Vision based process monitoring in wire arc additive manufacturing (WAAM)

Franke, Jan; Heinrich, Florian; Reisch, Raven T. Article — Published Version Vision based process monitoring in wire arc additive manufacturing (WAAM) Journal of Intelligent Manufacturing Provided in Cooperation with: Springer Nature Suggested Citation: Franke, Jan; Heinrich, Florian; Reisch, Raven T. (2024) : Vision based process monitoring in wire arc additive manufacturing (WAAM), Journal of Intelligent Manufacturing, ISSN 1572-8145, Springer US, New York, NY, Vol. 36, Iss. 3, pp. 1711-1721, https://doi.org/10.1007/s10845-023-02287-x This Version is available at: https://hdl.handle.net/10419/319241 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. http://creativecommons.org/licenses/by/4.0/ Journal of Intelligent Manufacturing (2025) 36:1711–1721 https://doi.org/10.1007/s10845-023-02287-x Vision based process monitoring in wire arc additive manufacturing (WAAM) Jan Franke1·Florian Heinrich1·Raven T. Reisch2,3 Received: 16 January 2023 / Accepted: 23 November 2023 / Published online: 1 March 2024 © The Author(s) 2024 Abstract Astableweldingprocessiscrucialtoobtainhighqualitypartsinwirearcadditivemanufacturing.Thecomplexityoftheprocess makes it inherently unstable, which can cause various defects, resulting in poor geometric accuracy and material properties. This demands for in-process monitoring and control mechanisms to industrialize the technology. In this work, process monitoring algorithms based on welding camera image analysis are presented. A neural network for semantic 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. Using these algorithms, the process stability is evaluated in real time and the analysis results enable the direction independent closed-loop-control of the manufacturing process. This significantly improves geometric fidelity as well as mechanical properties of the fabricated part and allows the automated production of parts with complex deposition paths including weld bead crossings, curvatures and overhang structures. Keywords Wire arc additive manufacturing ·Vision based monitoring ·Machine learning ·Nozzle-to-work distance monitoring ·Contact tube wear off detection ·Spatter detection Introduction Wirearcadditivemanufacturing(WAAM)isaprocessfor3D printing of large, near-net-shape metal parts layer-by-layer, using arc welding technologies. It offers significant time and cost advantages for various applications compared to conventional subtractive methods (Martina & Williams, 2015; Williams et al., 2016) and has been widely used with steel, aluminium, titanium, and nickel-based alloys (Rodrigues et Florian Heinrich and Raven T. Reisch have contributed equally to this work. BJan Franke [email protected] Florian Heinrich [email protected] Raven T. Reisch ra[email protected] 1FORWISS, University of Passau, Passau, Germany 2Technical University of Munich (TUM), Munich, Germany 3Siemens AG, Munich, Germany al., 2019). Currently most WAAM processes are based on Cold Metal Transfer (CMT), a modified Metal Inert Gas (MIG) welding process based on short-circuiting transfer (Selvi et al., 2018). Due to an oscillating wire, it allows a lower heat input in comparison to other welding technologies, and thus reduces residual stresses. Besides several advantages, WAAM is a multi-scale and multi-physics coupling process with complicated and unbalanced physical, chemical, thermal and metallurgical characteristics (Chen et al., 2021). The forming of a workpiece is intensely associated with the dynamic fluid characteristic of the melt pool, which is significantly influenced by several process parameters such as welding current and voltage, welding-speed, welding-angle or wire feed rate. Improper parametrization leads to a series of defects such as fluctuation effects, oxidation, pores, lack of fusion or slag inclusion (Hauser et al., 2020,2021a,2021b; Wu et al., 2018). Also, varying heat dissipation conditions may result in morphological changes of deposited beads (Li et al., 2021). Furthermore, for certain materials such as steel, the wires’ contact tube wears off over time, which could lead to material deposition defects or spattering. Heterogeneous error sources lead to various defects, 123 1712 Journal of Intelligent Manufacturing (2025) 36:1711–1721 negatively affecting both geometric accuracy and material properties. Morphological errors propagate to the next layer and accumulate over time. This may lead to a continuous deterioration of the overall geometric accuracy of the part, so that it eventually has to be rejected. In order to ensure high component quality and productivity, in-process monitoring and control strategies are indispensable. Process deviations are associated with the change of various signals, which can be used for online defect detection. Several sensor data evaluation algorithms were proposed in recent years. Tang et al. (2017) used an industrial camera behind the welding torch to capture weld bead images