Advances in Online Detection Technology for Laser Additive Manufacturing: A Review

Abstract

In additive manufacturing (AM), the mechanical properties of manufactured parts are often insufficient due to complex defects and residual stresses, limiting their use in high-value or mission-critical applications. Therefore, the research and application of nondestructive testing (NDT) technologies to identify defects in AM are becoming increasingly urgent. This article reviews the recent progress in online detection technologies in AM, a special introduction to the high-speed synchrotron X-ray technology for real-time in situ observation, and analysis of defect formation processes in the past 5 years, and also discusses the latest research efforts involving process monitoring and feedback control algorithms. The formation mechanism of different defects and the influence of process parameters on defect formation, important parameters such as defect spatial resolution, detection speed, and scope of application of common NDT methods, and the defect types, advantages, and disadvantages associated with current online detection methods for monitoring three-dimensional printing processes are summarized. In response to the development requirements of AM technology, the most promising trends in online detection are also prospected. This review aims to serve as a reference and guidance for the work to identify/select the most suitable measurement methods and corresponding control strategy for online detection.

Highlights

• (1)This review includes all the important contributions in the current area of interest, especially online detection technology and process monitoring and feedback control algorithms.
• (2)The formation mechanism of different defects and the influence of process parameters on defect formation, important parameters such as defect spatial resolution, detection speed, and scope of application of common nondestructive testing methods, and the defect types, advantages, and disadvantages associated with current online testing methods for monitoring three-dimensional printing processes are summarized.
• (3)This review especially introduced the high-speed synchronous X-ray technology used for real-time in situ observation and analysis of defect formation processes in the past 5 years.
• (4)We discussed the significant impact that machine learning and artificial intelligence have had on the online monitoring technology of additive manufacturing with respect to advancing the use of algorithms for process control in additive manufacturing.
• (5)We provided a prospective discussion of the most important trends in additive manufacturing that are set to drive the development of nondestructive testing methods in the near future.

Introduction

Additive manufacturing (AM), such as three-dimensional (3D) printing and rapid prototyping, is a “bottom-up” modern manufacturing technology based on the discrete-stacking principle. Lasers are used as the energy source for several implementations, including selective laser sintering, selective laser melting (SLM), laser cladding, and laser engineered net shaping (LENS).¹ The application of AM is gradually transforming traditional lifestyles and production methods. Owing to its characteristics of rapidity, customization, digitization, and connectivity, it has been labeled as the core technology of the “third industrial revolution,” with its use becoming widespread in medicine, aerospace engineering, and other important fields.^2,3

Several factors influence the printing quality and reusability of AM-molded structures, such as shape accuracy, size, surface quality, and property control.⁴ At present, roughness, internal porosity, low density, fracturing, and delamination of materials are the most notable defects encountered in the AM process, which can adversely affect the accuracy and reliability of the molded parts, thereby hindering the development and application of AM technology.⁵ In addition, nonequilibrium solidification leaves a large amount of residual stress in metallic structures, which leads to deformation, affecting their size and shape accuracy. Compared with those materials manufactured using traditional mechanical processes, properties such as mechanical strength and durability of AM-molded structures, particularly load-bearing parts, are unsuitable for extensive use.⁶ Consequently, to overcome these shortcomings, quality control is recognized as an important aspect of AM processes and can be improved via high-precision nondestructive testing (NDT).

Recently, there have been many reviews on AM.^7–9 This article reviews and analyzes the online detection technology in laser AM and supplements the latest developments to provide a more comprehensive background reference for condition monitoring and quality control. On the contrary, in recent years, the research on defect detection technology in AM has focused primarily on simulation and offline detection.^10–12 However, the monitoring of defect formation mechanism and its formation process and control lacks intuitive evidence. In situ high-speed X-ray technology has been applied in the in situ monitoring of defects in AM in recent years, which can monitor the formation process of defects in real time and facilitate subsequent control and elimination.^13–19 This technology has not been mentioned in previous related reviews.

Based on a literature review, the scope of this study lies in the identification of the measurement science needs for AM. The purpose of this article is to serve as a background reference and guidance for determining the most suitable measurement methods and corresponding control measures for online inspection of the AM process. The article is structured as follows: In the Common Defect Types and Formation Mechanism in AM section, the formation mechanism of different defects and the influence of process parameters on defect formation in the AM process are briefly summarized. Subsequently, the important parameters such as defect spatial resolution, detection speed, and scope of application of common NDT methods are tabulated. The Signal Types and Online Monitoring of AM Process section reviews recent progress in online detection technologies in AM. Following which, monitoring signal types and sensing technology in the AM process and the advantages and disadvantages associated with current online detection methods for monitoring AM processes are summarized. The Process Monitoring and Feedback Control Algorithms section summarizes the latest research efforts involving process monitoring and feedback control algorithms. Finally, a prospective overview of trends in the development of online detection technologies is provided in the Conclusions section to address the developmental requirements of AM technology.

Common Defect Types and Formation Mechanism in AM

Although AM offers numerous advantages in developing complex workpieces, it is affected by many factors, such as the laser energy input, scanning speed, scanning strategy, powder material, and size. Therefore, defects can occur at every stage of the process. There are four stages in 3D printing: preparation, printing, cooling, and service. Among them, the working conditions of the printing stage are bad and the most difficult to control, and so, it is the most concerned. The types of defects are different, and the content to be detected is also different. During preparation, the raw materials, that is, the powder size, particle shape, and physical and chemical properties, require testing. In the printing stage, the main areas requiring inspection are the molten state, stress state, material properties, holes, residual stress, deformation, and the distortion of parts, whereas in the cooling stage, the geometric shape deviation, part anisotropy, holes, cracks, inclusions, surface and internal defects, and residual stress are inspected. Potential defects in the service stage are surface defects, cracks, and deformations.^20–24

Defect formation mechanism in the AM process

AM involves complex physical processes, including laser energy absorption and transfer, rapid melting and solidification of materials, and microstructure evolution. The splashing, spheroidization, holes, cracks, and residual stress generated in the process are its most common defects and are detrimental to the mechanical and physical properties and the fatigue life of the prefabricated product parts, thereby limiting their application. The formation mechanism of these defects is summarized below.

Splashing

Splashing is also one of the most common defects in the AM process. Splashing is mainly caused by lateral protection airflow, fluctuations in the molten pool, and recoil pressure. Splashes falling on the powder will form larger metal particles, which will result in under-fusion and pore defects.²⁵ Splashes falling on the surface of the solidified layer will affect the powder spreading of the next layer, causing the next powder layer to be uneven and even damage the powder spreading roller.²⁶ To prevent the metal powder bed from being polluted by splashes, a high-speed protective airflow can be used to remove splashes.²⁷ However, excessive airflow will affect the surface quality of the powder layer. Figure 1²⁷ is the splashing phenomenon.

FIG. 1.

Splashing occurs at the interface of the powder molten pool at the front of the molten pool. Arrow indicates: splash direction. (Reprinted with permission. Ref.²⁷ Copyright 2017, Elsevier B.V.) MP, Molten pool.

