ArticlePublisher preview available
To read the full-text of this research, you can request a copy directly from the authors.

Abstract and Figures

As a promising modern technology, additive manufacturing (AM) has been receiving increasing research and industrial attention in the recent years. With its rapid development, the importance of quality monitoring in AM process has been recognized, which significantly affects the property of the manufactured parts. Since the conventional hand-crafted features for quality identification are generally costly, time-consuming and sensitive to noises, the intelligent data-driven automatic process monitoring methods are becoming more and more popular at present. This paper proposes a deep learning-based quality identification method for metal AM process. To alleviate the requirement for large amounts of high-quality labeled training data by most existing data-driven methods, an identification consistency-based approach is proposed to better explore the semi-supervised training data. The proposed method is able to achieve promising performance using limited supervised samples with low quality, such as noisy and blurred images. Experiments on a real-world metal AM dataset are implemented to validate the effectiveness of the proposed method, which offers a promising tool for real industrial applications.
This content is subject to copyright. Terms and conditions apply.
Journal of Intelligent Manufacturing (2020) 31:2003–2017
Quality analysis in metal additive manufacturing with deep learning
Xiang Li1,2 ·Xiaodong Jia1·Qibo Yang1·Jay Lee1
Received: 24 October 2019 / Accepted: 14 February 2020 / Published online: 25 February 2020
© Springer Science+Business Media, LLC, part of Springer Nature 2020
As a promising modern technology, additive manufacturing (AM) has been receiving increasing research and industrial
attention in the recent years. With its rapid development, the importance of quality monitoring in AM process has been
recognized, which significantly affects the property of the manufactured parts. Since the conventional hand-crafted features
for quality identification are generally costly, time-consuming and sensitive to noises, the intelligent data-driven automatic
process monitoring methods are becoming more and more popular at present. This paper proposes a deep learning-based
quality identification method for metal AM process. To alleviate the requirement for large amounts of high-quality labeled
training data by most existing data-driven methods, an identification consistency-based approach is proposed to better explore
the semi-supervised training data. The proposed method is able to achieve promising performance using limited supervised
samples with low quality, such as noisy and blurred images. Experiments on a real-world metal AM dataset are implemented
to validate the effectiveness of the proposed method, which offers a promising tool for real industrial applications.
Keywords Additive manufacturing ·Process monitoring ·Quality identification ·Deep learning ·Low-quality data
In the recent years, additive manufacturing (AM) techniques
have been emerging as one of the most promising manufac-
turing technologies in a wide variety of application scenarios
(Kwon et al. 2018; Tapia and Elwany 2014;Huetal.2019;
Zhao and Guo 2019; Gonzalez-Val et al. 2019), such as auto-
motive, aerospace, robotics, electronics etc. AM which is also
popularly known as 3D printing, free-form fabrication and
rapid prototyping, has been receiving increasing academic
and industrial attention, since it is capable of manufacturing
highly sophisticated and fully functional 3D objects that are
difficult to be accomplished by conventional manufacturing
approaches (Everton et al. 2016). This novel technology has
the potential to largely reduce the existing constraints on the
geometrical design, accelerate the production time and min-
imize the manufacturing cost.
BXiang Li
1Department of Mechanical and Materials Engineering,
University of Cincinnati, Cincinnati, OH 45221, USA
2Key Laboratory of Vibration and Control of Aero-Propulsion
System Ministry of Education, Northeastern University,
Shenyang 110819, China
Currently, different AM methods have been developed,
including material extrusion, sheet lamination, directed
energy deposition, material jetting, powder bed fusion etc.
Among them, the powder bed fusion (PBF) approach has
been one of the most popular methods in metal part addi-
tive manufacturing (Chua et al. 2017; Zhang et al. 2018).
Generally, metallic powder is spread over the previous lay-
ers during the AM process, and a laser is usually used as
the energy source to achieve the metallic bonding between
layers. In this study, the PBF process is investigated.
Despite the promising development of AM methods, the
main existing quality-related problems lie in the limited sta-
bility and repeatability of the mechanical properties of the
manufactured parts in production (Shevchik et al. 2019).
Generally, the part quality is directly influenced by the mate-
rial microstructure in the AM process, and the defects such
as porosity, residual stress and cracks are caused by a num-
ber of factors, including laser scanning speed, laser power,
hatch distance etc. Timely and accurate quality assessment
in AM process is of great importance in improving manu-
facturing quality and reducing industrial cost (Kwon et al.
