Method and a system for real-time monitoring of additive manufacturing (AM) building process of a three-dimensional (3D) product generated by a 3D building device

Abstract

This invention relates to a method and system for real-time monitoring of the additive manufacturing (AM) building process of a three-dimensional (3D) product generated by a 3D building device. The building process comprises alternately distributing a material on a substrate by a coating mechanism and forming a layer by fusing a portion of the material according to a planned geometry. An imaging device captures image data for each build layer, including a pair of images representing the distributed material prior to layer formation and the corresponding build layer after formation. A processor processes the image data by performing a localized-image evaluation to detect local surface defects before and after layer formation and determine a localized-image failure metric, and / or by performing a whole-image evaluation using statistical learning models to determine a whole-image failure metric. Based on the comparison of the failure metrics with reference values, each layer is categorized as critical or non-critical. The method further includes determining a forward-looking failure risk indicator using the evaluation results and a defined number of most recent, to predict whether the building process is trending toward failure.

Description

[0001] METHOD AND A SYSTEM FOR REAL-TIME MONITORING OF ADDITIVE MANUFACTURING (AM) BUILDING PROCESS OF A THREE-DIMENSIONAL (3D) PRODUCT GENERATED BY A 3D BUILDING DEVICE

[0002] FIELD OF THE INVENTION

[0003] The present invention relates to a method and a system for real-time monitoring of additive manufacturing building process of a three-dimensional (3D) product generated by a 3D building device, where the building process comprises alternatively distributing a material on a substrate by a coating mechanism and subsequently building a layer by means of fusing a portion of the material by a fusing device according to a planned geometry of the 3D product.

[0004] BACKGROUND OF THE INVENTION

[0005] Additive manufacturing (AM) is the general term for those technologies that successively join material in an additive manner to create physical objects as specified by 3D model data. AM has paved its way during the last decades, offering significantly more design freedom compared to subtractive methods. These technologies are presently used for various applications in the engineering industry such as the aerospace, automotive industry as well as other areas of society, such as medicine, education, architecture, toys and entertainment.

[0006] Metal AM has disruptive potential within several industries. In particular the aviation industry, where multi component assemblies can be combined into a single part. Industrial companies have documented 20:1 component reduction which results in manufacturing savings of up to $3M per specific component per airplane. In addition, they document 25% fuel efficiency and 25% lightweighting of components enabled by the technology. This has therefore significant financial and environmental impact.

[0007] Despite these benefits, the lack of quality assurance solutions is hindering widespread adoption of the technology. As these are heavily regulated industries, the need for a process and quality monitoring solution is substantial. Traditional measurement methods simply do not apply due to the complex geometrical freedom offered.

[0008] There are seven types of AM processes defined by the ISO / ASTM 52900:2021 standard, one of which is Powder Bed Fusion (PBF). Within this category, a common method is Laser Powder Bed Fusion (LPBF). A typical LPBF printer uses a high-power laser at a very small point, to melt powder which is spread out by a powder distribution / coating mechanism. The laser profile is drawn on the flat powder surface using a XY galvanometer scanning mirror system and necessary optics to focus the laser beam uniformly across the build plane. This process continues in a layer wise manner until the full component has been created, and commonly takes several days up to weeks, depending on the volume of the physical object or the amount of material to be fused, to print the physical object.

[0009] Existing solutions for monitoring of AM processes are mainly focused on acquiring images from a camera installed on the printer, with an unobstructed view of the build area. The images are then stored in a folder on a dedicated computer where an operator can review the images captured. Commonly, two images are captured for each layer that is produced. First an image of the powder spread is saved before the laser starts its melting procedure. After the layer has finished melting the layer then a second image is taken. These are stored in separate folders, labeled with the corresponding layer number.

[0010] Several types of defects may occur during the AM process, such as protruding part defects, which occur when a solidified part warps due to thermal stresses and as a result protrudes above the surface of the powder bed. Such a defect can cause a collision between the coater (powder spreader) and the part which can hinder successful redistribution of powder and damage the coater itself or the mechanical assembly. This can cause significant downtime for the printer and costs associated with servicing the printer.

[0011] Other types of defects include spattering defect, spreading defect, hole defect, porosity etc.. The origin of the defect can stem from various factors such as poor climate conditions (thermal, inert gas flow (argon / nitrogen) etc.), faulty hardware components and poor selection of process parameters. A considerable number of defects can also stem from a poor geometry design of the part to be printed, such as lack of support structures. These defects can lead to non-conformance of the printed part.

[0012] Quality assurance and control in AM is challenging due to the process's inherent complexity, involving layer-by-layer construction that introduces variability. The unpredictability of material properties, coupled with high precision requirements, adds to the difficulty.

[0013] In-process monitoring is limited, as products are often inaccessible until completion, making real-time quality checks challenging. Post-processing steps can also introduce variability, and the lack of standardized procedures in this relatively new field further complicates consistent quality control. AM's sensitivity to machine-specific parameters, like temperature and printing speed, and the unpredictability of complex designs, pose additional hurdles. Moreover, analyzing the vast data generated during printing requires advanced software and expertise, while a skills gap in the evolving workforce can lead to production inconsistencies, underlining the need for enhanced training and technology development in AM. The postprocess chain is long and can include several processes such as thermal annealing, CNC machining, EDM machining and polishing. This can take days and involve several experts per stage. Unfortunately, the defects are often not visible until the very end of the processing chain (i.e. after polishing). In many cases these defects could have been spotted in the LPBF stage and substantial time, resources and money saved.

[0014] More importantly, in severe cases the defects may even not be identified at all, which may cause risk of more fragile objects than expected, which obviously may cause high risk when the object is critical and is e.g. used in the aviation industry.

[0015] US2023 / 0294173 discloses a method for additive manufacturing method using statistical learning model for quality control. Each layer is created by generating a material surface formed from a particulate (powder) material. An image is taken of the particulate material surface material and fed into a previously trained statistical learning model to determine if defect is present or not. If no defect is detected, the layer is solidified and the process of making a subsequent layer of particulate material is repeated. If, however, a defect is identified to be present, the layer of particulate material is reworked or reproduced, and the process is subsequently repeated for this particulate layer.

[0016] This approach presents several technical drawbacks. One major limitation is the narrow defect detection window: because the decision-making process evaluates only the current powder layer, defects are not identified until they become visibly significant. By that point, the defect may be too advanced to correct or to salvage the build.

[0017] Additionally, the inspection is restricted to imaging and analyzing the powder layer before solidification. This poses two key issues: (1) it captures less image data as no image is taken after solidification of the layer, and (2) certain defects only become apparent after solidification, rendering them undetectable with this technique.