Spheroidization

As shown in Figure 2,²⁸ spheroidization is a unique metallurgical defect in the manufacturing process of the metal-based powder bed.²⁹ Spheroidization occurs when the liquid metal solidifies into a spherical shape under the action of surface tension. A laser beam energy density too high or too low will cause spheroidization; when the energy is too low, the metal powder will not be completely fused, which will cause spheroidization; when the energy is too high, the liquid metal will splash on the unmelted metal powder to form spheroidization. Spheroidization will affect the laying quality of the next layer and the surface quality of the component; it will also cause defects such as slag inclusion and poor fusion, which will affect the tensile strength and fatigue performance of the component.

FIG. 2.

Example of spheroidization of TNM-B1 titanium aluminide alloy in the SLM process. (Reprinted with permission. Ref.²⁹ Copyright 2014, Elsevier B.V.) SLM, selective laser melting.

Hole defects

Porosity, incomplete fusion holes, and pores are typical hole-related defects in AM workpieces. Porosity is governed predominantly^30–33 by the cooling rate during solidification; holes are formed when the dissolved gas cannot escape from the surface of the molten pool before the solidification is complete. These holes are usually small and approximately spherical. In addition, when a powerful laser is used as the input energy source, the temperature of the molten pool is generally higher, which leads to increased gas solubility and enrichment. Therefore, unmoderated laser energy input and unstable process conditions are the primary reasons for porosity.

Incomplete fusion holes, also known as lack-of-fusion defects, are formed because of a lack of sufficient energy; incomplete powder melting causes a new layer of powder with sufficient overlap to be deposited on the previous layer,^32–34 as shown in Figure 3.³⁵ In other words, when the laser energy input is too low, the width of the molten pool is small, creating incomplete fusion holes due to an insufficient overlap between the scanning tracks or insufficient penetration depth of the molten pool.

FIG. 3.

Optical image of lack-of-fusion defects in an SLM-manufactured part: (a) poor bonding defects and (b) unmelted metal powders. (Reprinted with permission. Ref.³⁵ Copyright 2014, Trans Tech Publications, Ltd.)

In the AM process, the rapid melting and solidification of materials and the violent fluctuation of the molten pool will lead to the formation of pores. According to the formation mechanism of pores, they can be divided into pores related to raw materials and pores caused by laser action. The size, morphology, number, and location of the pores will affect the mechanical properties of the component. A higher porosity will shorten the fatigue life of the molded part, and the pores close to the surface have a greater impact on the fatigue performance of the molded part.^36,37

In reference, Mohammad Hojjatzadeh et al.³⁶ directly observed the formation mechanisms of six different pores in the laser powder bed fusion (LPBF) AM process through in situ high-speed, high-energy X-ray imaging experiments: (1) keyhole-induced pores; (2) pores formed by feedstock powder; (3) pores formed along the melting boundary during the laser melting process caused by the evaporation of volatile substances or the expansion of tiny trapped gases; (4) pores that are trapped by surface fluctuations; (5) when the depression zone is shallow, the pores formed due to the fluctuation of the depression zone; and (6) pores formed by cracks.

Cracks

The formation of cracks in AM process is related to temperature distribution, residual stress, and poor fusion. With a high local-energy input, the powder undergoes rapid melting and solidification, and the cooling rate of the molten pool can reach 10⁸ K/s,³⁸ resulting in a high-temperature gradient and excessive residual thermal stress. This combination often results in cracks forming in prefabricated parts, which typically start from the initial surface attached to the partially melted powder. Cracks reduce the service life of products,^38–40 as shown in Figure 4.³⁹

FIG. 4.

Crack morphology in SLM-manufactured TC4 parts.

In addition, under-fusion formed by poor fusion is also a common type of cracks in AM components, which will have a fatal impact on the mechanical behavior and fatigue life of the components. Under-fusion cracks are mainly caused by the incomplete melting of metal powder, mostly appearing between adjacent scanning welding passes or between deposited layers. Severe cracks may also lead to stratification defects.⁵

Residual stress

As shown in Figures 5 and 6,⁴¹ residual stress is caused by two factors: temperature gradient and phase change.⁴¹ This is the principal mechanism driving the high-temperature gradient between the layers that induces large residual stress in the vertical direction. Phase change occurs during solidification. Applying a high-energy density input causes residual stress mainly in the horizontal direction while simultaneously increasing the stress effected by martensitic transformation. This confirms that the process parameters have a strong influence on the generation and distribution of residual stress. As the laser energy density increases, the vertical and horizontal residual stresses along the layer direction increase, leading to the formation of cracks and reducing product quality.

FIG. 5.

(a) Longitudinal horizontal and (b) transverse horizontal residual stress profiles in the build direction of Ti-6Al-4V samples produced at various laser energy densities. (Reprinted with permission. Ref.⁴¹ Copyright 2020, Elsevier Ltd.)

FIG. 6.

(a) Vertical and (b) horizontal residual stress profiles in the layer direction of Ti-6Al-4V samples produced at various laser energy densities. (Reprinted with permission. Ref.⁴¹ Copyright 2020, Elsevier Ltd.)

The above summarizes the causes of common defects in the AM process and their important influencing factors. Table 1 shows the relative frequency and importance of common defect types. For defect detection and intelligent control in the laser AM process, studying the formation of defects and suppressing the formation mechanism of defects are of great significance to improving the quality of printing and enhancing the service life of components.

Table 1.

The Relative Occurrence Frequency and Importance of Common Defects

Defect type	Influencing factors	Occurrence frequency	Harmfulness
Splashing	Laser power, spot size, energy density, scanning speed, scanning distance, molten pool instability.	One of the most common defects.	It directly affects the interaction between the laser and the material, leading to other defects and reducing the tensile strength and fatigue performance of the component.
Spheroidization	Scanning speed, laser power, powder layer thickness.	Metallurgical defects peculiar to metal-based powder bed manufacturing processes.	Increasing the porosity and surface roughness of the components and seriously affects the quality of the components.
Pores	Laser power, energy density, scanning speed and direction, powder layer thickness, and powder size.	Common defects in the laser powder bed additive manufacturing process.	The pores have the greatest impact on the mechanical properties of the components and severely reduce the fatigue life of the components.
Cracks	Laser power, energy density, temperature gradient, scanning speed.	Occasionally	Cracks have a fatal effect on the components and reduce the service life of the components.
Residual stress	Temperature gradient and phase change.	Has always existed and may evolve.	Higher residual stresses can lead to deformation, geometric changes, and microcrack formation.

The influence of process parameters on defect formation

AM relies on several process parameters, which can be divided into four categories: laser related, scanning related, powder related, and temperature related. Some of these factors can be predetermined, while others are generated during the process and cannot be predetermined. Figure 7 presents a block diagram of the process parameters involved in the AM process. Among these, the laser energy input, powder materials, and scanning strategy are the main factors that affect the formation of defects.

FIG. 7.

Influencing factors involved in the AM process. AM, additive manufacturing.