and trained a deep convolutional neural network (DCNN) combining it with a support vector machine (SVM) to classify weld bead images into five patterns as normal, pore, hump, depressionandundercut.Infraredthermographydatawasutilized by Chen et al. (2019) to identify weld bead deviation, hump and flow defects using a neural network. Reisch et al. (2020) developed a distance based anomaly detection based on the evaluation of current, voltage and welding camera images with a LSMT model, a convolution neural network and an autoencoder, respectively. The proposed approach showed the capability of detecting anomalies due to oxidation, polluted surfaces and form deviations. Lee et al. (2021) used a high dynamic range camera to extract features metal transfer and weld-pool that are used for an artificial intelligence model to classify normal and abnormal statuses of arc welding. Li (2021) developed two intelligent WAAM defect detection modules. The first module takes welding arc current and voltage signals during the deposition process as inputs and uses algorithms such as SVM and incremental SVM to identify disturbances and continuously learn new defects. The second module takes CCD images as inputs and uses object detection algorithms to predict the unfused defect during the WAAM manufacturing process. Although these studies contribute to the understanding of the complex process dynamics and provide methods for detecting typical defects, it is crucial to automatically compensate for weld bead deviations during production to industrialize the technology. Especially, fluctuations in the weld bead height are critical when accumulating over time. Since the welding torch is raised by a predefined layer height after each applied bead, large deviations in the distance between the nozzle of the welding torch and the workpiece, the nozzle-to-work (NtW) distance dNtW, can occur after several applied layers, which destabilizes the welding process.Alongdistance may generate bad gas protectiveeffects, resulting in porosities or bad formation of the layer. In contrast, a short distance can lead to a higher spatter rate and cause weld spatter to stick to the nozzle or even cause a collision between the welding nozzle and the workpiece (Xiong & Zhang, 2014). Thus, it is crucial to have a stable layer height resp. dNtW throughout the whole build process. In recent years, there has been increasing research into methods for measuring and controlling the layer height. Mostly a laser scan is used after each deposited layer to deduce height informations, which increases the inter-layer waiting time. Han et al. (2018) used the extracted features to control the layer height by adjusting the voltage, based on experience rules for a specific steel. The developed height controller was tested for a multilayer multi bead cuboid structure. Mu et al. (2022) compared the measurements against the CAD model. Geometric errors are then compensated and a new set of welding parameters for the next layer is created. Li et al. (2021) introduced an interlayer closed-loop-control (ICLC) algorithm for multi-layer multi-bead deposition of cuboid components with straight, parallel paths. Height errors are determined by comparing the measurements to the bead geometrymodel.Tangetal.(2021)establishedamulti-sensor system to monitor process parameter and proposed a method for weld bead modelling by means of a deep neural network. The feedback of the weld bead shape was received by means of an offline weld bead scan. Šˇcetinec et al. (2021) proposed a layer height controller based on the average arc current of the last deposited layer. Height deviations are compensated by replanning the tool path. Xiong et al. (2021) developed a closed-loop-control (CLC) of the deposition height by visual inspection of previous and current layer using a passive vision system. Classic image processing techniques are used to track the height. Deviationsareautomaticallycompensatedviacontrollingthe wire feed speed. The solution was tested for a single track unidirectional wall in one welding direction. Kissinger et al. (2019)and Hallam et al.(2022) used coherent range-resolved interferometry(CO-RRI)forlayerheightmeasurementwhen building straight walls made of steel. The sensor is mounted directly to the welding torch and measures the distance to the weld bead immediately behind the melt pool. There is no inherent sensitivity to the arc light and it is a cost effective and simple to setup solution. Mahfudianto et al. (2019) developed an artificial neural network with one hidden layer using several characteristic process parameter such as travel and feeding speed, current, voltage, etc. as input sources to estimate the distance of the wires’ contact tube to the work piece in real time. Hölscher et al. (2022) matched different process parameters to that distance. It has been shown that the electrical resistance during short circuit monitored the distance best for WAAM with and without using a CMT