2018). However, it is challenging for real implementation of
the assessment due to the high complexity of the involved
physics covering a wide range of practical factors.
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
... Based on the findings of the literature review, it is observed that medical 3DP depends upon operational barriers such as production inconsistency (Pal and Gander 2018;Li et al. 2020), lack of appropriate performance measurement (Primo et al. 2017;De Schutter et al. 2018), lack of stability and reliability of 3DP process (Zhang, Liu, and Shin 2019;Niaki, Torabi, and Nonino 2019), product size constraints (Zhao et al. 2019), low production speed (Baumers et al. 2016;Teubler et al. 2019), poor quality finished products (Baumers et al. 2016;Qiu et al. 2019), unskilled labour and technical expertise in manufacturing and troubleshooting issues (Pan et al. 2018;Illmann, Hosking, and Harris 2020), low maintenance-related facilities (Yılmaz 2020;Powell et al. 2020;Priyadarshini, Singh, Mishra, and Kamal 2022). ...
... More research is required to overcome the identified barriers of legal and policy support and management support (Choonara et al. 2016;Ashima et al. 2021;Stephens et al. 2021). The practitioners need to overcome alternative financial support (Li et al. 2020;Tyagi, Spinks, and Jager 2021), the latest technological option (Shilo et al. 2018;Sedlakova et al. 2022), improved legal and policy (Dodgson 2020;Daly et al. 2021), and management focus on measurable goals (Pietrabissa et al. 2020; Ravi, Antoline, and Rybicki 2021) for the successful adoption of SMAM. Stakeholders need to overcome the fear of losing revenue and the problems of obtaining bank financing to help the implementation of SMAM with the help of partners in the supply chain. ...
Sustainable medical additive manufacturing (SMAM) is becoming the next Industry 4.0 technology to revolutionise the medical industry. The adoption of SMAM offers several advantages and brings a paradigm shift in complex manufacturing geometry with improved quality, speed, cost, and sustainability in medical sectors. This research aims to identify the adoption barriers of SMAM in the Indian context. The research design involves two-step procedures: First, a literature review was conducted to determine the barriers to SMAM technology adoption. Later, these were validated by a panel of experts from industry and academia. Second, a hybrid ISM-DEMATEL methodology was deployed to establish and evaluate the cause-effect relationship between the validated barriers. Our findings suggest that among the identified barriers, infrastructural barriers were the most important for adopting SMAM in India, followed by a lack of long-term planning, operational barriers, and supply-demand barriers. Further, it also identified net cause driving barriers, including financial barriers, legal and policy barriers , technological barriers, and management barriers to SMAM adoption. As the adoption of SMAM offers several advantages, including a shift in complex manufacturing geometry with improved quality, speed, cost, and sustainability in medical sectors, these findings assume significance and will help decision-makers overcome complex barriers to SMAM adoption.
... When both time-sequential and timeinvariant data were involved, Ren et al. combined LSTM with an improved Wide-and-Deep model to capture the key quality information [12]. When image data was involved, convolutional neural networks (CNNs) were used to pre-process and detect defects from various processes, including welding and additive manufacturing [13,14]. These approaches effectively delegate feature engineering to DNNs and sidestep the need for human expertise-based feature design. ...
... These features may affect the functional performance of the printed articles. Achieving high-quality products requires laborious monitoring procedures involving instrumental control and computer-assisted engineering, including machine learning [36]. Adopting AM for medical products is impossible without the implementation of specialized approaches allowing to highly precisely print defect-free, thoughtfully architectured articles. ...
Full-text available
We overview recent findings achieved in the field of model-driven development of additively manufactured porous materials for the development of a new generation of bioactive implants for orthopedic applications. Porous structures produced from biocompatible titanium alloys using selective laser melting can present a promising material to design scaffolds with regulated mechanical properties and with the capacity to be loaded with pharmaceutical products. Adjusting pore geometry, one could control elastic modulus and strength/fatigue properties of the engineered structures to be compatible with bone tissues, thus preventing the stress shield effect when replacing a diseased bone fragment. Adsorption of medicals by internal spaces would make it possible to emit the antibiotic and anti-tumor agents into surrounding tissues. The developed internal porosity and surface roughness can provide the desired vascularization and osteointegration. We critically analyze the recent advances in the field featuring model design approaches, virtual testing of the designed structures, capabilities of additive printing of porous structures, biomedical issues of the engineered scaffolds, and so on. Special attention is paid to highlighting the actual problems in the field and the ways of their solutions.