Influence of laser energy input

The laser energy input directly determines the melting conditions of the metal powder and has a significant impact on the type and size of defects in the AM process. The input of energy in the material is mainly related to process parameters such as laser power, scanning speed, hatch spacing, and layer thickness; at lower scanning speed and higher laser power, the energy input is higher, and more powder is melted at a higher temperature, resulting in pore defects. In addition, the low melting point components in the alloy, such as aluminum and magnesium, may evaporate into gas and form bubbles. During the rapid solidification process, the bubbles failed to escape from the molten pool in time and stayed in the molten pool, forming spherical pore defects.^34,42 On the contrary, when the energy input is high, the molten pool becomes larger, causing the powder around the molten pool to be eroded. The denudation process leads to insufficient molten metal to fill the gap between adjacent tracks, thereby forming a large porosity.³⁰

In addition, the relatively low scanning speed and high-energy input may result in high residual thermal stress during rapid melting and solidification. The higher the energy input, the more severe the shrinkage of the molten metal during solidification. High residual stresses are generated during the solidification process.^38,39,43 Under higher scanning speed and lower laser power, the input energy is too low to completely melt the powder, which will lead to the formation of incomplete melting defects. Too large a powder thickness will result in insufficient penetration of laser energy input, and an effective overlap between layers may not be formed, resulting in the formation of incomplete fusion defects between layers.^32,34,43,44

Influence of powder materials

The morphology and size of the powder have a great influence on the flatness and fluidity of the powder bed. Therefore, there are strict requirements on the morphology and size of the powder in the AM process. There are various powder preparation methods, which will have different effects on the formation of defects. Smaller size powders are easier to reduce the porosity of prefabricated parts. In addition, the gas contained in the powder also increases the possibility of defect formation. Literature⁴⁵ studied the influence of different powder sizes of 316L stainless steel on the quality of parts during the SLM process. The report showed that smaller metal powders could reduce the porosity of manufactured parts, and the relative density of the average particle size of 26.36 μm can reach 99.75%. Literature⁴⁰ studied the densification behavior of gas-atomized and water-atomized 316L stainless steel powder. The results showed that compared with water-atomized powder, the parts made by gas-atomized powder had a higher relative density and lower porosity.

Influence of scanning strategy

The different scanning strategies directly affect the heat transfer, melting, and solidification of the powder and ultimately affect the location and distribution of defects. Three different scanning methods are usually used in the AM process: “unidirectional,” “zigzag,” and “cross-hatching.” For unidirectional and zigzag scanning methods, the laser power is unstable at the beginning and the end of the scanning track and the scanning speed is gradually reduced, often leading to relatively high laser energy input, thereby forming defects.^46,47 In the actual processing process, incomplete fusion defects are also prone to occur between the scanning track and the layer.^48,49 The cross-hatching method can make the laser energy input more balanced in the entire layer, effectively avoiding the accumulation and propagation of defects.

Influence of temperature gradient

The temperature gradient is an important cause of cracks and residual stress, and temperature history plays a vital role in the change of residual stress and the structural direction of density and material phase structure. The cracks formed by residual stress can be divided into solidification cracks and liquefaction cracks. The solidification cracks are caused by the large temperature gradient between the molten pool and the solidified metal, which leads to greater deformation of the molten pool. However, the fluidity of the liquid is insufficient to supplement the deformation caused by the molten pool. In addition, Hojjatzadeh et al.¹⁷ combined high-speed synchrotron radiation X-ray and tracer labeling technology and found that the movement behavior of the pores in the molten pool is controlled by the competition between the thermal capillary force caused by the temperature gradient and the resistance caused by the melt flow. Moreover, for the first time, an essential method of eliminating pores is proposed, that is, by adjusting the 3D printing process parameters to obtain a high-temperature gradient in the molten pool, the thermal capillary force caused by it drives and discharges the pores, to obtain 3D-printed metal components without pore defects.

Residual stress has a huge impact on the service life of 3D-printed components; it is most concentrated at the connection between the part and the substrate. There is large compressive stress at the center of the part and large tensile stress at the edge. Therefore, the methods to reduce the residual stress mainly include the following:

• (1)Add support structures because they are warmer than the substrate alone. Once the part is removed from the substrate, the residual stress will be released, but the part may be deformed in the process.⁵⁰
• (2)To control temperature fluctuations, the length of the scanning vector can be reduced instead of continuous laser scanning. Rotating the orientation of the scan vector based on the largest section of the part may work.⁵¹
• (3)Heat the substrate and material before printing. Due to the lower operating temperature, preheating is more common in electron beam melting processes than SLM or directed energy deposition (DED) processes.⁵²

Online detection and process monitoring requirements for AM technology

AM is a discrete, stacked manufacturing method, which is essentially different from traditional “subtractive” manufacturing.⁵³ In AM, the production cost of workpieces is high, as it requires online monitoring and control as well as a quality audit of the printed parts. Therefore, AM workpieces cannot be effectively tested for defects using traditional destructive methods; they require a unique, nondestructive method selected based on their individual characteristics to help standardize the process and improve the overall part quality.

NDT technologies must be compatible with the materials, design, and testing methods used in the AM process, including the implementation and optimization of process testing, product quality evaluation, and quality monitoring during service. They are also expected to last throughout the life cycle of the materials. Based on the detection principle, NDT technologies can be divided into (1) imaging, for example, computed tomography (CT) testing, (2) ultrasonic, for example, laser ultrasonic testing, (3) electromagnetic, for example, eddy current testing, and (4) thermal imaging, for example, infrared thermal imaging. In terms of inspection, NDT can be divided into online and offline inspection.

Online inspection forms a crucial part of the production process; high real-time performance enables rapid feedback of process information to the control system, thereby facilitating system decision-making and modification. Offline detection has a relatively high lag and is often unable to form a closed-loop feedback control system. However, offline detection is usually more accurate and can perform comprehensive detection; it could serve as an important benchmark or supplement online detection methods. Some important parameters, advantages, and disadvantages of common NDT techniques are described in Table 2.

Table 2.