source. Both solutions verified their measurement techniques on single-track experiments by continuously modifying the weld distance. The above studies present different solutions for either measuring the weld bead height or determining the weld distance. Also, layer height control strategies are proposed, mostly validated by fabricating straight walls or cuboids. Mu et al. (2022) and Šˇcetinec et al. (2021) verified the 123 Journal of Intelligent Manufacturing (2025) 36:1711–1721 1713 adaptiveness of the designed control strategy by fabricating inclined and free-form structures. However, when printing complex geometries, material agglomerations for example at weld bead crossings could lead to sudden changes in the welddistance,severelyinfluencing theformingquality,when building up over time. If such short-term deviations are not properly accounted for in the control strategy, this can lead toinaccuratecompensation.Furthermore,optical systemsfor direct measurement of the weld bead dimensions behind the molten pool have an eccentric guidance and thus a directional dependency on the welding path. Reisgen et al. (2019) utilized visual monitoring of the processing zone to determine height informations in-situ, independently of process parameters, the selected material and the welding direction, by measuring the length of the visible wire (stickout), using no CMT welding technology. Workpiece and welding torch height control strategies were implemented based on an elevation map and showed the capability to compensate for surface errors and to control the distance between welding torch and workpiece. Reisch et al. (2021) investigated and evaluated the usage of several sensors for in-process measurement of the nozzle-to-work distance dNtW,usingaCMT welding source. It has been shown, that a welding camera is mostsuitable for process anddirection independent measurements. A CLC with in-situ deposition control was developed, where layer height deviations are compensated by adjusting the welding speed according to the evaluations of welding camera frames with a DCNN, predicting the visible length of the oscillating wire (stickout). Validation-parts showed significantly reduced weld bead deviations and thus a higher geometrical accuracy. Humping effects at weld bead crossings in particular could be compensated. Even a free-form object with incline structure was printed correctly on the first try without manual intervention, while the build up without deposition control had to be stopped after 20 deposited layers due to severe oxidation, misalignment and unstable process behaviour. This paper presents image based WAAM process monitoring algorithms for the CMT welding technology. The goal is to obtain information about the process stability in real time. Monitoring of the dNtW is realized by measuring the wire stickout. Therefore, the latter procedure using a DCNN, is extended to a wire segmentation. The additional semantic information makes the procedure more robust and not only the wire length but also the horizontal position can be extracted, which is used to identify deflection of the wire. Deflection occurs if the wire contact tube is worn off or dNtW is too large. This could lead to unstable arc behaviour, increased spattering or weld seam irregularities such as lack of fusion (Henckell et al., 2020) and the resulting offset affects the software supported path planning with the risk of dimensional deviations from the CAD model and the work piece (Zhan et al., 2017). Spattering indicates instabilities in the welding process for various reasons, like improper feedstock, incorrect welding parameters, insufficient shielding gas or an incorrect dNtW. This may result in reduced surface quality, strength, durability and functionality (Serrati et al., 2023). To capture this phenomenon, a novel spatter monitoring algorithm, based on detecting changes between consecutive frames, is presented. The collected spatter statistic is a valuable data-source for ex-situ tests of the fabricated part and optimization strategies concerning design and deposition paths. Thereminder ofthis paper isorganizedasfollows:Section “Vision system” describes the experimental setup. WAAMand monitoring-components as well as a semi-automatic system calibration method are introduced. Section “Wire monitoring” focuses on the deduction of the dNtW from wire stickout and discusses in detail the developed machine learning solution for wire monitoring. A semantic wire segmentation and a classification of the results as well as a following feature extraction is proposed. Section “Spatter monitoring”presents atwo-stagealgorithm torobustlydetect spatterformation in welding camera frames. Finally,the findings of this paper are summarized in section “Conclusion”. Vision system Experimental setup Experiments and validation tests were conducted on a robotbased WAAM setup with Fronius TPS CMT 4000 advanced system as welding source which is capable of using CMT wire feed technology. The