[0018] Moreover, since the method only detects defects in the current unsolidified layer and triggers intervention only when these defects become substantial, there is a high risk that corrective action comes too late to be effective. SUMMARY OF THE INVENTION

[0019] It is an object of the invention to overcome the above-mentioned problem by use of real-time monitoring for enabling early stopping or intervention of a build that is likely to fail, which saves a significant amount of down time. This is of a particular relevance when printing multiple of separate parts in the same build because a single misprinted / defective component can affect the others around it and compromise the entire build. Early detection of this occurring thus saves the rest of the entire build if e.g. a user is alerted. The user could then e.g. manually delete the defective part from the building instructions, or the defective part could be automatically deleted, thus saving the rest of the build.

[0020] For the purposes of the present description, the terms 'building a layer' and 'printing a layer' are to be construed as referring to the same process step within the layer-by-layer additive manufacturing sequence, specifically within laser powder bed fusion (LPBF) and similar technologies. Both terms describe the deposition and solidification of a single cross-sectional layer of the digital build model, wherein powdered material is selectively fused, typically by a laser or electron beam according to the geometric pattern defined for that particular layer. Although the terminology may vary depending on user preference or industry standards, no distinction is intended between the two expressions in the context of the present invention.

[0021] In general, the invention preferably seeks to mitigate, alleviate or eliminate one or more of the above-mentioned disadvantages of the prior art singly or in any combination. In particular, it may be seen as an object of embodiments of the present invention to provide a method and a system that solves the above-mentioned problems, or other problems.

[0022] To better address one or more of these concerns, in a first aspect of the invention a method is provided for real-time monitoring of additive manufacturing (AM) building process of a three dimensional (3D) product generated by a 3D building device, where the building process comprises alternatively distributing a material on a substrate by a coating mechanism and subsequently building a layer by means of fusing a portion of the material by a fusing device according to a planned geometry of the 3D product.

[0023] The method comprises obtaining, from an imaging device, image data for each build layer while building the 3D product, and processing, by a processor, the image data for each of said build layer. The image data for each build layer while building the 3D product comprises a pair of images including image data representing the distributed material before building a layer, and image data representing the corresponding build layer after formation from the distributed material resulting in a build layer.

[0024] The step of processing the image data of each of said build layers comprises:

[0025] • performing a localized-image-evaluation of the image data, by means of: o detecting local surface-defects in the distributed material before building the layer, o detecting local surface-defects in the resulting build layer after building the layer from the distributed material, and based thereon o determining a localized-image failure metric for the localized-image- evaluation, and / or

[0026] • performing a whole-image evaluation of the image data, by means: o comparing the whole image data of distributed material before building the layer with previously trained statistical learning model, and o comparing the whole image data after building the layer with previously trained statistical learning model, and based thereon o determining a whole-image failure metric for the whole-image evaluation.

[0027] If the determined failure metrics are below reference failure metric(s) the build layer is categorized as non-critical, and the method further comprises the steps of:

[0028] • utilizing the results of said localized-image-evaluation and / or the whole-image evaluation, as well as the result together with the most recent n number of build layers, as input into a prediction model where a forward-looking failure risk indicator is determined indicating probability that the building process of the 3D product is trending toward failure or not.

[0029] A key technical advantage of the present invention lies in its ability to leverage cumulative data across multiple build layers to enhance the predictive reliability of process trend analysis. Specifically, when the failure metrics determined for a given layer fall below one or more predefined reference thresholds, the layer is classified as non-critical. In such cases, the method does not discard the associated evaluation data; instead, it utilizes the localized and / or whole-image analysis results from the current layer in combination with those of the most recent n preceding layers. This aggregated dataset is then input into a forward-looking prediction model.

[0030] The most recent n number of build layers may as an example include layers that are categorized as non-critical, or a combination of non-critical and critical layers.

[0031] In one embodiment, the number n may be a fixed positive integer. In another embodiment, n may be a continuously increasing number until a subsequent n-th layer is categorized as critical. In yet another embodiment, if layer n is categorized as critical, the layer count may continue until the fixed positive integer value is reached, after which the counting resets to zero.

[0032] By incorporating historical into the failure risk evaluation, the method achieves improved sensitivity to gradual process degradation trends that may not be evident from a single-layer analysis. This multi-layer, data-accumulative approach enhances the accuracy of forecasting potential future build failures and enables earlier and more reliable intervention decisions, reduces the risk of catastrophic build termination, and minimizes wasted machine time and material, thereby contributing to more efficient and economically viable additive manufacturing operations.

[0033] The method according to the present invention thus allows automatically monitoring builds in real time using e.g. Al and computer vision to detect anomalies, predict failures, and classify issues as they arise. This not only reduces the reliance on manual inspection but also ensures faster response times, minimizing waste and maximizing productivity.

[0034] In an embodiment, the step performing the localized-image-evaluation of the image data and / or the whole-image evaluation of the image data comprises:

[0035] • processing the image data together with reference data, where the reference data comprises reference image data of a distributed material before building the layer and reference image data after building a layer.

[0036] In an embodiment, the method further comprises utilizing the result of the processing in determining a statistical matching likelihood score between the image data and the reference data, where said localized-image failure metric and / or said whole-image failure metric are determined based on the statistical matching likelihood score. By processing the image data together with reference image data captured both before and after building a layer, the method enables a more robust and context-aware evaluation of the distributed material. This dual-reference approach enhances defect detection accuracy by providing temporal and spatial baselines for comparison.

[0037] Furthermore, calculating a statistical matching likelihood score allows for a quantitative assessment of deviations, which improves the objectivity and reliability of the localized and / or whole-image failure metrics. As a result, the method facilitates earlier and more precise identification of potential manufacturing defects, thereby improving overall process quality and reducing the risk of downstream errors.

[0038] In a preferred embodiment, the results of said localized-image-evaluation and said wholeimage evaluation are both utilized as input into the prediction model for determining the forward-looking failure risk indicator.

[0039] In an alternative embodiment, if one or more of the determined failure metrics exceed said reference metric(s), or if the determined probability indicates that the building process of the 3D product is trending toward failure, an electrical command, such as an alert signal, is triggered indicating that subsequent build layers are under potential failure risk. In that way, an operator may immediately act on it and prevent the whole build from being damaged.

[0040] By utilizing both the results of the localized-image-evaluation and the whole-image evaluation as input to the prediction model, the method leverages complementary insights, fine-grained local anomalies and broader global patterns thereby enhancing the predictive accuracy of the forward-looking failure risk indicator. This dual-level input approach improves the model’s ability to detect subtle but critical failure precursors, allowing for more reliable early warnings and preventive actions. Consequently, it supports more proactive quality assurance and reduces the likelihood of undetected defects propagating through subsequent production stages.

[0041] In collaboration with an external partner, the method according to the present invention successfully deployed its real time process monitoring platform across four LPBF systems. Over 54 builds, totaling approximately 47 days of print time, where the method or method platform continuously monitored and classified build quality in real time, flagging build layer statuses as OK, minor issues, or critical issues. The results revealed that 27.8% of total print time was wasted on failed builds. However, had the prints been halted at the point of critical failure detection by the method according to the present invention, 77% of total wasted time - equivalent to roughly 10 days - could have been saved. That translates to $115k machine operating cost savings per year. If material and pre- and post-processing costs are included this number becomes significantly higher.