Common Nondestructive Testing Techniques

Detection method	Defect spatial resolution	Detection speed	Penetration depth	Application scope	Online detection?	Advantages (A) and disadvantages (D)
Ultrasonic testing	100 μm54	9 m/min	1 mm to 3.5 m	Suitable for area-type defect detection in metals, nonmetals, and composite materials.55	Yes	A: Strong penetration, surface- and internal-level defect detection, accurate positioning of defects, fast detection. D: Requires a coupling agent for detection; limitations in workpiece and defect size and shape affect sensitivity.56,57
Ray detection	25 μm	1 min	5 mm	Suitable for detecting internal volumetric defects with relatively low thickness.55	Yes	A: Intuitive defect display, surface and internal defect detection. D: Workpiece thickness, defect position, and defect orientation affect sensitivity; low detection speed; causes radiation.57,58
Magnetic particle inspection	0.1 μm (crack width)	10 min	6 mm weld defect depth	Suitable for detecting defects such as small cracks close to the surface of ferromagnetic materials.55	No	A: Intuitive display, easy operation, low cost. D: Workpiece requires high surface smoothness, detection limited to surface and near-surface defects of ferromagnetic materials, difficulty in detecting deep holes.55
Penetration testing	1 μm (crack width)	10 min	1–2 mm	Suitable for detecting surface-opening defects. Works best at ambient temperatures of 10–50°C.55	No	A: Intuitive display, easy operation, low cost. D: Detection limited to surface-opening defects, workpiece requires high surface smoothness, its unquantifiable results, unsuitable for porous materials.57
Eddy current testing	100 μm59	20 m/min	1–2 mm	Suitable for automatically detecting defects on and near the surfaces of conductive materials in high-temperature environments.57	Yes	Noncontact, no coupling agent, fast detection speed; however, only conductive materials can be detected, which are affected by surface roughness and edge effects, and limited by complex-shaped components.57
Magnetic flux leakage testing	1 mm60	Pipe 5 m/s61	4–32 mm of pipe wall thickness61	Suitable for detecting surface and subsurface defects of ferromagnetic materials.62	Yes	High detection efficiency, good reliability, easy to realize automation; however, it can only detect surface and near-surface defects of ferromagnetic materials, not applicable to the detection of complex components, irregular defects, and fatigue cracks with very narrow cracking.62
Acoustic emission testing	10 μm crack propagation	Sampling frequency 10 m/s	n/a/	Suitable for real-time analysis of dynamic damage characteristics and other material information.55	Yes	Overall monitoring, dynamic measurement, very sensitive to materials, but poor sensitivity to geometric shapes; easy to pick up the signal, but the noise is large; defects cannot be qualitatively and quantitatively defined accurately.63
Computed tomography imaging	0.5 μm58	0.2 s/layer	5 mm	Suitable for internal defect detection and thickness analysis of complex structures.57	No	Not limited by geometry and offers high precision; however, it is not applicable to the field detection of planar thin-plate components and large components and has low detection efficiency and high cost.57,64
Infrared thermal imaging testing	1 μm wide,65 cracks at a depth of 3 mm66	/	0.5–3 mm	Suitable for online detection of surface and near-surface defects in high-temperature environments.55	Yes	Overall monitoring, noncontact, no pollution, fast detection, real-time, efficient, intuitive, high resolution, and high sensitivity; however, it requires a relatively large camera data processing capacity for image acquisition.55

Owing to the particularity of AM technology and the time lag associated with offline detection technology, existing offline detection methods can be used neither for part inspection nor for in situ inspection in the AM process. Therefore, the development of online detection technology is a hot topic within the AM research community. Online detection can monitor the entire laser–substance interaction, extracting a wealth of useful information. In particular, online detection can provide crucial insights regarding the scientific issues in AM technology as well as help to establish process monitoring.

Signal Types and Online Monitoring of AM Process

Improving the quality of additively manufactured products is a challenging task, mainly because of the lack of key experimental data and mathematical and statistical models that provide a basis for the control process. These data are derived through monitoring the state of the melting process. The key issue concerning workpiece condition monitoring in AM is the identification of the correlation between process parameters, process characteristics, and product quality. These parameters can be used to inform a monitoring program that provides real-time control of the AM process. Here, we define the process parameters, which determine the rate of energy transfer to the powder surface and the interaction of the energy with the materials, as “inputs.” The process parameters are divided into controllable (e.g., laser power and scanning speed) and predefined material parameters (e.g., powder size and distribution). Process characteristics refer to the dynamics of the powder heating, melting, and solidification processes, which can be categorized as observable (direct observation, such as the shape and temperature of the molten pool) or derivable (determined by analysis and simulation, such as defects and residual stress); both observable and derivable characteristics can directly affect the molding quality of the final product (e.g., geometric size, physical properties, and mechanical properties).

Owing to the different physical or chemical properties of the raw materials used in AM, or certain characteristics of the part structure, it is necessary to select detection methods that are sensitive to these characteristics. The AM process involves complex physical processes, such as laser absorption, heating, melting, evaporation, recoil, piston effects, plasma formation, laser-supported absorption waves, Marangoni convection, and Kelvin–Helmholtz instability.⁶⁷ These complex interactions make the stability of the melting process difficult to control, with extreme environments such as high temperature and arc interference further complicating the task. A variety of signals are generated during these interactions, including sound, vibrational, heat, and light waves. More specifically, the five most frequently encountered signal types are infrared light emitted from the molten pool, visible and ultraviolet light emitted from the plasma plume, audible sound propagating in the air, ultrasonic waves propagating in the solid structure, and electric signal in the plasma plume. To overcome these challenges, many recent studies have been devoted to researching AM process monitoring and control with the aim of improving part quality. Figure 8 shows the types of signal monitoring and sensing technology in AM.

FIG. 8.

Process monitoring and sensing technology classification in AM.

Infrared signal online monitoring

Coaxial detection architecture

Infrared signals can directly reflect changes in the molten pool state, making infrared detection a common feature of current monitoring methods.^68–78 Elsewhere, Lott et al.⁷⁰ used a coaxial optical path, dichroic mirror, and beam splitter to build an optical system for monitoring high-scan speed and molten-pool flow dynamics. This system collected comprehensive visible and infrared light information from the molten pool area.

A similar approach was followed by Kruth and Mercelis^72,73 to develop an available method to obtain relevant information from the SLM process. Their method is based on a coaxial optical path and uses CCD cameras and photodiodes to monitor the entire melting process within the SLM system area, while they also designed a feedback control system to monitor the temperature distribution within the molten pool.^72,74,75 Moreover, the Kruth research team extracted the molten pool information along the horizontal plane to detect the thermal stress and deformation of the overhanging structure caused by overheating. Thus, the molten pool process inspection device enabled the area, length, and width of the molten pool to be recorded for the SLM process, thereby clarifying that the change of state in the molten pool is related to the internal porosity of the molded part.⁷⁹

Bernhard et al.^80,81 used a high-speed, single-point thermometer and a photodiode coaxial optical path to obtain a molten pool temperature mean tomogram and near-infrared radiation, based on temperature or radiation abnormalities, and developed scripts for visualization, real-time monitoring print quality. The results showed that the method could detect voids and hole defects. In particular, artifacts related to changes in laser power and feed rate are identifiable, and the correlation between visualization and errors detected by microscopy is possible. This technology can be used as a supplement to data visualization and quality control systems in AM. Hooper⁸² took the lead in using two ultrahigh-speed cameras to build a dual-band colorimetric temperature measurement system based on a coaxial optical path on SLM equipment in 2018, with high time resolution (image acquisition speed reached 100 KHz), using this system to characterize the temperature distribution and cooling rate of the molten pool. This method provides a new tool for optimizing scanning strategies and parameters, identifying the causes of defect-prone parts, and controlling the cooling rate of local tissue development.

Paraxial detection architecture

The optical path design of the paraxonic architecture is relatively simple, and there is no need for a large number of modifications to the existing AM equipment, but as it adds subsequent image preprocessing and algorithmic research, it has received significant research attention. For example, Kleszczynski et al.⁸³ used a high-resolution monochromatic CCD camera in a laser melting system for image monitoring and defect analysis, which were performed through a paraxial observation window to obtain high-resolution images. Consequently, they proposed a series of measures to reduce process failures. Separately, Grasso et al.⁸⁴ reported that, in the SLM process, the accumulation of defects in the previous layer leads to the failure of the next layer and that traditional micro-CT and ultrasonic NDT can only be processed offline, meaning that they cannot be corrected in real time. In response to this problem, Grasso adopted a paraxial structure to improve the defect identification speed. A statistical descriptor of the image data based on principal component analysis (PCA) was proposed, which was used to identify the defect area in the image layer and perform more rigorous analysis.