... • The results of the numerical simulation and the predictions made by the RNN-DNN showed an accuracy level of more than 95% when they were compared. Li, Jia [76] To propose a DL-based quality identification method for the metal AM process. ...
Full-text available
This paper critically examined the research landscape and the impact of machine/deep learning on additive manufacturing (MDLAM) through publication trends, bibliometric analyses, and literature review. The Elsevier Scopus database was selected to identify and recover publications on MDLAM research published from 2013 to 2022 based on the Prisma approach. The recovered bibliographic data was analyzed using VOSViewer software to examine the co-authorship, keyword, and citation networks on the MDLAM research. Results showed that the publications output (and citations count) increased progressively from 1 (19) to 375 (980) from 2013 to 2022, which exhibits the high TC/TP ratio typically characteristic of highly impactful fields with future growth potentials. Analysis of top performers on the topic revealed that Prahalada K. Rao (US), Pennsylvania State University (US), and National Science Foundation (US) are the most prolific authors, affiliations, and funder of MDLAM research, respectively. Hence, the most active nation on MDLAM research is the United States (US), although China and the United Kingdom have also made significant contributions over the years. Keyword occurrence revealed the existence of several research hotspots with researchers' interests directed at basic research, optimization studies, industrial applications, and novel learning systems. The paper showed that MDLAM is a broad, complex, and impactful research area that will continue to experience scientific growth and technological development, mainly due to the growing demands for accurate computational methods for AM prototypes, processes, and products.
... Using real-time, AM process data as inputs to use machine learning (ML) models for defect detection has become popular. Li et al. [10] proposed measuring part quality using image data collected during the process to build a deep learning-based quality identification method. Mohammadi et al. [11] studied reliable product quality through dimensional error prediction, investigating the performance of several different ML methodologies to detect defects in real-time. ...
Full-text available
Metal Additive Manufacturing (MAM) produces complex, part geometries from a variety of materials in powder and wire form. Due to complexities of MAM processes that create those geometries, especially powder bed fusion, quality assurance, and qualification remain an ongoing challenge. Quality assurance involves assessing the quality of a part’s geometry, surface finish, and mechanical properties. Currently, quality assurance is not easily achieved due to variations in the powder inputs, the MAM process itself, and environmental factors, such as temperature. Many efforts are underway to develop a new quality system that includes 1) planning, measuring, and qualifying parts and 2) enhancing quality through a processing-monitoring-qualifying framework. Creating this new system requires building the complex relationships between requirements, processes, and quality. These relationships are needed to specify, measure, analyze, and optimize variables to ensure final part quality. Thus, a processing-monitoring-quality framework could provide critical steps to identify those relationships and help meet stakeholder needs. The paper describes how to adapt the “software and systems engineering” V-model to a “metal AM quality assurance” V-model that can provide a framework for quality assurance in MAM.
... To address these research problems, researchers have proposed the use of semisupervised learning methods that require partial labels, and unsupervised learning methods that require no label. Li et al. (2020) developed a semi-supervised deep learning method for robust quality monitoring using metal AM images. However, it still requires manual efforts to label partial data. ...
Full-text available
With the rapid development of additive manufacturing (AM) technology, quality inspection has become one of the most crucial research topics in additive manufacturing. Although numerous image-based deep learning methods have been successfully developed to monitor and inspect AM product quality effectively, many require substantial labels in order to achieve satisfactory training, which is often impractical in real-life AM processes. In this article, a novel image feature-based self-supervised learning (IFSSL) model is proposed for effective quality inspection in AM. Through a feature-based image fusion approach based on defect-relevant feature extraction, the IFSSL model is able to guide machine vision to focus on highlighted defect-relevant regions in the AM product image. In addition, the defect-relevant features are used to generate pseudo-labels for self-supervised learning. With self-supervision, the IFSSL model leverages the advantages of supervised learning and unsupervised learning by requiring no sample label while retaining defect-relevant information. The effectiveness of the proposed IFSSL method is demonstrated through a real case study of fused deposition modeling product image dataset. Results show that the IFSSL model can guide machine vision to pay more attention to potential defective regions, enabling it to detect and locate faults effectively and automatically for machine vision guided quality inspection.