[0042] Accordingly, a reliable solution is provided that enables intervention in a potentially failed build process, thereby significantly reducing downtime. Early intervention may include stopping the build entirely or dynamically adjusting one or more printing parameters to mitigate the issue before failure occurs. The step of issuing the alert signal may as an example be done in real-time, e.g. via text message, email, automatic phone call etc..

[0043] Moreover, it is now possible to identify defects during the build that might otherwise not be identified at all, which in turn could therefore result in a defected 3D product which may have severe consequences when such a 3D build is utilized in various applications.

[0044] Also, in the case of a successful build, the present invention enables issuing data showing that the build was successful.

[0045] The reference data may as an example include original raw reference data from previous image data and / or data derived therefrom from previous builds. The reference data may continuously be extended / updated / improved while the building process takes place. It should be noted that the reference data may be collected from multiple printers and users and may include prints produced with a variety of powder types. Incorporating such diverse data into the reference set is preferred for capturing process variability and material diversity for increasing the robustness and stability of the solution.

[0046] The statistical matching likelihood score may be understood as, but not limited to, as a classification of the status of the current build layer, for example but not limited to, "support structure", “ok”, “potential issue”, “minor issue”, “critical issue”. A classification model can as an example determine the confidence score from e.g. 0 to 1 for each layer status type and the overall status assigned to the current layer being the one with the highest confidence score.

[0047] Said step of calculating the potential failure metric can be obtained with sequential data models such as, but not limited to, recurrent neural networks (RNN). RNN is a deep learning model capable of evaluating time-series data, in this case the failure likelihood score based on the identified potential defects and layer status analyzed from the image data for each build layer. A RNN model takes sequential data such as time-series data as input and can output a specific data output. The model can be trained to output a failure likelihood score for a future build layer (for a certain number of layers after the current layer) based on the failure likelihood of a certain number of previous layers. Other layer results can also be used as input for the sequential model, such as, but not limited to, raw or processed image data, defect type and number, size and position of each defect type detected per layer.

[0048] As already mentioned, the term build or building process may, according to the present invention, also be understood as print or printing process. Also, the distributed material on the substrate may in one embodiment comprise a metal powder and the fusing device may comprise a laser, and where the 3D product is a metal product. The distributed material may also be of any other type, e.g. plastic material where the 3D product is a plastic product.

[0049] In an embodiment, the processing further comprises identifying potential defects for each build layer, where the potential defect has one or more different characteristic property, wherein the step of determining the failure metric for the building process is further based on using the identified potential defects. The step of using the identified potential defects and the statistical matching likelihood scores in calculating the potential failure metric may in an embodiment comprise calculating a statistical likelihood of a critical failure for the forthcoming build layers.

[0050] Said step of identifying potential defects may in one embodiment comprise comparing the characteristic properties of the identified potential defects with characteristic properties of previously detected defects obtained from previously built layers of said 3D product and / or from previous 3D products.

[0051] The identified potential defects may in an embodiment further comprise grouping the identified potential defects into different characteristic property groups, each group containing potential defects of similar or identical nature.

[0052] In an embodiment, the method further comprises calculating the number of potential defects within each group, wherein the calculated number is configured to be used in presenting in real time a frequency plot for different types of potential defects. The step of calculating may in an embodiment comprise determining a number of potential defect matches and / or calculating areal of the potential defects and utilize the areal in converting it into representation of the number of potential defects.

[0053] In addition to the frequency plot of the build layers, the images for each build layer within the frequency plot may additionally be shown where the potential defects are highlighted via e.g. colored markups, e.g. colored box, circle or other shape, to visually point to where the potential defect is located for each build layer. As an example, the user may scroll within the frequency plot and by doing so, the build layer where the user is scrolling through is shown by showing the image data pair for the build layer together with said colored markups. If the user selects, via a selection command, a given defect group, the correspondingcolored markup is hidden in the displayed image pair. The user can later select the colored markup again if a further evaluation of the potential defect group is required.

[0054] Accordingly, by identifying the potential defects the reliability of enabling early stopping of a potentially failed build may be further improved.

[0055] The potential defects are selected from, but are not limited to:

[0056] • a protruding fused material defect occurring due to warpage of the fused material, which may be due to excessive heat buildup or overhangs that lead to part deformation, where severity of protrusion may also be grouped into several defect groups,

[0057] • a spatter defect occurring when laser melt pool conditions are poor causing unwanted particles to be formed such as due to oxidized powder or melt pool ejections,

[0058] • a streaking defect occurring when the coating mechanism is defective causing insufficient powder distribution during coating or due to lose solidified particles being dragged along with the coating mechanism, i.e. the coating mechanism leaves streaks, causing uneven powder distribution,

[0059] • a powder hole or cavity caused due to insufficient distribution of powder,

[0060] • a hopping or ringing defect occurring during powder distribution process where the coating mechanism may collide with a protruding part causing e.g. a mechanical ringing or vibration resulting from collision, • a smoke defect caused by suboptimal build chamber environment parameters, e.g. gas composition or gas flow, where excessive spatter or overheating can cause smoke,

[0061] • a burn defect caused when the fusing device, subsequent to distribution of the material, strikes already fused material,

[0062] • a porosity defect caused by voids or holes inside fused portions or layers.

[0063] The porosity in AM, often referred to as 3D printing, relates to the presence of small voids or holes within the manufactured 3D product which may be a metal part. The voids in the porosity defect may be detrimental to the structural integrity and mechanical properties of the component, affecting its density, strength, ductility, and fatigue life.

[0064] Accordingly, identifying said defects such as the porosity, preferably in real time, is important for applications where the mechanical properties and reliability of the metal parts are critical, such as in aerospace, automotive, and medical devices.

[0065] In an embodiment, the method further comprises receiving an input command from a user via a user interface, and displaying, in response to the input command, one or more of the different types of potential defects or regions of interest. The received input command may be done via touch button command from the user, or via mouse click command, or via speech command, just to mention a few examples. The user may as an example select one potential defect at a time that may e.g. be displayed on a computer screen, or two or more potential defects, or all of, the potential defects simultaneously, where the user may visually see the development of the potential defects as a function of time or as a function of the built layers, starting from the first built layer. This may be presented in various formats, such as, but not limited to, as a line where each point on the line illustrates potential defects for each built layer. Accordingly, a very user-friendly way is provided for visually identifying how the building process is developing from the first built layer until the present built layer.

[0066] The term “localized-image evaluation” may according to the present invention involve dividing the captured image of a build layer into smaller sub-regions or tiles and independently analyzing each region for specific surface defects. This approach is particularly effective for identifying discrete, spatially confined anomalies such as insufficient powder distribution, recoater streaks, spatter, or protruding parts. For example, before building a layer, the processor may analyze small patches of the powder bed image to detect areas where the powder is unevenly spread or missing. Similarly, after building the layer, localized evaluation can detect small melt-related defects. Based on the detection in these specific regions, a localized-image failure metric is computed, providing a detailed assessment of the layer’s quality.