In AM, the quality of the produced parts relates closely to the temperature distribution of the molten pool. In recent years, researchers have usually monitored print quality online by measuring the temperature field of the molten pool. However, the small size, ultrahigh temperature, rapid formation, and solidification of the molten pool make measuring the temperature of the molten pool difficult. To address this issue, Liu and colleagues⁸⁵ designed a single-camera, temperature-field high-speed measurement optical path based on the dual-wavelength temperature measurement principle. They proposed a dual-band image matching method with subpixel accuracy and a multiparameter collaborative optimization calibration method for the scale coefficient K and wavelengths ${\lambda }_1$ and ${\lambda }_2$. The optical path for the measurements is shown in Figure 9.⁸⁵ In addition, an online measurement system for the temperature of the molten pool was developed, with the verification experiment showing that, above 600°C, the temperature measurement error of the system was <1%. The temperature distribution of the melting pool in the DED process was measured online. The measurement system is shown in Figure 10.⁸⁵ The method reduces the system development cost significantly and provides real-time monitoring of the temperature distribution of the melting pool during the DED process.

FIG. 9.

(a) Optical path for the dual-channel filter with a single camera. (b) Design principle for the optical path in (a). (Reprinted with permission. Ref.⁸⁵ Copyright 2020, Elsevier Ltd.)

FIG. 10.

Melting pool temperature measurement system. (Reprinted with permission. Ref.⁸⁵ Copyright 2020, Elsevier Ltd.)

Moreover, in response to quality and stability issues, such as splash, Liu et al.⁸⁶ conducted cutting-edge research on splash detection in the SLM process. Furthermore, Repossini et al.⁸⁷ used a high-speed camera with a paraxial structure to collect images and extracted statistical descriptors related to SLM splash behavior using image segmentation and feature extraction methods. These descriptors were used to examine the correspondence between splash descriptors and SLM processes with different laser energy densities via a logical regression model.

Online monitoring of visible light and ultraviolet signal

In the AM process, splashes and plasma plumes contain a large amount of ultraviolet and visible light, while the recoil pressure generated by the metal vapor above the melting pool contributes to the molten material (droplet splash) and powder splash near the melting pool.⁸⁸ Based on the important role of recoil pressure in the melting process, Andani et al.⁸⁹ used a high-speed camera to observe the formation mechanism and dynamic behavior of splash particles in the SLM process and proposed a computational image analysis framework to evaluate the size and quantity of splashes and determine the influence of splash morphology and composition on component surfaces. The results showed that the nonmelted areas in the powder deposition or solidified layer are a harmful factor that affects the mechanical properties. In addition, compared with the energy input, changing the laser scanning speed has a greater impact on the formation of spatter, providing a new monitoring method for optimizing process parameters.

Liu et al.⁸⁶ used a high-speed camera to observe the splash dynamics of 316L stainless steel powder with different energy inputs. The results showed that the magnitude of the energy input affects the splash size, scattering state, and spray height. The splash particles are revealed to be predominantly spherical, and the higher the energy input, the more intense the splash behavior. Figure 11⁸⁶ shows the splash behavior corresponding to different energy inputs.

FIG. 11.

Splash behavior under different energy inputs (fixed speed of 50 mm/s, layer thickness of 0.04 mm): (a) ψ = 0.26 × 10⁶ W/cm³, (b) ψ = 0.52 × 10⁶ W/cm³, (c) ψ = 0.78 × 10⁶ W/cm³, and (d) ψ = 1.04 × 10⁶ W/cm³. (Reprinted with permission. Ref.⁸⁶ Copyright 2015, Elsevier Ltd.)

Acoustic signal online monitoring

Using acoustic signals for online monitoring offers several advantages, including simplicity, low cost, a small amount of data collection, and depth prediction, for the manufactured parts.⁹⁰ It is an effective means for real-time online monitoring during AM.⁹¹ Ye et al.^92–94 have studied the application of acoustic signal monitoring in the SLM process and analyzed acoustic signal formation mechanisms, thus elucidating the mapping relationships between acoustic signals, laser power, and laser scanning speed. Based on the feature extraction of acoustic signals and the use of machine learning (ML) models, the classification and prediction of five typical melt channel states in the SLM process were realized. Among them, state A is undermelting, B is slightly undermelted, C is a normal state, D is slightly overmelted, and E is overmelted.

Ultrasonic signal online monitoring

The transmission of the ultrasonic signal is generally completed by the ultrasonic transducer, with only part of the scattered wave generated at the defect reaching the transducer, which is then converted into an electric pulse.⁹⁵ The ultrasonic transducer should be mounted directly on either the metal plate or on the lens of the laser guide path.⁹⁶ Rieder et al. applied ultrasonic signal monitoring to the dynamics of the SLM process and developed a set of ultrasonic online monitoring systems. By evaluating the echo signal, the residual stress can be evaluated qualitatively,⁹⁷ while estimating the ultrasonic velocity can be used to predict the porosity, although for parts with simple geometrical structures only.⁵⁴ The detection of various defects depends on their position within the manufactured component, and the porosity-based evaluation of ultrasonic signals requires further study.

Importantly, ultrasonic transducers can detect defects quickly and are applicable at high temperatures. However, although ultrasonic monitoring can evaluate defects such as residual stress, porosity, and cracks, its effectiveness is affected by the surface roughness of the component, with reduced sensitivity when detecting microstructures.

Online monitoring of electronic and other signals

Online monitoring can be performed using electronic signals, eddy currents,⁹⁸ and magnetic fields,⁹¹ all of which are highly sensitive but are yet to be widely adopted by AM systems. For example, Smith et al.⁹⁹ used the acoustic wave generated by absorbed laser heat to image the pulsed laser on the surface of the structure, with the properties and defects inside and outside the forming layer detected via spatially resolved acoustic spectroscopy. By contrast, Van Belle et al.¹⁰⁰ installed thermocouples and strain gauges at the bottom of the melting platform to measure the thermal stress generated during the melting process and used a detection system to evaluate the residual stress generated during the melting process. Kleszczynski et al.¹⁰¹ used an acceleration sensor to monitor the vibration signal during the SLM process. The system used a piezoelectric accelerometer integrated on the powder spreading mechanism to collect the vibration signal between the powder spreading mechanism and the manufacturing platform and determined the rising critical value that maintains the stability of the powder spreading device during the manufacturing process. In addition, certain signal monitoring instruments, such as spectrometers, can provide details of the properties and composition of the plasma plume above the molten pool. Although spectrometers have been used in laser processing technology, their use is not widespread in 3D printing.¹⁰²

Each online detection method offers advantages and disadvantages. Table 3 lists the principles, advantages, and disadvantages of infrared radiation, ultraviolet radiation, ultrasonic signals, acoustic signals, electrical signals, heat signals, and vibration signals for monitoring 3D-printed molded parts.

Table 3.