... These models can be used in AM DTs for quality control and inspection, predictive maintenance, process monitoring, simulation and virtual prototyping and automated surface reconstruction. In addition, other types of neural networks can be used in combination with CNNs to improve the performance of the model [20]. ...
Digital Twin (DT) is a virtual replica of a physical system with two-way data communication between the cyber and physical domains. It is an innovative concept in smart manufacturing, envisioning improved efficiency, productivity, flexibility and quality control through real-time data-driven and collaborative digital systems. DT systems are enabled through technologies like Artificial Intelligence (AI), Human Machine Interface (HMI), simulation models, AR/VR, online sensing, big data analytics and the Internet of Things (IoT). Additive Manufacturing (AM) is a fast-growing and revolutionary technology with immense applications in aerospace, biomedical, automobile and marine applications. Some of the fundamental challenges of additive manufacturing processes are part irregularities, indispensable post-processing, unanticipated process anomalies, design and metrology disintegration and lack of standardization. Currently, considerable time and cost are being devoted to offline metrology, defect identification, and process optimization. DT-driven additive manufacturing (AM) is still in its nascent stage; however, the technology has demonstrated its immense potential to transform the emerging world of additive manufacturing. A DT system looks to systematically address these fundamental shortcomings through computational intelligence and real-time data communication. Key application domains include anomaly detection, online condition monitoring, feed-forward process control, intelligent post-processing and process optimization. The chapter critically analyses the current state-of-the-art of DTs in additive manufacturing processes and further discusses its future prospects and research directions. The chapter highlights the capacity of a DT to broaden the acceptability of additive manufacturing in various industrial applications by improving robustness and efficiency.
Conference Paper
Edge-enabled Deep Learning (DL) solutions for Predictive Quality Inspection (PQI) of products in Industry 4.0 are mostly designed for static manufacturing environments. In general, modern manufacturing processes are dynamic in nature. In this context, continual learning-based model retraining accommodates the dynamism for PQI of multiple processes (tasks) using a single DL model. However, the impact of the task ordering in sequentially arriving tasks and solution to reduce this impact on the overall PQI is yet to be solved. To this end, a novel mechanism using a light-weight similarity analysis module is introduced in the quality prediction system at the resource-limited edge. Sequential training of tasks above a similarity threshold (γ) is preferred, and dissimilar tasks are overlooked to train a separate model. This enables a PQI system to hover over training efficiency and model sustainability. The experimental results validate the impact of task order and the effectiveness of the proposed similarity-based analysis to reduce this impact by 70% on the model's overall performance in the real-world use case of plastic bricks.
The process health monitoring of wire arc additive manufacturing (WAAM) is significant for product quality. Most existing additive manufacturing process monitoring is based on image data such as temperature and spatters. However, these monitoring methods do not reflect status information promptly. Moreover, the issue of limited cross-domain diagnosis generalization ability is faced by traditional neural networks for health state discrimination. To address these issues, this work puts forward a bi-classifier and orthogonal constraints jointly guided domain adaptation method based on acoustic emission signal for wire arc additive manufacturing health monitoring. Specifically, we first build a min-max optimization strategy using bi-classifier discrepancy loss to achieve feature adaptation of different domains. Meanwhile, the orthogonal loss increases the dispersion of inter-class features and the aggregation of intra-class features. Moreover, a non-local module is attached for obtaining the remote dependence relationship between sample pixels of acoustic emission signal. Finally, based on the acoustic emission signals from the WAAM process, the performance of the method is evaluated, and the comprehensive results prove its effectiveness and superiority.
Full-text available
In robotic GMAW-based additive manufacturing, the surface evenness of the deposited layer was significant to the dimensional accuracy and the stable fabrication process, and it was determined by the multi-bead overlapping distance. To obtain the optimal overlapping distance, a group of two-bead overlapping experiments was conducted with different overlapping ratio. The cross-section shape was observed and the variation of the bead profile caused by the damming up of the previous bead was investigated. The second bead profile could be fitted by a rotated varying parabola or circular arc function with the decreasing of the overlapping distance from the initial single bead width (w) to 0. A varying cross-section profile overlapping model was developed based on the actual forming characteristics of the overlapping experiment, through which the varying profile of two overlapping beads with arbitrary distance could be predicted. Then, the optimal overlapping distance was calculated under some principles to achieve a relatively flat top surface and stable overlapping process, and the multi-bead overlapping experiments were performed to validate the model. The results showed that the model could achieve an excellent approximation to the actual overlapping experiment, and the good surface evenness and stable overlapping process was obtained, which was significant to the research into the appearance optimization in GMAW-based additive manufacturing.