[0067] In contrast, the term “whole-image evaluation” may be understood as treating the image of the entire build layer as a single unit of analysis. This method typically involves comparing the overall image, either before or after layer construction, to a previously trained statistical or machine learning model that characterizes acceptable layer conditions. Whole-image evaluation is particularly suitable for detecting broader process anomalies that may not be localized but instead reflect systemic deviations, such as global thermal gradients, improper exposure patterns, or progressive drift in the process environment. The resulting wholeimage failure metric provides a holistic indication of whether the entire layer deviates from expected norms.

[0068] Together, these approaches enable a robust monitoring method capable of detecting both localized and global deviations in real time, thereby improving part quality, enhancing process reliability, and supporting adaptive corrective actions during the build process.

[0069] In a second aspect of the invention, a system is provided for real-time monitoring of additive manufacturing (AM) building process of a three-dimensional (3D) product, comprising:

[0070] • a 3D building device comprising a coating mechanism configured to alternatively distributing a material on a substrate by the coating mechanism and subsequently building a layer by means of fusing a portion of the material by a fusing device according to a planned geometry of the 3D product,

[0071] • an imaging device for providing image data for each build layer while building the 3D product, where the image data comprise image data of the distributed material before building a layer and image data after building a layer,

[0072] • a processor for processing the image data, for each of said build layers, where the processing comprises: o performing a localized-image evaluation of the image data, by means of:

[0073] ■ detecting local surface-defects in the distributed material before building the layer,

[0074] ■ detecting local surface-defects in the resulting build layer after building the layer from the distributed material, and based thereon

[0075] ■ determining a localized-image failure metric for the localized-image evaluation, and / or o performing a whole-image evaluation of the image data, by means:

[0076] ■ comparing the whole image data of distributed material before building the layer with previously trained statistical learning model, and

[0077] ■ comparing the whole image data after building the layer with previously trained statistical learning model, and based thereon

[0078] ■ determining a whole-image failure metric for the whole-image evaluation, wherein, if the determined failure metrics are below reference failure metric(s) the build layer is categorized as non-critical, the processor is further configured to utilize the results of said localized-image evaluation and / or the whole-image evaluation, as well as the result together with the most recent n number of printed layers, as input into a prediction model where a forward-looking failure risk indicator is determined indicating probability that the building process of the 3D product is trending toward failure or not.

[0079] The processor is further in an alternative embodiment, configured to trigger an electrical command such as an alert signal, if one or more of the determined failure metrics exceed said reference metric(s), or if the determined probability indicates that the building process of the 3D product is trending toward failure, indicating that subsequent build layers are under potential failure risk, and where the system further comprises an alert unit operable connected to the processor for triggering an alert command in response to said real-time alert signal.

[0080] The technical advantage of this embodiment lies amongst others in its capacity for proactive failure prevention, reducing material waste, machine downtime, and post-processing inspection efforts. The processor may, in another embodiment, further be configured to utilize the determined failure metric by issuing a rating score for the 3D building device indicating the reliability of the 3D building device. Thus, the system enables predictive maintenance and quality assurance, facilitating data-driven decisions for equipment usage, servicing schedules, and process optimization. Such a scoring system can also be used to compare performance across multiple machines or over time.

[0081] In an embodiment, the 3D building device comprises Laser Powder Bed Fusion (LPBF) printer and where the fusing device comprises a laser. This particular implementation benefits from the described real-time monitoring and failure prediction capabilities, as LPBF processes are sensitive to thermal gradients and layer-wise deviations. The integration of failure detection and alert mechanisms enhances process stability and part quality in high- precision manufacturing environments.

[0082] In a preferred embodiment, the 3D building device is configured to simultaneously build a plurality of 3D products on the substrate, and wherein the step of processing the image data by the processor comprises determining the localized-image failure metric and / or the wholeimage failure metric and / or the forward-looking failure risk indicator individually for each of the plurality of 3D products.

[0083] By incorporating real-time monitoring of failure metrics and probabilistic failure prediction, the system enables early detection of process anomalies before they propagate across subsequent layers. This proactive feedback mechanism allows for immediate corrective actions, such as halting the build, adjusting process parameters, or notifying the operator, thereby enhancing process stability, improving dimensional and structural integrity of the final part, and reducing the likelihood of costly part rejections or build failures. Moreover, the integration of such closed-loop control in LPBF systems supports the production of high- tolerance, mission-critical components, particularly in aerospace, medical, and industrial applications where build reliability is paramount.

[0084] An example of the imaging device used for monitoring the AM building process is a high- resolution digital camera, such as a CMOS or CCD camera, which may as an example be positioned within the 3D building device comprising a closed build chamber, configured for coaxial or oblique viewing of the build area. The camera may be equipped with an adjustable optical zoom lens and mounted on a fixed or movable gantry to enable image capture of different regions of the build plane.

[0085] To facilitate imaging both before and after each layer is formed, the camera may be synchronized with the recoater blade motion and the laser exposure process. The imaging system may include a lighting unit, such as an LED ring light or laser-based illumination, which may be arranged coaxially with the camera to provide uniform lighting conditions for capturing surface features of the powder bed and the built layer.

[0086] Accordingly, a system is provided whereby using failure metrics from a sequence of previous layers provides continuous data across the build, allowing the model to recognize evolving trends rather than isolated events. In additive manufacturing, each layer’s quality is typically highly correlated with previous layers. By feeding the current layer’s localized and wholeimage evaluation results along with the last n layer metrics into the model, the system maintains continuity of information over time. This cumulative view enables detection of gradual drifts or degradation that single-layer analysis might miss. For example, if a minor defect (e.g. slight overheating or misalignment) begins at one layer and incrementally worsens with each subsequent layer, a sequential model will discern that upward trend in the failure metric. In contrast, a one-shot model might overlook such a small anomaly until it becomes severe. Trend recognition is thus a key advantage of the system that is capable of identifying when a series of “okay” (non-critical) layers is showing a pattern of increasing anomaly that precedes a failure.

[0087] In general, the various aspects of the invention may be combined and coupled in any way possible within the scope of the invention. These and other aspects, features and / or advantages of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter.

[0088] BRIEF DESCRIPTION OF THE DRAWINGS Embodiments of the invention will be described, by way of example only, with reference to the drawings, in which

[0089] Figure 1 depicts a flowchart of one embodiment of a method according to the present invention for real-time monitoring of additive manufacturing (AM) building process of a three- dimensional (3D) product generated by a 3D building device,

[0090] Figure 2 depicts a flowchart of another embodiment of a method according to the present invention for real-time monitoring of additive manufacturing (AM) building process of a three- dimensional (3D) product generated by a 3D building device,

[0091] Figure 3 depicts a flowchart of yet another alternative embodiment according to the present invention for real-time monitoring of additive manufacturing (AM) building process of a three- dimensional (3D) product generated by a 3D building device,

[0092] Figures 4 and 5 illustrate graphically a frequency plot of number of defects per layer, where different types of potential defects or regions of interest are displayed to a user in real-time,

[0093] Figures 6 to 11 illustrate two examples of actual builds and serve to illustrate the practical benefits of the method and system according to the present invention, including the ability to predict and visualize the progression of build failures and to highlight the consequences of early versus late intervention, and

[0094] Figure 12 depicts a block diagram of a system according to the present invention.