Principles, Advantages, and Disadvantages of Various Signal Monitoring Defects

Signal	Defect types	Theory	Advantages and disadvantages
Infrared radiation	Thermal stress, defect morphology, and location	The temperature field distribution of the target is detected.	Direct observation. Low accuracy and sensitivity of the diode signal; large amount of camera data processing.103
Ultraviolet radiation	Holes	The plasma plume distribution and its absorption of laser energy are detected.
Ultrasonic signal	Depth and location of the defect	Defects reflect the ultrasonic signals.	Restricted by part geometry and high signal-to-noise.104
Acoustic signal	Holes or internal cracks	Holes and internal cracks emit acoustic waves of specific frequency.	Easy signal detection, but high signal-to-noise.105
Electric signal	Holes	Measures the potential difference and electron distribution between the workpiece and the laser scanner.	Devices are unsuitable for installation.106
Heat signal	Microstructure, residual stress, deformation	The temperature history of the target is monitored by thermograph or thermocouple.	Observe the phenomenon directly, but the thermocouple can only measure the temperature of the fixed point, and needs to touch the measured object.
Vibration signal	Cracks, holes	The vibration signal is monitored by the acceleration sensor to determine the formation of small holes and the fluctuation of the molten pool.	There is little research at present, and further research is needed.

High-speed synchrotron X-ray technology

Unlike other monitoring methods, X-rays can directly reflect the 3D morphology and location of internal defects, such as the size and location of pores and cracks. In particular, high-speed synchrotron X-ray technology has played a vital role in the study of real-time in situ observation and analysis of defect formation mechanisms in the AM process.^13–19

For example, Chu et al.¹³ used in situ and operando high-speed synchrotron X-rays to reveal the underlying physical phenomena during the deposition process of the first and second layer melt tracks of the laser AM. The study found that the laser-induced gas/vapor jet through spattering (at a velocity of 1 ms⁻¹) promoted the formation of melt tracks and denuded zones. They also uncovered mechanisms of pore migration (recirculating at a velocity 0.4 ms⁻¹) driven by Marangoni fluid, the dissolution, and dispersion of pores during laser remelting.

Martin et al.¹⁴ conducted research on the formation of pores in laser powder bed melting. Using in situ X-ray imaging and multiphysical simulation methods, the formation mechanism of pores during the melting of the laser powder bed at the laser turning point was clarified. During the turning process, the deceleration and acceleration of the scanning mirror based on the galvanometer lead to the appearance of deep keyhole depression and subsequent collapse, and the localized normalized enthalpy of the material surface changes drastically, forming a small hole at the laser turning point. When the laser accelerates away from the corner, the keyhole depression collapses, the molten metal fills the gap, traps the argon gas, and finally forms a small hole when the material solidifies. To eliminate this kind of pores, a general power mitigation strategy has been developed to eliminate the formation process of this pore and improve the geometric quality of the melt trajectory.

Ren and Mazumder¹⁵ used a combination of artificial intelligence and plasma emission spectroscopy to identify the porosity of additive manufactured parts. During the DED manufacturing process of 7075 aluminum alloy specimens, time and position synchronization spectra were collected, and 18 features were extracted from the spectra to couple with deposition quality. Three-dimensional X-ray tomography (CT) is used to characterize the deposition quality and used to train a random forest classifier, which has an accuracy of 83% in identifying deposition porosity. Zhao et al.¹⁶ used high-speed X-ray imaging technology to observe the formation of pores in Ti-6Al-4V caused by the critical instability of the keyhole tip. In the power-velocity space, the pore porosity boundary is sharp and smooth, with little change between the bare board and the powder layer. Figure 12¹⁶ shows the keyhole porosity boundary and role of powder in laser melting. Figure 13¹⁶ shows the megahertz X-ray images of a keyhole pore-formation process.

FIG. 12.

Keyhole porosity boundary and role of powder in laser melting. (Reprinted with permission. Ref.¹⁶ Copyright 2020, American Association for the Advancement of Science.)

FIG. 13.

Megahertz X-ray images of a keyhole pore-formation process. (Reprinted with permission. Ref.¹⁶ Copyright 2020, American Association for the Advancement of Science.)

Multisensor and multisignal fusion online monitoring

At present, the single-signal monitoring system cannot evaluate the product quality with sufficient accuracy. Therefore, to meet different requirements and measurement parameters, precision instrument detection systems often require multiple sensors, each designed to measure a specific parameter; then, these independent measurements are combined using an algorithm to provide the final measurement result. Multisensor technology can be used to monitor the surface quality, monitor process parameters effectively, compensate for processing errors, and improve the processing quality. Therefore, the use of multisensor systems is gaining popularity. Coordinating the signals from multiple sensors is key to successfully implementing multisensor systems.^107–110 The measurement data recorded by the individual sensors are integrated, and a special algorithm is used to describe the integrity and consistency of the measurements—this process is called multisensor data fusion.^111–117 This technology can improve detectability and reliability, expand the scope of spatiotemporal perception, reduce inference ambiguity, improve detection accuracy, increase target feature dimensions, increase the spatial resolution, and enhance system fault tolerance. Figure 14¹¹⁸ shows the intelligent multisensor system framework.

FIG. 14.

Intelligent multisensor system framework.

For AM monitoring systems, mainstream sensors and field data measurement equipment can be divided into the following categories: noncontact temperature measurement, visible imaging, and low-coherence interference imaging. Figure 15⁷⁶ illustrates an example of a multisensor monitoring system used in an AM process. The optical monitoring device comprises NIR CMOS cameras and photodiodes. Sensors 1 and 2 are sensitive to wavelengths from 400 to 1000 nm and measure the molten pool radiation. An NIR CMOS camera measures the shape and temperature distribution of the molten pool during the manufacturing process.

FIG. 15.

Schematic diagram of a multisensor monitoring system for AM.

The principal use for DED is to manufacture large-scale functional metal parts. Because the DED process runs in an open-loop environment, problems involving production loss and insufficient reproducibility arise. Ongoing research projects are seeking to develop single closed-loop control systems that can control material delivery, laser energy, and heat distribution. Typically, the monitoring system can control one criterion only, although a new multisensor method capable of monitoring at least two continuous processes has been proposed.¹¹⁹ The focus of this research was to develop both a thermal closed-loop and a geometric control system. Infrared cameras were installed in wire arc AM and laser metal deposition units to provide global thermal imaging during the manufacturing process. An infrared camera was installed far away from the melting area to enable a sufficiently large imaging field.

Recent years have seen the research interest in molded parts produced via AM intensify. To date, such research has demonstrated that the surface roughness, density, defects, and residual stress of manufactured parts are related to changes in the acoustic signal, infrared light, molten pool shape, and temperature gradient in the AM process. Most studies have shown a direct relationship between process parameters and product quality, but it can be seen from Figure 16 that changing a process parameter may affect multiple process features linked to the part quality. With respect to process parameters, process features, and product quality, multiple inputs lead to multiple outputs. Therefore, if the quality of the manufacturing process can be monitored more accurately and in real time, then analyzed and fed back to the process control algorithm enabling the control strategy to be adjusted, the product quality of the AM process can be improved.

FIG. 16.