Full-text available
The extraction of meaningful features from the monitoring of laser processes is the foundation of new non-destructive quality inspection methods for the manufactured pieces, which has been and remains a growing interest in industry. We present ConvLBM, a novel approach to monitor Laser Based Manufacturing processes in real-time. ConvLBM uses a Convolutional Neural Network model to extract features and quality indicators from raw Medium Wavelength Infrared coaxial images. We demonstrate the ability of ConvLBM to represent process dynamics, and predict quality indicators in two scenarios: dilution estimation in Laser Metal Deposition, and location of defects in laser welding processes. Obtained results represent a breakthrough in the 3D printing of large metal parts, and in the quality control of welding processes. We are also releasing the first large dataset of annotated images of laser manufacturing.
Full-text available
Single image dehazing has been a challenging problem which aims to recover clear images from hazy ones. The performance of existing image dehazing methods is limited by hand-designed features and priors. In this paper, we propose a multi-scale deep neural network for single image dehazing by learning the mapping between hazy images and their transmission maps. The proposed algorithm consists of a coarse-scale net which predicts a holistic transmission map based on the entire image, and a fine-scale net which refines dehazed results locally. To train the multi-scale deep network, we synthesize a dataset comprised of hazy images and corresponding transmission maps based on the NYU Depth dataset. In addition, we propose a holistic edge guided network to refine edges of the estimated transmission map. Extensive experiments demonstrate that the proposed algorithm performs favorably against the state-of-the-art methods on both synthetic and real-world images in terms of quality and speed.
Full-text available
AM, generally known as 3D printing, is a promising technology. Robotic AM enables the direct fabrication of products possessing complex geometry and high performance without extra support structures. Process planning of slicing and tool path generation has been a challenging issue due to geometric complexity, material property, etc. Simple and robust planar slicing has been widely researched and applied. However, support structures usually result in time-consuming and cost-expensive. Notwithstanding multi-direction slicing and non-planar slicing (curved layer slicing) have been proposed respectively to decrease support structures, capture some minute but critical features and improve the surface quality and part strength. There is no slicing method aiming at features of part’s sub-volumes. A comprehensive literature review is given first to illustrate the problems and features of available slicing methods better. Then, in order to combine the merits of planar and non-planar slicing to realize intelligent manufacturing further, this paper reports the concept and implementation of a mixed-layer adaptive slicing method for robotic AM. Different from applying planar slicing in any cases or adopting the decomposing and regrouping based multi-direction planar slicing for finding the optimal slicing directions, the proposed method mainly focuses on how to apply planar and non-planar slicing for each sub-volume according to the geometrical features. Additionally, the requirements for robotic AM equipment in possessing multi-mode of printing and slicing are investigated.
Despite the recent advances of intelligent data-driven fault diagnosis methods on rotating machines, balanced training data for different machine health conditions are assumed in most studies. However, the signals in machine faulty states are usually difficult and expensive to collect, resulting in imbalanced training dataset in most cases. That significantly deteriorates the effectiveness of the existing data-driven approaches. This paper proposes a deep learning-based fault diagnosis method to address the imbalanced data problem by explicitly creating additional training data. Generative adversarial networks are firstly used to learn the mapping between the distributions of noise and real machinery temporal vibration data, and additional realistic fake samples can be generated to balance and further expand the available dataset afterwards. Through experiments on two rotating machinery datasets, it is validated that the data-driven methods can significantly benefit from the data augmentation, and the proposed method offers a promising tool on fault diagnosis with imbalanced training data.
In the recent years, data-driven machinery fault diagnostic methods have been successfully developed, and the tasks where the training and testing data are from the same distribution have been well addressed. However, due to sensor malfunctions, the training and testing data can be collected at different places of machines, resulting in the feature space with significant distribution discrepancy. This challenging issue has received less attention in the current literature, and the existing approaches generally fail in such scenarios. This paper proposes a domain adaptation method for machinery fault diagnostics based on deep learning. Adversarial training is introduced for marginal domain fusion, and unsupervised parallel data are explored to achieve conditional distribution alignments with respect to different machine health conditions. Experiments on two rotating machinery datasets are carried out for validations. The results suggest the proposed method is promising to address the fault diagnostic tasks with data from different places of machines, further enhancing applicability of data-driven methods in real industries.