[0095] DESCRIPTION OF EMBODIMENTS

[0096] Figure 1 shows a flowchart of one embodiment of a method according to the present invention for real-time monitoring of additive manufacturing (AM) building process of a three- dimensional (3D) product generated by a 3D building device. The building process comprises alternatively distributing a material on a substrate by a coating mechanism and subsequently building a layer by means of fusing a portion of the material by a fusing device according to a planned geometry of the 3D product. The distributed material may be a material including plastic or any plastic type material and where the 3D product is made of plastic or any plastic type material. In another alternative embodiment, the distributed material is a metal powder and where the fused portion is a metal portion and where the 3D product is a metal product. In a first step (S1) 101 , image data obtained from an imaging device is received for each build layer while building the 3D product, where the image data comprises image data of the distributed material before building a layer and after building a layer. The imaging device may include any type of digital camera that may be placed within or outside the chamber where the building process takes place.

[0097] In a second step (S2) 102, the image data of each build layer is processed, by a processor, preferably in real time. The processing for each build layer comprises processing the image data together with reference data, where the reference data comprises reference image data of a distributed material before building the layer and reference image data after building a layer.

[0098] In a third step (S3) 103, a statistical matching likelihood score between the image data and the reference data is determined.

[0099] In a fourth step (S4) 104, the statistical matching likelihood score is utilized in determining a failure metric for the building process, where the determined failure metric is compared with a reference failure metric, where in case the determined failure metric exceeds the reference failure metric, issuing a real-time alert signal indicating that subsequent build layers are under potential failure risk.

[0100] The step of processing the image data in S2 may in an embodiment further comprise identifying potential defects having one or more different characteristic properties and utilize the identified potential defects as additional input in determining the failure metric.

[0101] The potential defects include, but are not limited to, one or more of the following: a protruding fused material defect occurring due to warpage of the fused material, a spatter defect occurring when laser melt pool conditions are poor causing unwanted particles to be formed such as due to oxidized powder or melt pool ejections, a streaking defect occurring when the coating mechanism is defective causing insufficient powder distribution during coating or due to loose solidified particles being dragged along with the coating mechanism, a powder hole or cavity caused due to insufficient distribution of powder, a hopping or ringing defect occurring during powder distribution process where the coating mechanism may collide with a protruding part causing e.g. a mechanical ringing or vibration resulting from collision, a smoke defect caused by suboptimal build chamber environment parameters, e.g. gas composition or gas flow, a burn defect caused when the laser, subsequent to distribution of metal powder, strikes an already solidified part, and porosity defect caused by voids or holes inside fused portions or layers.

[0102] The identified potential defects may in an alternative embodiment be used as an additional input in calculating, e.g. via deep learning or any type of machine learning or mathematical modeling, a statistical likelihood of a critical failure for a current or future build layer based on identified potential defects in previous build layers. This may in an alternative embodiment comprise identifying potential defects by comparing the characteristic properties of the identified potential defects with characteristic properties of previously detected defects obtained from previously built layers of said 3D product and / or from previous 3D products. Based on the comparison, the identified potential defects may be grouped into different characteristic property groups where each group contains potential defects of similar or identical nature.

[0103] Moreover, the number of potential defects within each group may be calculated, where the calculated number is configured to be used in presenting in real time a frequency pattern for different types of potential defects in real-time. The step of calculating may include determining a number of potential defect matches and / or calculating area of the potential defects and utilizing the area in converting it into a representation of the number of potential defects. An input command may in an embodiment be received from a user via a user interface, and in response to the received command one or more of the different types of potential defects or regions of interest are displayed.

[0104] Figure 2 shows a flowchart of another embodiment of a method according to the present invention for real-time monitoring of additive manufacturing (AM) building process of a three- dimensional (3D) product generated by a 3D building device, where the building process comprises alternatively distributing a material on a substrate by a coating mechanism and subsequently building a layer by means of fusing a portion of the material by a fusing device according to a planned geometry of the 3D product.

[0105] Step (S1) 201 includes obtaining, from an imaging device, image data for each build layer while building the 3D product. The image data, for each build layer, comprises a pair of images including image data representing the distributed material prior to building the layer, and image data representing the corresponding build layer after formation from the distributed material. Step (S2) 202 includes processing, by a processor, the image data for each of said build layers by means of: (i) performing a localized-image evaluation of the image data and / or (ii) performing a whole-image evaluation of the image data.

[0106] In an alternative embodiment, (i) includes detecting local surface-defects in the distributed material before building the layer, detecting local surface-defects in the resulting build layer after building the layer from the distributed material, and based thereon determining a localized-image failure metric for the localized-image evaluation.

[0107] In an alternative embodiment, (ii) includes comparing the whole image data of distributed material before building the layer with previously trained statistical learning model and comparing the whole image data after building the layer with previously trained statistical learning model, and based thereon determining a whole-image failure metric for the wholeimage evaluation.

[0108] In step (S3) 203 it is checked, e.g. by the processor, if the determined failure metrics are below reference failure metric(s) the build layer is categorized as non-critical.

[0109] Step (S4) 204 includes, in case the determined failure metrics are below reference failure metric(s), utilizing the results of said localized-image evaluation and / or the whole-image evaluation, as well as the result together with the most recent n number of printed layers, as input into a prediction model where a forward-looking failure risk indicator is determined indicating probability that the building process of the 3D product is trending toward failure or not.

[0110] The term localized-image evaluation may, according to the present invention, be understood as focusing on identifying specific, local surface defects by analysing subregions or regions of interest within the layer’s images. This approach scrutinizes the distributed powder material before fusion and the solidified layer after fusion for any irregularities at fine spatial resolution. For example, each captured layer image may be scanned for discrete anomalies such as powder recoating irregularities, spatter deposits, or localized delamination cracks. Irregular powder deposition defects, e.g. a streak or bare spot left by the coating mechanism, powder clumps, or part protrusions, can be detected on the pre-fusion layer image as localized out-of-norm regions. Likewise, after the layer is melted, the post-fusion image is examined for small-scale defects on the new surface, such as balling (spheroidal beads of ejected metal), isolated pores or ditches, and micro-cracks or warping at specific sites. The localized evaluation may employ computer-vision algorithms or machine learning object detectors that locate these defect indications within the image (for instance, using bounding boxes or segmentation to mark defect areas). Each detected defect contributes to a localized-image failure metric, a quantification of layer quality at a granular level (e.g., the number, size, or severity of defects in that layer). This localized metric can provide immediate, targeted feedback on any spot that deviates from expected layer conditions. By focusing on individual defect sites, the localized-image evaluation enables precise identification of defect types (such as powder bed voids, spatter-induced protrusions, or delamination of a small region) and their coordinates on the layer, facilitating targeted corrective actions or layer-wise rejection if a critical flaw is found. It thereby excels at capturing fine-grained, stochastic defects in real-time, before they propagate into more severe flaws in subsequent layers.