welding torch was attached to the head of a calibrated six axis robot (COMAU NJ130 2.0), see Fig. 1. An AlSi12 wire ( 1.2 mm ) and Argon as inert gas at a gas flow rate of 10 Lmin−1in an air-conditioned room was used. Camera To visually monitor the processing zone, a welding camera (Cavitar C300) was mounted to the robot head, see Fig. 2.It enables the visualization of build-part and welding process at the same time, using active laser-illumination. It provides a 8-bit grayscale video stream and was operated at a sample rate of 20 Hz with a resolution of 960 ×740 pixels. Corner detection on a checkerboard pattern was used to calibrate the welding camera to the processing zone, see Fig. 3. This automatically compensates for the monitoring angle. 123 1714 Journal of Intelligent Manufacturing (2025) 36:1711–1721 Fig. 1 WAAM system overview: CMT welding source, robot with attached welding torch and substrate table Fig. 2 Welding-camera at a working distance of 200 mm and an angle of 10◦to the horizontal Wire monitoring Welding stability for the CMT process with respect to uniform material deposition is monitored by observing the oscillating wire. The perspective on the processing zone is fixed throughout the build process and process related, the oscillating wire dips into the weld bead. Then, the stickout lwcorresponds to the dNtW when added-off a known hidden portion lh. This enables the measurement of the dNtW by measuring the stickout lw, see Fig. 4. From a classic image processing point of view, measuring the stickout is difficult to achieve due to permanently Fig. 3 Camera calibration: a 3 mm ×3 mm checkerboard pattern at working distance and the calibration results (blue) - Detected inner checkerboard corners Fig. 4 Welding camera frame. Weld distance dNtW composed of stickout lwand hidden portion lh changing specular reflections and thus badly posed, see Fig. 5. To overcome the problem of optical measurement in an environment with highly fluctuating lighting conditions, a data-driven approach was adopted. To robustly extract wire information in real time, a three stage architecture is proposed. First, a neural network for semantic segmentation localizes the wire, followed by a classification that checks the integrity of the prediction, validating it for the final feature extraction. Wire segmentation Semantic segmentation is used to locate objects in an image. The goal is to label each pixel of an image with a corresponding class of what is being represented. This can be 123 Journal of Intelligent Manufacturing (2025) 36:1711–1721 1715 Fig. 5 Processingzoneofa test-print atthreedifferenttimesduringproduction. Motion and continuously changing lighting conditions cause unpredictable specular reflections from metal surfaces. It is not easy, even for the human eye, to see the stickout achieved using the well known U-net architecture, originally developed for biomedical image processing by Ronneberg et al. (2015). The proposed architecture was modified to detect and locate the visible wire, utilizing a feature vector v=4,8,16,32,64, a single output class (wire), zero padding and additional dropout layers with 25% dropout, as modifications, see the left side of Fig. 6. It consists of a downsampling part (left) and a symmetric upsampling part (right). A downsampling step consists of a back to back execution of two convolutions, a rectified linear unit (ReLU), max pooling for downsampling and dropout to prevent overfitting and improve the generalization error. By repeating this, more compact and increasingly abstract representations (feature maps) of the input image are created, according to the feature vector. The network learns to detect a wire using these detailed feature maps. As a result, the information about where the wire is located is lost. The spatial information is recovered by upsampling, i.e. by converting the low-resolution representations into a high-resolution image. Every up step consists of a transposed convolution thathalvesthe number and doubles the dimensions of thefeature maps, dropout, a concatenation with the corresponding feature maps from the downsampling part and two convolutions, each followed by a ReLU. The corresponding feature maps from the downsampling step help to localize the features more precisely during upsampling. At the final layer a 1x1 convolution is used to map each 4-component feature vector to one class. The architecture returns a full resolution prediction of a wire. That means, each pixel is assigned a probability, called precision score, of belonging to a wire. The semantic segmentation model is trained using segmentation maps. Therefore, welding camera videos of various test-prints served as data source. Interesting subsets from different printing stages were annotated using a semi-automatic annotation tool. In total, 3000 frames were labelled, where 20% of them were set aside for validation. The network was trained on cut-outs of dimensions 256x512 pixels which contain the