The relationship among process parameters, process characteristics, and product quality.

Process Monitoring and Feedback Control Algorithms

AM technology has great potential in the fields of aerospace, biomedicine, and machinery manufacturing. However, it has many influencing factors and a high degree of coupling, which presents challenges when trying to improve the quality stability of molded parts. Therefore, current research has focused on establishing an online monitoring system that can control and adjust printing process parameters through real-time feedback. Therefore, process monitoring and feedback control algorithms are the key to realizing online monitoring and feedback control in AM.

In the online detection process of AM, the original signal is first obtained through the detection framework. After feature analysis and extraction, the original signal is used as the input signal for algorithm recognition. Then, the original data are trained, and finally compared and evaluated by the database. Figure 17 shows the specific flow of feedback control in the online monitoring system. Therefore, when the actual processing signal is used as the input, defect information is acquired through comparative analysis with the database formed by the algorithm framework, and timely feedback can be provided. Research on such monitoring and feedback control algorithms generally divides the algorithm model into two parts: analysis using the statistical level as the entry point, and analysis using the ML level. The following classification and analysis can be applied to summarize the similarities and differences in practical applications.

FIG. 17.

The specific data flow of the feedback control algorithm.

Application of statistical models and algorithms

Statistical process control (SPC) is a traditional quality control method. The evaluation and monitoring of various stages of the processing process ensure that the process is maintained at a controllable and stable level and that the product meets quality requirements. As a basic analysis tool for monitoring the AM process, SPC can be used to provide regression analysis between characteristic values, correlation analysis, histogram analysis, and descriptive statistical analysis. In addition, the SPC chart is also of great significance for the stability and quality control of the AM process.

Owing to the large amount of data generated during the AM process, such as infrared image data of the molten pool temperature field, acoustic signal data, and ultrasonic signal data, the relevant statistical algorithms can be used as a reference in the process monitoring system. The analysis, extraction, and data integration provided by algorithms facilitate improvements to the reliability and stability of monitoring.^{86,93,94,120–127} To begin, these algorithms collect and input a large amount of raw data to build a monitoring system database. Simultaneously, the original data are characterized according to different performance indicators, enabling the process monitoring system to index, match, and compare them with the database according to the characteristics of the original data in real time. Consequently, the processing state can be assessed in real time. Among them, the statistical process model and algorithm analysis are shown in Table 4.

Table 4.

Statistical Model and Algorithm Analysis

Statistical process control model
Eigenvalue analysis tool	Temperature field analysis data of molten pool	Analysis goal
Regression analysis Related analysis Scatter diagram analysis Histogram analysis Descriptive statistics Subanalysis	Infrared image data Near-infrared image data Acoustic signal data Ultrasound signal data Electromagnetic signal data Photodiode data	Molten pool stability120
		Pros and cons of fusion quality121

Repossini et al.⁸⁷ adopted image acquisition, segmentation, and feature extraction to use SPC methods to estimate the different feature descriptors of the splashing behavior on the laser scanning path and established a logistic regression model. Thus, the ability of spatter correlation descriptors to classify different energy density conditions corresponding to different mass states was determined. The results showed that using spatter as the driving factor of process characteristics can significantly improve detection ability under penetration and excessive melting conditions. It also showed that the use of splash features for analysis and modeling could effectively improve the stability of the AM process. By detecting and extracting features of the melting pool, spatter, and flying feather in the SLM process, Grasso and Colosimo¹²⁸ proposed an algorithm for monitoring and evaluating the SLM process that used relevant methods of SPC in combination with machine-learning-related models (support vector machine [SVM]) and verified the effectiveness and reliability of the algorithm in practical applications. Specifically, they extracted plume features from infrared video imaging of the high-speed molten pool, and combined them with the scale coefficient control chart to construct an online monitoring system for SLM with zinc powder, as well as an evaluation system to assess the processing quality.^128,129

Application of ML models and algorithms

ML is a manifestation of artificial intelligence. It studies the selection of appropriate algorithms from data, automatically summarizes logic or rules, and makes predictions based on this inductive model and new data. Therefore, data, algorithms, and models are the three core elements of ML. The general algorithm framework is shown in Figure 18. Applying the ML model to the AM process can perform performance prediction, parameter optimization, defect identification, classification, regression, and prediction on the characteristics of multiple information sources and judge the stability of the process. Following the development of artificial intelligence, deep learning algorithms, a subset of ML algorithms, have gradually emerged in AM research, promising broad application prospects. To date, several research teams have applied ML to process monitoring systems for AM.^130–134 Table 5 summarizes the AM process monitoring based on ML, which will be introduced separately.

FIG. 18.

Machine learning algorithm framework.

Table 5.

Summary of Additive Manufacturing Process Detection Based on Machine Learning

Machine learning methods	Data	Goals	References
Principal component analysis +K-means clustering	Optical image	Automatic detection of delamination defects, geometric structure error	84
Self-organizing map clustering algorithm	Similarity of heat distribution in molten pool	Microstructure abnormalities and pore defects	135
Deep belief network	Acoustic signal	Defect detection	93
Improved deep belief network	Plume and splash characteristics	Molten state, Part quality	130
Multilayer classifier for process monitoring	Molten pool parameters	Print quality	138
Convolutional neural network	Morphological characteristics of molten pool	Porosity and pore instability	131
Convolutional neural network	SLM raw video images	Defect abnormal	132
Convolutional neural network	SLM raw video images	Print quality	133
Support vector machines	Molten pool temperature data	Defect abnormal	140

SLM, selective laser melting.

In one such example, Grasso et al.⁸⁴ proposed a layered process defect detection and spatial recognition method based on a machine vision system and visual range and constructed a feature descriptor based on PCA, which was suitable for identifying image data. Then, the authors applied the image K-means clustering method to detect defects automatically and discussed examples of simple and complex geometric structures in SLM to verify the performance of the method. Alternatively, Khanzadeh¹³⁵ proposed an online monitoring method for porosity prediction based on the similarity of the thermal distribution of the molten pool, which was aimed at alleviating the problems caused by microstructure anomalies or pores in DED. In addition, a new data processing program was developed to convert the poorly structured molten pool image flow into a continuous heat distribution model with the same function. After the molten pool images were defined in the same domain, the self-organizing map (SOM) clustering algorithm was used to group the molten pool heat distribution based on similarities, thus identifying similar and different molten pools. In this approach, the difference between two clusters depends on the correlation between them; if a particular cluster has low correlation with other clusters, or only contains a small amount of the molten pool, it is considered an anomaly, with this anomaly predicted as the porosity of the corresponding position. Through experimental verification, when selecting a suitable SOM model, the method based on the temperature distribution of the molten pool can predict the porosity distribution with an accuracy rate of 96%.

Considering the SLM process, Ye et al.^{93,94,130,136,137} collected the acoustic signals of the overlapping states of different melting channels and near-infrared images of the molten pool moving at high speed and used both sound and image processing algorithms to extract multiple sets of effective features of each typical mode (e.g., sound signal characteristics, the state of the high-speed moving molten pool, splash, flying feathers, and other quantitative characteristics) to form a corresponding training set. A multilayer perceptron, SVM, deep belief nets, and other ML models were used to train the effective online identification of the weld state, splash, and weld pool during the weld formation process. The results showed that the fusion of acoustic signals with near-infrared information, which provides complementary advantages, enhances the accuracy of melting process monitoring.