[0111] In contrast, the whole-image evaluation may, according to the present invention, be understood as treating the entire layer image as a single data set to assess overall layer integrity. In one embodiment, this approach involves comparing the full image (both of the distributed powder before fusion and of the fused layer after fusion) against a previously trained statistical learning model representing normal build conditions. The model, for example, a machine learning classifier or an anomaly detection algorithm, is trained on numerous images of acceptable and non-acceptable layers to capture the global appearance and texture of normal and abnormal powder bed layers and properly and in- properly fused layers, respectively. For each new layer, the system feeds the complete image into this model (or extracts global features such as overall brightness uniformity, texture patterns, or thermal signatures) and obtains a prediction or anomaly score indicating whether the layer is within normal variation. In essence, the whole-image evaluation asks: “Does this entire layer look anomalous as a whole?” The image is classified to the category most similar, and if the image is classified as abnormal , it will be flagged. This could detect subtle or diffuse issues that might not be confined to one small region. For instance, an overall darker powder-bed image might indicate insufficient powder across the layer, or a widespread thermal glow pattern might suggest an overheating condition. Using e.g. a holistic analysis, global defects may be caught such as a systematic coating mechanism malfunction affecting the entire layer or a large-area delamination (e.g., an entire section of the layer lifted) that imparts a distinct signature across the image. The outcome of this analysis is a whole-image failure metric, reflecting the layer’s quality in aggregate (e.g., a probability of layer failure or an anomaly score). It should be noted that prior research in LPBF monitoring has demonstrated that machine learning models can effectively perform such layer-wide anomaly detection by evaluating the entire layer image at once. Convolutional neural network classifiers, for example, have been used to rapidly classify each layer’s image as normal or anomalous based on learned patterns. These whole-image methods are well-suited to detect distributed or emergent process deviations: for example, a trained model can recognize the overall temperature distribution in a thermal image of the layer and flag disturbances in the thermal profile that correlate with defect formation. By analysing the cumulative visual or thermal features of the layer, the whole-image evaluation provides a broad quality assessment, ensuring that no large-scale or subtle trend goes unnoticed even if individual defects are below local detection thresholds.

[0112] Similarly, the detection of an abnormal powder bed layer in an optical image may be performed using the same whole-image analysis approach. An example of a suitable imaging device for monitoring the additive manufacturing (AM) build process is a high- resolution digital camera, such as a CMOS or CCD sensor, positioned within the 3D printing apparatus, which includes a sealed build chamber. The camera is configured for either coaxial or oblique viewing of the build area and may be equipped with an adjustable optical zoom lens. It can be mounted on a fixed or movable gantry to facilitate image acquisition from different regions of the build surface.

[0113] Said step of performing the localized-image evaluation of the image data and / or the wholeimage evaluation of the image data may in an embodiment include processing the image data together with reference data, where the reference data comprises reference image data of a distributed material before building the layer and reference image data after building a layer. This may include determining a statistical matching likelihood score between the image data and the reference data, where said localized-image failure metric and / or said whole-image failure metric are determined based on the statistical matching likelihood score.

[0114] Step (S5) 205 includes, if one or more of the determined failure metrics exceed said reference metric(s), or if the determined probability indicates that the building process of the 3D product is trending toward failure, triggering an electrical command such as an alert signal indicating that subsequent build layers are under potential failure risk.

[0115] Figure 3 depicts a flowchart of yet another alternative embodiment of a method according to the present invention. The steps discussed previously in relation to Figure 1 and / or 2 may be performed by software, which may be considered as prediction software. In step (ST) 301 , image data obtained from an imaging device are received for each build layer while building the 3D product, where the image data comprise image data of the distributed material before building a layer and after building a layer, as discussed previously.

[0116] In this embodiment, three parallel processing steps 302, 303, 304 are performed.

[0117] Step 302 includes two sub-steps, where in sub-steps 302’ a prediction model analyzes the chance of future failure based on the status and defects of the last n layers, and sub-step 302” displays a chance of failure in a gauge or by other means of display in build job overview.

[0118] Step 303, which is a whole-image evaluation, includes two sub-steps, where in sub-step 303’ a classification model analyzes the image pair for general status of build, i.e. and sub-step 303” displays the status of the build job on a status bar.

[0119] Step 304, which is a localized-image evaluation, includes three sub-steps, where in sub-step 304’ a detection model analyzes the image pair for defects, sub-steps 304” displays the detected defects on the image pair to aid a user in evaluating the build job, and step 304’” the number of defects per type is shown to the user in an analyzing tool.

[0120] If any of the above-mentioned three parallel processing steps 302, 303, 304 result in that a real-time alert signal is triggered 307 an alert is sent to the user 305, and the build job may automatically or via the user be stopped. If not, it is checked if the build is completed 308, and if yes, a quality report is generated for the build 306.

[0121] If the build is not completed, the above-mentioned steps are repeated starting from step ST 201.

[0122] Figures 4 and 5 illustrate an example of user friendly way to graphically present a frequency plot of number of defects per layer, where different types of potential defects or regions of interest are displayed to a user in real-time. As shown here, over 550 layers have been built, but the number of defects for each build layer is displayed in real time, starting with the first build layer on the far-left side. Such a building process can take hours, days or even weeks, depending on the volume and complexity of the 3D product being built.

[0123] In Figure 4, a user 401 selects a single defect type (“Metal”), which triggers displaying a frequency plot of the selected defect 402, whereas in Figure 5 two defects are selected, “Metal” and “Spatter”.

[0124] The vertical axis represents the number for potential defects selected and the horizontal axis represents the layer number.

[0125] In Figure 4, each circle in the plot presents the number of “Metal” defects for a single build layer before building the layer and after building the layer, whereas already mentioned, the image pairs for each layer are analyzed, e.g. using deep learning algorithms, to determine defective areas in the images. This may be done by a separate deep learning model that is trained for the coating layer images and the solidified layer images.

[0126] In Figure 5 the solid line additionally represents the “Spatter” defect that is displayed together with the “Metal” defects.

[0127] A slider module enables the user to slide the image from layer to layer as depicted by the arrow, where upon sliding as depicted here the image pairs for each layer 403, 404 are displayed on the display screen, where e.g. the left images 403, 503 represent coat image and the right images 404, 504 represents solidification image, for the same layer.

[0128] Below the horizontal axis is a visual real-time status of the build process 405, 406, 407, 408 and 505, 506, 507, 508, which may be visually shown via different coloring where e.g. green color indicates that the build process is accordion to plan, a yellow color indicates possible future failure, and a red color indicates a critical potential failure. This example serves merely as one illustrative approach for presenting the information in a user-friendly manner and should not be interpreted as limiting the implementation to the depicted colour scheme.