Rotating machinery fault diagnosis problems have been well addressed when sufficient supervised data of the tested machine are available using the latest data-driven methods. However, it is still challenging to develop effective diagnostic method with insufficient training data, which is highly demanded in real industrial scenarios since high-quality data are usually difficult and expensive to collect. Considering the underlying similarities of rotating machines, data mining on different but related equipments potentially benefit the diagnostic performance on the target machine. Therefore, a novel transfer learning method for diagnostics based on deep learning is proposed in this paper, where the diagnostic knowledge learned from sufficient supervised data of multiple rotating machines is transferred to the target equipment with domain adversarial training. The experimental results on four datasets validate the effectiveness of the proposed method, and show it is feasible and promising to explore different datasets to improve diagnostic performance.
Additive manufacturing (AM) is considered as a revolution in manufacturing. However, the high expectations face technical difficulties that prevent further penetration into wider industries. The main reason is the lack of process reproducibility and the absence of a reliable and cost-effective process monitoring. This paper is a supplement to existing studies in this field and proposes a unique combination of highly sensitive acoustic sensor and machine learning for process monitoring. The acoustic signals from a real powder-bed fusion AM process were collected using a fiber Bragg grating. The process parameters are intentionally tuned to achieve three levels of quality categories, which are related to the porosity contents inside the workpiece. The quality categories are defined as high, medium, and poor quality and their corresponding porosity contents are 0.07%, 0.30%, and 1.42%, respectively. Wavelet spectrograms of the signals and their encoded label representations, obtained from spectral clustering, are taken as features. A deep convolutional neural network is used to classify the features from each category and the classification accuracy ranges between 78% and 91%. Hence, the proposed method has significant industrial potentials for in situ and real-time quality monitoring of AM processes since it requires minimum modifications of commercially available industrial machines.
Recent years have witnessed increasing popularity and development of deep learning spanning through various fields. Deep networks, and in particular convolutional neural network (CNN) have also achieved many state-of-the-art competition results in the intelligent fault diagnosis of mechanical systems. However, most of the existing studies have been performed with the assumption that the same distribution holds for both the training data and the test data, which is not in accord with situations in real diagnosis tasks. To tackle this problem, a transfer learning framework based on pre-trained CNN, which leverages the knowledge learned from the training data to facilitate diagnosing a new but similar task, is presented in this work. First, the CNN is trained on large datasets to learn the hierarchical features from the raw data. Then, the architecture and weights of the pre-trained CNN are transferred to new tasks with proper fine-tuning instead of training a network from scratch. To adapt the pre-trained CNN in a specific case, three transfer learning strategies are discussed and compared to investigate the applicability as well as the significance of feature transferability from the different levels of a deep structure. The case studies show that the proposed framework can transfer the features of the pre-trained CNN to boost the diagnosis performance on unseen machine conditions in terms of diverse working conditions and fault types.
Purpose Part quality inspection is playing a critical role in the metal additive manufacturing (AM) industry. It produces a part quality analysis report which can be adopted to further improve the overall part quality. However, the part quality inspection process puts heavy reliance on the engineer’s background and experience. This manual process suffers from both low efficiency and potential errors and, therefore, cannot meet the requirement of real-time detection. The purpose of this paper is to look into a deep neural network, Convolutional Neural Network (CNN), towards a robust method for online monitoring of AM parts. Design/methodology/approach The proposed online monitoring method relies on a deep CNN that takes a real metal AM part’s images as inputs and the part quality categories as network outputs. The authors validate the efficacy of the proposed methodology by recognizing the “beautiful-weld” category from material CoCrMo top surface images. The images of “beautiful-weld” parts that show even hatch lines and appropriate overlaps indicate a good quality of an AM part. Findings The classification accuracy of the developed method using limited information of a small local block of an image is 82 per cent. The classification accuracy using the full image and the ensemble of model outputs is 100 per cent. Originality/value A real-world data set of high resolution images of ASTM F75 I CoCrMo-based three-dimensional printed parts (Top surface images with magnification 63×) annotated with categories labels. Development of a CNN-based classification model for the supervised learning task of recognizing a “beautiful-weld” AM parts. The classification accuracy using the full image and the ensemble of model outputs is 100 per cent.