immediate processing zone, together with their corresponding segmentation maps with value 1 inside an enclosing quadrangle for the visible wire and 0 elsewhere, see Fig. 7. σ>ρ TRUEFALSE break Decision thresholding: τ Segmentation Input: Image 512x256 Output:Probability P 512x256 14 4 84 4 1 256x128 8 8 16 8 128x64 16 16 32 16 64x32 32 32 64 32 32x16 64 conv 1x1 copy conv 3x3, ReLU max pool 2x2, dropout trans-conv 2x2, dropout Classification Input: Probability P Histogram analysis of P Output: Binary wire Fig. 6 The proposed architecture for computing a valid wire representation for feature extraction with segmentation network and subsequent classification 123 1716 Journal of Intelligent Manufacturing (2025) 36:1711–1721 Fig. 7 Training data: Sample datum (left) and corresponding annotated segmentation map (right) Fig. 8 Learning curve of the proposed U-Net implementation ThenetworkwasimplementedwiththeTensorflowframework1using its Keras API and trained for 80 epochs with 32 samples per iteration (batch size), binary cross-entropy loss and default RMSprop optimizer achieving a training and validation accuracy of ∼99.5%. Figure 8shows the learning curve. Classification The quality of a specific wire prediction is determined by the accuracy of the network and the actual visual conditions present. The data situation can occasionally be poor, leading thenetworkto providewire predictions containing areas with low confidence, whereas in other regions, the confidence is high. To decide if the information content of a prediction is suitable for feature extraction, it is classified into one of three categoriesbymeasuringitscertainty.Therefore,theprobability distribution of the prediction is considered to construct a confidence marker. It is assumed that precision scores below 0.1 are irrelevant for the determination of a wire instance. To measure the certainty of the prediction, a dichotomy of 1https://www.tensorflow.org/. the relevant precision scores is performed, using the standard decision threshold for semantic segmentation τ=0.5. A value above τis classified as high, otherwise as low. High certainty of a prediction is indicated by a high proportion of high precision scores. For a prediction P, consider the sets M={P(x,y)|P(x,y)≥0.1}and (1) N={P(x,y)|P(x,y)>τ}.(2) Then σ=|N| |M|is used as a confidence marker and Pis validated by setting a suitable threshold ρfor σ.If σ>ρ, (3) then the corresponding prediction Pis considered to be of high certainty and a binary wire representation is computed by thresholding Pusing τ. This results either in a fragmented or a full wire representation, as depicted within the yellow and green rectangle in Fig. 6.IfEq.3does not hold, this indicates that the corresponding prediction Pis of low certainty and thresholding Presults in an unsuitable wire representation for feature extraction, see the red rectangle in Fig. 6. Images corresponding to invalid classified wires served as data source for re-teaching of the model. In this way, the network became more robust with respect to the permanently changing specular reflections from the wire. Feature extraction The stickout is determined by the lowest detected pixel in a binary wire returned by the classification. Figure 9shows the monitoring result for the stickouts of a specific build job part with adverse visual conditions along with σ, identifying phases during fabrication, where the feature extraction should be suspended. The benefit of this approach is that, compared to the DCNNregressionmodel used inReisch et al. (2021),nowthe prediction results are explainable. This breaks up the black box characteristic of the regression model and only suitable predictions are used to deduce the stickout for process control. Figure 10 illustrates the problem of using the regression model to predict the stickout. Under adverse visual conditions, it may return plausible values, but they are not correct. Now that a prediction is comprehensible, not only the stickout can be robustly determined, but also the center of gravity and thus the position of the wire. It is deduced from non-fragmented wire representations, which is ensured by an additional connected component analysis (CCA). Figure 11 shows three valid wire representations of the same build job and the derived features, superimposed on the corresponding welding camera frames. The wire position is used to track the wear off of the contact tube by measuring deviations in its horizontal com123 Journal of Intelligent Manufacturing (2025) 36:1711–1721 1717 Fig. 9 Confidence marker σ(red) and stickouts (blue), deduced with the proposed approach. The proportion of high confidence values drops when entering adverse visual conditions at the end of the observation, identified using a confidence threshold ρ=0.75, indicated by a horizontal line (Color figure online) Fig. 10 Stickouts predicted with the regression model and an underestimated stickout for a wire, indicated by a horizontal line ponent. Wear off causes the