Additive manufacturing (AM) has emerged as a disruptive digital manufacturing technology. However, its broad adoption in industry is still hindered by high entry barriers of design for additive manufacturing (DfAM), limited materials library, various processing defects, and inconsistent product quality. In recent years, machine learning (ML) has gained increasing attention in AM due to its unprecedented performance in data tasks such as classification, regression and clustering. This article provides a comprehensive review on the state-of-the-art of ML applications in a variety of AM domains. In the DfAM, ML can be leveraged to output new high-performance metamaterials and optimized topological designs. In AM processing, contemporary ML algorithms can help to optimize process parameters, and conduct examination of powder spreading and in-process defect monitoring. On the production of AM, ML is able to assist practitioners in pre-manufacturing planning, and product quality assessment and control. Moreover, there has been an increasing concern about data security in AM as data breaches could occur with the aid of ML techniques. Lastly, it concludes with a section summarizing the main findings from the literature and providing perspectives on some selected interesting applications of ML in research and development of AM.

Machine learning in additive manufacturing: State-of-the-art and perspectives

Additive manufacturing (AM), also known as 3D printing, has been increasingly adopted in the aerospace, automotive, energy, and healthcare industries over the past few years. While AM has many advantages over subtractive manufacturing processes, one of the primary limitations of AM is surface integrity. To improve the surface integrity of additively manufactured parts, a data-driven predictive modeling approach to predicting surface roughness in AM is introduced. Multiple sensors of different types, including thermocouples, infrared temperature sensors, and accelerometers, are used to collect temperature and vibration data. An ensemble learning algorithm is introduced to train the predictive model of surface roughness. Features in the time and frequency domains are extracted from sensor-based condition monitoring data. A subset of these features is selected to improve computational efficiency and prediction accuracy. The predictive model is validated using the condition monitoring data collected from a set of AM tests conducted on a fused filament fabrication (FFF) machine. Experimental results have shown that the proposed predictive modeling approach is capable of predicting the surface roughness of 3D printed components with high accuracy.

Prediction of surface roughness in extrusion-based additive manufacturing with machine learning

Data and knowledge mining with big data towards smart production

Machine vision intelligence for product defect inspection based on deep learning and Hough transform

The powder bed fusion additive manufacturing process enables fabrication of metal parts with complex geometry and elaborate internal features, the simplification of the assembly process, and the reduction of development time; however, its tremendous potential for widespread application in industry is hampered by the lack of consistent quality. This limits its ability as a viable manufacturing process particularly in the aerospace and medical industries where high quality and repeatability are critical. A variety of defects, which may be initiated during powder bed fusion additive manufacturing, compromise the repeatability, precision, and resulting mechanical properties of the final part. One approach that has been more recently proposed to try to control the process by detecting, avoiding, and/or eliminating defects is online monitoring. In order to support the design and implementation of effective monitoring and control strategies, this paper identifies, analyzes, and classifies the common defects and their contributing parameters reported in the literature, and defines the relationship between the two. Next, both defects and contributing parameters are categorized under an umbrella of manufacturing features for monitoring and control purposes. The quintuple set of manufacturing features presented here is meant to be employed for online monitoring and control in order to ultimately achieve a defect-free part. This categorization is established based on three criteria: (1) covering all the defects generated during the process, (2) including the essential contributing parameters for the majority of defects, and (3) the defects need to be detectable by existing monitoring approaches as well as controllable through standard process parameters. Finally, the monitoring of signatures instead of actual defects is presented as an alternative approach to controlling the process “indirectly.”

Common defects and contributing parameters in powder bed fusion AM process and their classification for online monitoring and control: a review

The objective of this work is to identify failure modes and detect the onset of process anomalies in additive manufacturing (AM) processes, specifically focusing on fused filament fabrication (FFF). We accomplish this objective using advanced Bayesian nonparametric analysis of in situ heterogeneous sensor data. Experiments are conducted on a desktop FFF machine instrumented with a heterogeneous sensor array including thermocouples, accelerometers, an infrared (IR) temperature sensor, and a real-time miniature video borescope. FFF process failures are detected online using the nonparametric Bayesian Dirichlet process (DP) mixture model and evidence theory (ET) based on the experimentally acquired sensor data. This sensor data-driven defect detection approach facilitates real-time identification and correction of FFF process drifts with an accuracy and precision approaching 85% (average F-score). In comparison, the F-score from existing approaches, such as probabilistic neural networks (PNN), naive Bayesian clustering, support vector machines (SVM), and quadratic discriminant analysis (QDA), was in the range of 55–75%.

Online Real-Time Quality Monitoring in Additive Manufacturing Processes Using Heterogeneous Sensors

Image analysis-based closed loop quality control for additive manufacturing with fused filament fabrication

Online non-contact surface finish measurement in machining using graph theory-based image analysis

The objective of this work is to identify failure modes and detect the onset of process anomalies in Additive Manufacturing (AM) processes, specifically focusing on Fused Filament Fabrication (FFF). We accomplish this objective using advanced Bayesian non-parametric analysis of in situ heterogeneous sensor data. The proposed method can ultimately lead to intelligent decision making and closed loop control in AM processes. Experiments are conducted on a desktop FFF machine (MakerBot Replicator 2X) instrumented with a heterogeneous sensor array including thermocouples, accelerometers, an infrared temperature sensor, and a real-time miniature video borescope. FFF process failures are detected online using the non-parametric Bayesian Dirichlet Process (DP) mixture model and evidence theory based on the experimentally acquired sensor data. This sensor data-driven defect detection approach facilitates real-time identification and correction of FFF process drifts with an accuracy and precision approaching 85% (average F-score). In comparison, the F-score from existing approaches, such as Probabilistic Neural Networks, Naive Bayesian Clustering, Support Vector Machines, and Quadratic Discriminant Analysis was in the range of 55% to 75%.Copyright © 2015 by ASME

Sensor-Based Online Process Fault Detection in Additive Manufacturing

A model of the laser powder deposition (LPD) process is presented, which predicts the cross-sectional geometry of parts that are made up of thin-walled and thick-walled features, deposited via multiple passes. The model builds up the part shape incrementally by predicting the cross-section of a bead of material deposited on the part, updating part shape to reflect the added material, and repeating for each additional deposition pass. The effects of laser power and deposition speed are accounted for empirically, and the effect of nozzle stand-off distance is accounted for via a powder catchment model suitable for coaxial deposition nozzles. The model was calibrated via deposition experiments using stainless steel 316L powder and via measurement of nozzle characteristics. Validation tests showed that the powder catchment model captured the effect of nozzle stand-off distance on deposited bead size. Validation tests also showed that the model predicted the overall shapes of both thin-walled and thick-walled features, including rounding present at the edges of some thick-walled features. Using calibration data from short thick-walled depositions, the average error in predicted feature height, after ten layers, was 9.3% and 9.5% for thin-walled and thick-walled features, respectively. The model was also shown to predict the effects of using a step-up distance per layer that is too small, resulting in inefficient deposition, or too large, resulting in deposition failure after a few layers.© 2016 ASME

David Roberson

Papers

Online Real-Time Quality Monitoring in Additive Manufacturing Processes Using Heterogeneous Sensors

Image analysis-based closed loop quality control for additive manufacturing with fused filament fabrication

Online non-contact surface finish measurement in machining using graph theory-based image analysis

Sensor-Based Online Process Fault Detection in Additive Manufacturing

Simulation of Cross-Sectional Geometry During Laser Powder Deposition of Tall Thin-Walled and Thick-Walled Features