In 2018, Amini and Chang¹³⁸ proposed a new multilayer classifier for process monitoring framework and conducted a framework test based on simulated data to achieve layer-by-layer supervision, a rapid feedback status signal, and early warning. Elsewhere, Scime and Beuth¹³¹ used the collaborative algorithm framework of computer vision technology and the unsupervised ML model to perform feature extraction learning and training on the image data collected by high-speed cameras. This was applied to assess the quality of the laser powder layer melting zone. Moreover, Scime and Beuth¹³² also investigated the use of deep learning algorithms, such as the multiscale convolutional neural network (MS-CNN), to realize online real-time monitoring of the characteristics of the molten powder pool during SLM. Their study demonstrated the practicality and superiority of this method. Similarly, Yuan et al.¹³³ also used an image processing algorithm based on a CNN to build the core framework of an online monitoring system for SLM. Using a vast quantity of data (nearly a 1000 SLM original video images), they constructed the original signal input of the algorithm database and used it as the basis for algorithm learning and training. Finally, it was applied to the actual processes, yielding high reliability.

In addition, Mukherjee and Debroy¹³⁹ proposed a challenging new concept: online monitoring and process control for SLM based on digital twins. The authors reported that the combination of a large amount of monitoring data used in the forming process, simulation methods, and ML algorithms is expected to produce a digital twin model for the AM process. This digital twin model can reduce defects and shorten the time between design and production, thereby reducing production costs.

The digital twin method¹⁴⁰ (gray box) combines the advantages of a physical model (white box) and a data-driven model (black box). Physical models typically comprise deterministic equations that encapsulate changes in physical properties. For example, a second-order partial differential heat equation is used to capture the temporal and spatial distribution of temperature in the 3D printing process. The physical model is essential to understanding the process mechanism, but it is not suitable for capturing randomness, and the computational cost increases with the appearance of multiscale phenomena.^141–143 In contrast, data-driven models are constructed using only empirical data and link the process features extracted from the sensors with the observation results. Data-driven models can adapt to the uncertainty in the process but, compared with physical models, are less capable of explaining the root causes of defects.^144,145 The digital twin method provides the following advantages.^139,146,147

• (1)Optimized laser power, scanning speed, and other process parameters to provide guidance for part design.
• (2)Combination of theoretical predictions and real-time sensor data as the basis for early process fault detection and monitoring, which guides model-based feedforward control in AM.
• (3)Multiscale modeling with reduced computational burden.

In 2020, based on previous work, Gaikwad¹⁴⁰ combined the predictions from physical models (simulations) and field sensor data to form an ML framework to detect defects and abnormalities generated in the AM process. Specifically, a computational heat transfer model based on graph theory was used to predict the instantaneous temporal and spatial distribution of local temperature in laser powder bed fusion and DED. This method is more efficient than the finite element method, while maintaining similar prediction accuracy. Then, these predictions were combined with the molten pool temperature data obtained from the two-color coaxial in situ high-temperature detection system in an easy-to-implement supervised ML framework (i.e., an SVM) to detect potential anomalies in the AM process. The effectiveness of this digital twin method was verified using experimental data in the LPBF and DED processes, with the accuracy of the defect predictions during the inspection process exceeding 80% in some cases.

Through the summary, induction, and analysis of the application of statistical models and ML models in AM process monitoring, we understand that based on the collected signals, extract features and use relevant model algorithms to establish the relationship between signals and defects or processes, which can classify or predict defects or processing status. However, the laser scanning speed in the AM process is fast. The data processing time is a huge challenge to realize real-time monitoring and control.

Conclusions

Outlook

To meet the requirements for online detection and process monitoring of AM technology, studying the formation mechanism of various defects in AM technology, realizing the scientific definition of various defect types, and developing scientific and effective online detection methods are highly important for improving the processing level of AM technology and ensuring the quality of product formation. At present, developments in online detection technology for AM are focused on the following aspects.

• (1)Fusion of multiple detection methods, such as multisignal fusion monitoring technology and special detection technology. Multisensor technology and information fusion methods are important research trends for further developments and are important requirements for improving the detection accuracy of AM technology. Therefore, on the basis of offline detection, the process monitoring of multiple signal sources is integrated to realize closed-loop control of the entire AM process. Also, the defect formation process can be observed and analyzed in situ in real time through high-speed synchrotron radiation X-ray technology.
• (2)Application of deep learning algorithms and artificial intelligence. Algorithms are essential to online monitoring and closed-loop control. The application of ML algorithms has provided considerable advantages in the evaluation of AM product quality. With the rapid development of deep learning and artificial intelligence algorithms, algorithms based on image features, pattern recognition, and sound features are highly adaptable for the formulation of control strategies, and represent the future of NDT.
• (3)Integration of detection methods and simulation analysis. The latter is indispensable for understanding the formation mechanisms involved in AM. Many current simulation methods, such as finite element analysis and multiphysics coupling, can provide scientific explanations that cannot be achieved by detection methods. Therefore, the in-depth integration of simulation methods and detection methods is an important goal for future studies, and there has already been notable progress.
• (4)Hardware optimization. Real-time performance is an important requirement for an online inspection system. Therefore, when building an online inspection system, the camera module, laser module, powder feeding platform, and signal processing control system must be optimized. Furthermore, the rapid transmission of large amounts of data acquired during the inspection process has increased the demands on the communication capacity of AM systems.

Summary

Based on a literature review, this study's scope is to identify the measurement science's needs for AM. This article reviews state-of-the-art detection technologies in the field of AM technology.

First, we introduced the types and causes of defects involved in AM, including summarizing the relationship between defects and process parameters and the necessity of NDT methods to identify defects and residual stress. Important parameters such as defect spatial resolution, detection speed, and scope of application of common NDT methods are discussed. Second, from the perspectives of online detection and process monitoring, the existing detection technology classifications, detection signal sources, and detection architectures are introduced comprehensively. In particular, the research progress of real-time in situ observation and analysis of defect formation process with high-speed synchronous X-ray technology in the past 5 years is introduced. The monitoring signal types and sensing technology in the AM process, and the defect types, advantages, and disadvantages of current online monitoring methods for 3D-printed parts are summarized. Finally, for process monitoring, the importance of process control and feedback algorithms was emphasized, and recent research results of process control algorithms were analyzed and summarized from the perspectives of statistics and ML, providing an insight into the current trends in AM research.

The review is intended to provide a background reference and guidance for the research and application of online detection technology related to AM. An effort has been made here to include all the important contributions in the current area of interest.

Author Disclosure Statement

No competing financial interests exist.

Funding Information

This work was financially supported by the National Key Research and Development Program of China (Grant No. 2017YFB1103900), the National Natural Science Foundation of China (11972084), the National Science and Technology Major Project (2017-VI-0003-0073), the Beijing Institute of Spacecraft Environment Engineering (CAST-BISEE2019-010), and the Beijing Natural Science Foundation (1192014).