wire to lose guidance over time. This leads to increased deposition defects, which becomes unacceptable at a certain point during fabrication. Then, the process must be stopped to exchange the wear part. The life cycle of that specific wear part can be determined adaptively in-process independent of a specific build job. Figure 12 shows two contact tube examples with the wire guide worn to different degrees. To monitor the positional offset of the wire and consequently the wear state of the contact tube, the median of the positional deviations from the original is calculated for every 200 frames along the processing. Figure 13 shows the monitoring result of a 45 minute build job. This approach was tested with a single camera on industrial WAAM data as stated in section “Vision system”. In general, it is a priori not clear in which direction the wear off will evolve. In the worst case, it evolves in the view direction of the monitoring system and thus cannot be detected. Fig. 11 Robust mapping of the wire. Monitoring results (blue) and extracted features: marker for end of wire (red) and horizontal barycenter (green) (Color figure online) Fig. 12 Wire contact tubes, 50% worn off (left) and 100% (right) Fig. 13 Deflection detection. The median of positional offsets, measured for every 200 images, increases over time In most cases, a suitable positioning of the camera is sufficient. For an optimal solution, a second camera could be used, mounted horizontal perpendicular, to ensure the capturing of positional deviations. However, this is accompanied by higher costs. 123 1718 Journal of Intelligent Manufacturing (2025) 36:1711–1721 Fig. 14 Spatter detection: image processing pipeline. The final result contains only potential spatter components Computational aspect The evaluation of a welding camera frame, using the proposed method, takes 40 ms on an Intel Core i7-10700K CPU and 25 ms on the Nvidia GeForce RTX 3070 GPU, with a setup as described in section “Vision system”. The real time monitoring of the welding process could be realized and the developed wire monitoring system enables the in-situ control of the weld bead height, when using the CLC for the dNtW designed by Reisch et al. (2021). Spatter monitoring Spatters in welding might indicate process instabilities. In the following, an automated spatter detection and quantification is described. Monitoring spatter formation does not face the same image processing challenges like monitoring the wire. Spatter components have similar shapes and brightness and thus are, as objects, easier to detect. But, specular reflections can have the same visual characteristics may be detected falsely as spatter. These are identified by determining changes between consecutive frames. The algorithm consists of two stages, spatter detection and a refinement stage to remove false positives. Spatter detection Spatter appears as small sized bright spots, is a temporary phenomenon and distributes beyond the immediate processingzone.Thusthefullframesareinspected.Thevideostream is a function in time with discrete arguments. Therefore, the first step is to compute the finite difference between consecutive frames for removing background and low intensity areas. Next, clusters of specular reflections, mostly residing around the immediate processing zone are identified using morphological closing. Finally, a CCA keeps only areas representing potential spatter. The image processing pipeline for a frame, togetherwiththecorrespondingintermediateresultsdepicted for a cut-out (blue) for visualization reasons, is shown in Fig. 14 . Background subtraction:Foraframe fi,i=0,1, ... consider its successor fi+1and the pixel-wise finite difference fD=fi+1−fi.(4) Using a suitable threshold T>0, this yields binary images f+resp. f−containing high intensity areas only visible in fi+1resp. fi, defined by: f+(x,y)=1,fD(x,y)>T 0,else (5) f−(x,y)=1,fD(x,y)<−T 0,else .(6) Spatter detection must not be performed twice for fi, since theinformationabout potential spatterisalreadyknownfrom the previous iteration, except if i=0. In the latter case, the rest of the spatter detection pipeline is performed for both f+and f−, else it is performed only for f+. The algorithm is insensitive to the choice of T. Any bright objects left after the background subtraction that do not represent spatter are removed by the following processing steps. Clustering: Assuming that individual spatter instances have a minimum distance dmin to each other, cluster of specular reflections, mostly residing on weld bead and melt pool, can beidentifiedusingmorphologicalclosing.Thisconnectshigh intensity areas which are close together, with respect to dmin and leaves the rest. This results in the morphological closure C(f+)of the binary image f+. Spatter identification: Non-spatter-like areas are finally removed considering their morphology and size, using a CCA. This yields a binary image S(f+)containing only spatter-like components visible in fi+1. 123