[0129] Figures 4 and 5 illustrate a single build job for a single machine, where after completing the build job the machine may be rated with a kind of quality indicator and / machine health statistic. As an example, at the end of each print, the overall build job resulting in the 3D part may be evaluated as a whole, and a quality rating is assigned to the part.

[0130] The method steps discussed in previous embodiments apply here, where at any instant of time, the build job may be terminated, either by the user or be automatically terminated.

[0131] Begin Example 1 :

[0132] Figures 6 and 7 illustrate an exemplary additive manufacturing build monitored using the method according to the present invention, wherein real-time image analysis and predictive failure detection are applied across successive layers of the build sequence.

[0133] In particular, Figure 6 presents a frequency-based visual plot 602 representing the calculated failure probability across layers up to and including “layer 160” 620. At this point in the build process, the processor determines, based on image-derived metrics such as 622, 623 (e.g., localized defect density, thermal signature deviation, and whole-image anomaly scoring), that the aggregated failure probability for upcoming layers exceeds a predefined threshold. This condition is reflected in the predictive layer status indicator corresponding to layer 160, which marks the first automated forecast of an impending failure state within the current build cycle.

[0134] The method identifies this condition not based on a present-layer failure event but based on cumulative indicators suggesting that downstream process instability is likely, e.g., due to increasing irregularities in powder distribution or melt pool consistency, or an emerging pattern of localized deviations suggestive of coating mechanism induced artifacts.

[0135] Figure 7 subsequently illustrates the condition at layer 195 720, wherein a critical failure is detected. At this stage, the image evaluation algorithms confirm that failure metrics, such as delamination signatures, thermal discontinuity, or severe surface discontinuities surpass critical intervention thresholds. This event corresponds to an actual breakdown in the build, manifesting as layer delamination or structural collapse in affected regions and significant damage to the recoater mechanism.

[0136] Had the predictive alert at layer 160 triggered a system response, such as automated build pausing, dynamic adjustment of process parameters, or operator intervention, the failure occurring at layer 195 could have been avoided. This example demonstrates how the method according to the present invention supports anticipatory failure management by identifying precursor conditions significantly in advance of critical build failure, thereby enabling proactive mitigation and improved build reliability.

[0137] End Example 1.

[0138] Begin Example 2:

[0139] Figures 8 through 11 illustrate a second representative build monitored using the method according to the present invention, demonstrating the system’s ability to predict and visualize the progression of failure over multiple layers in an additive manufacturing process.

[0140] In Figure 8, a frequency-based visual plot 802 is shown, capturing the evolution of calculated failure probabilities across sequential build layers. At this layer, the probability of failure is predicted as minimal as the power bed layer is nominal. The processor identifies layer 158, as shown in Figure 9, as the point at which the aggregated failure probability exceeds a predefined criticality threshold. This prediction is derived from a combination of real-time image data analysis, incorporating both localized-image evaluation (e.g., identifying increasing densities of powder bed anomalies and surface irregularities) and whole-image evaluation (e.g., deviation in global thermal distribution or pattern recognition via statistical models). The predictive layer status indicator generated at this stage signifies the onset of potentially destabilizing conditions in the build process.

[0141] Following this point, Figures 9 and 10 illustrate the physical manifestation and spatial escalation of build degradation in subsequent layers. In Figure 9, the appearance of dotted regions corresponds to early-stage physical indicators 922 of process instability, possibly caused by coating mechanism interference, insufficient powder coverage, or incipient delamination effects, which begin to appear and expand across the build surface. These regions are consistent with localized surface degradation flagged by the predictive analysis.

[0142] By layer 171 1021 , as shown in Figure 10, the build has entered a clearly deteriorating state, with the affected areas, e.g. 922 and 923, visibly expanding and becoming more pronounced. This layer represents the transition from predicted instability to confirmed defect formation, validating the reliability of the earlier prediction at layer 159 921.

[0143] Finally, Figure 11 presents the result of continued processing without intervention, culminating in a severely damaged build, e.g. 922 and 923. The surface condition at this stage is marked by substantial material irregularities, structural compromise, and extensive loss of build integrity, rendering the part unsalvageable. This outcome demonstrates the consequences of inaction following a system-generated alert, and further supports the efficacy of the method in providing early warnings of failure well in advance of the actual defect onset.

[0144] This example emphasizes the predictive depth and diagnostic resolution of the method according to the present invention, showcasing its ability to trace the development of critical failure from its statistical forecast through to its final physical manifestation across the layer sequence.

[0145] End Example 2.

[0146] Figure 12 depicts a block diagram of a system 1200 according to the present invention, comprising a 3D building device (3D_B) 1201 , an imaging device (l_D) 1202, a processor (P) 1203 and an alert unit (A_U) 1204 .

[0147] The 3D building device is configured to alternatively distribute a material on a substrate by a coating mechanism and subsequently building a layer by means of fusing a portion of the material by a fusing device according to a planned geometry of the 3D product.

[0148] The imaging device provides image data for each build layer while building the 3D product, where the image data comprises image data of the distributed material before building a layer by fusing a portion of the material and image data of the built layer.

[0149] The processor is configured for processing the image data of each of said build layer the first layer in real-time, where the processing includes: i) performing a localized-image evaluation of the image data, by means of: detecting local surface-defects in the distributed material before building the layer, detecting local surface-defects in the resulting build layer after building the layer from the distributed material, and based thereon determining a localized- image failure metric for the localized-image evaluation, and / or ii) performing a whole-image- evaluation of the image data, by means: comparing the whole image data of distributed material before building the layer with previously trained statistical learning model, and comparing the whole image data after building the layer with previously trained statistical learning model, and based thereon, determining a whole-image failure metric for the whole- image-evaluation.

[0150] The processor may further be configured to trigger an electrical command, such as an alert signal via the alert unit 1204, if one or more of the determined failure metrics exceed one or more reference metrics, or if the determined probability indicates that the building process of the 3D product is trending toward failure, thereby suggesting that subsequent build layers are at risk of failure. The system further comprises an alert unit operably connected to the processor and configured to issue an alert command in response to the real-time alert signal. Additionally, the processor is configured to utilize the determined failure metrics to generate a rating score representing the reliability of the 3D building device, which may comprise, for example, a Laser Powder Bed Fusion (LPBF) printer wherein the fusing device includes a laser. Furthermore, the 3D building device is configured to simultaneously manufacture a plurality of 3D products on a substrate, and the processing of image data by the processor includes determining the localized-image failure metric, the whole-image failure metric, and / or a forward-looking failure risk indicator individually for each of the plurality of 3D products.

[0151] Moreover, as discussed in relation to depicted in Figures 4 and 5, the system 1200 further comprises an user interface comprising an input unit or a slider configured to receive an input command from a user, and a display for displaying, in response to the input command, one or more of the different types of potential defects discussed in relation to the flowchart in Figures 1 to 3.

