Towards intelligent monitoring system in wire arc additive manufacturing: a surface anomaly detector on a small dataset 120 (7) 5225–5242. doi:10.1007/s00170-022-09076-5. URL https://doi.org/10.1007/s00170-022-09076-5 [6] H. Mu, J. Polden, Y. Li, F. He, C. Xia, Z. Pan, Layer-by-layer modelbased adaptive control for wire arc additive manufacturing of thin-wall structures 33 (4) 1165–1180. doi:10.1007/s10845-022-01920-5. URL https://doi.org/10.1007/s10845-022-01920-5 [7] A. ˇ Sˇcetinec, D. Klobˇcar, D. Braˇcun, In-process path replanning and online layer height control through deposition arc current for gas metal arc based additive manufacturing 64 1169–1179. doi:10.1016/j.jmapro.2021.02.038. URL https://www.sciencedirect.com/science/article/pii/ S1526612521001249 [8] J. Xiong, G. Zhang, Adaptive control of deposited height in GMAW-based layer additive manufacturing 214 (4) 962–968. doi:10.1016/j.jmatprotec.2013.11.014. URL https://www.sciencedirect.com/science/article/pii/ S0924013613003415 [9] N. Kozamernik, D. Braˇcun, D. Klobˇcar, WAAM system with interpass temperature control and forced cooling for near-net-shape printing of small metal components 110 (7) 1955–1968. doi:10.1007/ s00170-020-05958-8. URL https://doi.org/10.1007/s00170-020-05958-8 [10] M. Karmuhilan, A. k. sood, Intelligent process model for bead geometry prediction in WAAM 5 (11) 24005–24013. 16 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5007393 Preprint not peer reviewed doi:10.1016/j.matpr.2018.10.193. URL https://www.sciencedirect.com/science/article/pii/ S221478531832488X [11] A. Novelino, G. Carvalho, M. Ziberov, Influence of waamcmt deposition parameters on wall geometry, Advances in Industrial and Manufacturing Engineering 5 (2022) 100105. doi:https://doi.org/10.1016/j.aime.2022.100105. URL https://www.sciencedirect.com/science/article/pii/ S2666912922000320 [12] P. K. Nalajam, R. Varadarajan, A hybrid deep learning model for layer-wise melt pool temperature forecasting in wire-arc additive manufacturing process 9 100652–100664, conference Name: IEEE Access. doi:10.1109/ACCESS.2021.3097177. URL https://ieeexplore.ieee.org/abstract/document/9483890 [13] C. Xia, Z. Pan, Y. Li, J. Chen, H. Li, Vision-based melt pool monitoring for wire-arc additive manufacturing using deep learning method 120 (1) 551–562. doi:10.1007/s00170-022-08811-2. URL https://doi.org/10.1007/s00170-022-08811-2 [14] J. Xiong, G. Zhang, Online measurement of bead geometry in GMAWbased additive manufacturing using passive vision 24 (11) 115103, publisher: IOP Publishing. doi:10.1088/0957-0233/24/11/115103. URL https://dx.doi.org/10.1088/0957-0233/24/11/115103 [15] L. Yang, H. Sohn, Z. Ma, I. Jeon, P. Liu, J. C. Cheng, Realtime layer height estimation during multi-layer directed energy deposition using domain adaptive neural networks 148 103882. doi:10.1016/j.compind.2023.103882. URL https://linkinghub.elsevier.com/retrieve/pii/ S0166361523000325 [16] C. Halisch, T. Radel, D. Tyralla, T. Seefeld, Measuring the melt pool size in a wire arc additive manufacturing process using a high dynamic range two-colored pyrometric camera 64 (8) 1349–1356. doi:10.1007/ s40194-020-00892-5. URL https://doi.org/10.1007/s40194-020-00892-5 17 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5007393 Preprint not peer reviewed [17] C. Lee, G. Seo, D. B. Kim, M. Kim, J.-H. Shin, Development of defect detection AI model for wire + arc additive manufacturing using high dynamic range images 11 (16) 7541, number: 16 Publisher: Multidisciplinary Digital Publishing Institute. doi:10.3390/app11167541. URL https://www.mdpi.com/2076-3417/11/16/7541 [18] S. Adel, E. B. Njaastad, M. H. Arbo, F. Lindseth, Layer height estimation using laser line scanners for closed-loop directed energy deposition, in: 2023 11th International Conference on Control, Mechatronics and Automation (ICCMA), pp. 308–313, ISSN: 2837-5149. doi: 10.1109/ICCMA59762.2023.10374694. URL https://ieeexplore.ieee.org/document/10374694 [19] T. M. H. J. H¨olscher, Lennart Vincent Hassel, Detection of the contact tube to working distance in wire and arc additive manufacturing, The International Journal of Advanced Manufacturing Technology (2022). doi:10.1007/s00170-022-08805-0. URL https://doi.org/10.1007/s00170-022-08805-0 [20] H. F. . R. R. Franke, J., Vision based process monitoring in wire arc additive manufacturing (waam), Journal of Intelligent Manufacturing (2022). doi:https://doi.org/10.1007/s10845-023-02287-x. URL https://link-springer-com-443.webvpn.synu.edu.cn/ article/10.1007/s10845-023-02287-x#citeas 18 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5007393 Preprint not peer reviewed