[0152] While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive; the invention is not limited to the disclosed embodiments. Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from a study of the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

Claims

CLAIMS1. A method for real-time monitoring of additive manufacturing (AM) building process of a three-dimensional (3D) product generated by a 3D building device, where the building process comprises alternatively distributing a material on a substrate by a coating mechanism and subsequently building a layer by means of fusing a portion of the material by a fusing device according to a planned geometry of the 3D product, comprising:■ obtaining, from an imaging device, image data for each build layer while building the 3D product,■ processing, by a processor, the image data for each of said build layer, wherein the image data, for each build layer, comprises a pair of images including image data representing the distributed material prior to building the layer, and image data representing the corresponding build layer after formation from the distributed material, wherein processing the image data, for each of said build layers, comprises:• performing a localized-image evaluation of the image data, by means of: o detecting local surface-defects in the distributed material before building the layer, o detecting local surface-defects in the resulting build layer after building the layer from the distributed material, and based thereon o determining a localized-image failure metric for the localized-image evaluation, and / or• performing a whole-image evaluation of the image data, by means: o comparing the whole image data of distributed material before building the layer with previously trained statistical learning model, and o comparing the whole image data after building the layer with previously trained statistical learning model, and based thereon o determining a whole-image failure metric for the whole-image-evaluation, wherein, if the determined failure metrics are below reference failure metric(s) the build layer is categorized as non-critical, and the method further comprises:• utilizing the results of said localized-image evaluation and / or the whole-image- evaluation, as well as the result together with the most recent n number of printed layers, as input into a prediction model where a forward-looking failure risk indicator is determined indicating probability that the building process of the 3D product is trending toward failure or not.

2. The method according to claim 1, wherein the step performing the localized-image evaluation of the image data and / or the whole-image-evaluation of the image data comprises:• processing the image data together with reference data, where the reference data comprises reference image data of a distributed material before building the layer and reference image data after building a layer.

3. The method according to claim 2, further comprising utilizing the result of the processing in determining a statistical matching likelihood score between the image data and the reference data, where said localized-image failure metric and / or said whole-image failure metric are determined based on the statistical matching likelihood score.

4. The method according to any of the preceding claims, wherein the results of said localized-image evaluation and said whole-image-evaluation are both utilized as input into the prediction model for determining the forward-looking failure risk indicator.

5. The method according to any of the preceding claims, wherein if one or more of the determined failure metrics exceed said reference metric(s), or if the determined probability indicates that the building process of the 3D product is trending toward failure, an electrical command such as an alert signal is triggered indicating that subsequent build layers are under potential failure risk.

6. The method according to any of the preceding claims, wherein the step of determining the failure metric(s) for the building process is continuously repeated in real-time during the building process.

7. The method according to any of the preceding claims, wherein the step of detecting local surface-defects before and after building the layer comprises:• identifying potential defects having one or more different characteristic properties, where the step of determining the localized-image failure metric is further based on using the identified potential defects.

8. The method according to claim 7, further comprising grouping the identified potential defects into different characteristic property groups, each group containing potential defects of similar or identical nature.

9. The method according to claim 8, further comprises calculating the number of potential defects within each group, wherein the calculated number is configured to be used in presenting in real time a frequency plot for different types of potential defects, where the step of calculating may e.g. comprise determining a number of potential defect matches and / or calculating areal of the potential defects and utilize the areal in converting it into representation of the number of potential defects.

10. The method according to any of the claims 9 to 11 , wherein the potential defects include:• a protruding fused material defect occurring due to warpage of the fused material,• a spatter defect occurring when laser melt pool conditions are poor causing unwanted particles to be formed such as due to oxidized powder or melt pool ejections,• a streaking defect occurring when the coating mechanism is defective causing insufficient powder distribution during coating or due to lose solidified particles being dragged along with the coating mechanism,• a powder hole or cavity cause due to insufficient distribution of powder,• a hopping or ringing defect occurring during powder distribution process where the mechanism may collide with a protruding part causing e.g. a mechanical ringing or vibration resulting from collision,• a smoke defect caused by suboptimal build chamber environment parameters, e.g. gas composition or gas flow,• a burn defect caused when the laser, subsequent to distribution of metal powder, strikes already solidified metal,• porosity defect caused by voids or holes inside fused portions or layers.

11. A system is provided for real-time monitoring of additive manufacturing (AM) building process of a three-dimensional (3D) product, comprising:• a 3D building device comprising a coating mechanism configured to alternatively distributing a material on a substrate by the coating mechanism and subsequentlybuilding a layer by means of fusing a portion of the material by a fusing device according to a planned geometry of the 3D product,• an imaging device for providing image data for each build layer while building the 3D product, where the image data comprise image data of the distributed material before building a layer and image data after building a layer,• a processor for processing the image data, for each of said build layers, where the processing comprises: o performing a localized-image evaluation of the image data, by means of:■ detecting local surface-defects in the distributed material before building the layer,■ detecting local surface-defects in the resulting build layer after building the layer from the distributed material, and based thereon■ determining a localized-image failure metric for the localized-image evaluation, and / or o performing a whole-image-evaluation of the image data, by means:■ comparing the whole image data of distributed material before building the layer with previously trained statistical learning model, and■ comparing the whole image data after building the layer with previously trained statistical learning model, and based thereon■ determining a whole-image failure metric for the whole-image-evaluation, wherein, if the determined failure metrics are below reference failure metric(s) the build layer is categorized as non-critical, the processor is further configured to utilize the results of said localized-image evaluation and / or the whole-image-evaluation, as well as the result together with the most recent n number of printed layers, as input into a prediction model where a forward-looking failure risk indicator is determined indicating probability that the building process of the 3D product is trending toward failure or not.

12. The system according to claim 13, wherein the processor is further configured to trigger an electrical command such as an alert signal, if one or more of the determined failure metrics exceed said reference metric(s), or if the determined probability indicates that the building process of the 3D product is trending toward failure, indicating that subsequent build layers are under potential failure risk, and where the system further comprises an alert unit operable connected to the processor for triggering an alert command in response to said real-time alert signal.

13. The system according to claim 11 or 12, wherein the processor is further configured to utilize the determined failure metric by issuing a rating score for the 3D building device indicating the reliability of the 3D building device, where the 3D building device may e.g. comprise Laser Powder Bed Fusion (LPBF) printer and where the fusing device comprises a laser.

14. The system according to any of claims 11 to 13, wherein the 3D building device is configured to simultaneously build a plurality of 3D products on the substrate, and wherein the step of processing the image data by the processor comprises determining the localized- image failure metric and / or the whole-image failure metric and / or the forward-looking failure risk indicator individually for each of the plurality of 3D products.

15. The system according to any of the claims 11 to 14, wherein the imaging device comprises one or more of: a high-resolution digital camera, such as a CMOS or CCD camera, positioned within the 3D building device, and / or a heat camera.

Images