Process monitoring for material extrusion additive manufacturing: a state-of-the-art review

Qualitative uncertainties are a key challenge for the further industrialization of additive manufacturing. To solve this challenge, methods for measuring the process states and properties of parts during additive manufacturing are essential. The subject of this review is in-situ process monitoring for material extrusion additive manufacturing. The objectives are, first, to quantify the research activity on this topic, second, to analyze the utilized technologies, and finally, to identify research gaps. Various databases were systematically searched for relevant publications and a total of 221 publications were analyzed in detail. The study demonstrated that the research activity in this field has been gaining importance. Numerous sensor technologies and analysis algorithms have been identified. Nonetheless, research gaps exist in topics such as optimized monitoring systems for industrial material extrusion facilities, inspection capabilities for additional quality characteristics, and standardization aspects. This literature review is the first to address process monitoring for material extrusion using a systematic and comprehensive approach.


Introduction
Additive manufacturing is already an accepted technology for special applications and prototype production. However, it has considerable potential for further expansion in the future [1]. Examples of future applications are small-batch productions in the automotive [2] and aerospace [3] sectors as well as the production of customized medical devices [4]. Additive manufacturing can further be used in the jewelry [5] and construction industries [6]. Niche applications include mouthpieces for musical instruments [7] or textiles for clothing [8].
Solving the challenge of qualitative uncertainties in terms of materials, processes, and products, as well as process knowledge deficits, is vital to further incorporate additive manufacturing in the industry [9,10]. Therefore, providing tools for comprehensive quality management is essential [11,12]. Means of measuring process states and part properties during additive manufacturing are particularly relevant to achieving this aim [9,[13][14][15].
Process monitoring enables the assessment of whether a product satisfies certain requirements. In-situ inspection techniques fundamentally increase customer confidence in a product and reduce costs due to rejection, because process anomalies are detected immediately after they occur. Furthermore, information from process monitoring is the basis for implementing a closed-loop quality control [16]. A significant challenge for testing technologies in the field of additive manufacturing is the complex geometries of parts that contain infill structures and process-specific defects [17,18]. This review aims to identify and analyze the existing literature on in-situ process monitoring for material extrusion (MEX), as it is one of the most widely used additive process categories [1,19].
Former reviews, specifically on the additive manufacturing of metal parts, have already been published [17,20,21]. Their focus lies on monitoring techniques for powder bed [22] fusion and directed energy deposition [16,[23][24][25][26][27]. The results of these studies are not directly transferable to MEX because additive process categories are significantly different due to dissimilar processing principles being applied [9]. However, a number of reviews which comprise a wider range of additive process categories have been published: Vora and Sanyal [28] investigated the usability of different conventional inspection techniques for process monitoring in additive manufacturing. Their focus was the analysis of general functional principles. Process monitoring in MEX was merely minimally addressed. Charalampous et al. [29] discussed the research on sensor-based quality monitoring before, during, and after the additive manufacturing process. They presented nine different projects on MEX in-situ process monitoring. Controlling the additive processes using sensor technologies was the focus of a study [30] that listed commercially available solutions in addition to research work. It included eleven references regarding MEX. Lu and Wong [14] presented fundamental challenges and developed principles for monitoring with thermography, and acoustic emissions. However, MEX was only considered to a very limited extent. A review on ultrasonic testing by Honarvar and Varvani-Farahani [31] discussed two MEX projects. Furthermore, applications of machine learning have already been discussed in various publications [32][33][34]. One of their topics was process monitoring, but the presentation of MEX projects was marginal.
In summary, the studies on hand provide only a rather limited insight into the subject matter of MEX in-situ process monitoring. A comprehensive and systematic analysis of the state of knowledge has yet to be conducted. Therefore, the aim of this study is to compile and structure the current state of research using an approach that is as objective and comprehensive as possible. The following three central questions will be answered: • How much activity is involved in the field of process monitoring? • What methods and technologies are used for the process monitoring of which quality characteristics? • What are the research gaps?
After an overview of the fundamentals of MEX in Sect. 2, the methodology for the literature search and analysis is introduced in Sect. 3. Subsequently, in Sects. 4, 5, and 6 the results are presented and discussed, structured according to the abovementioned questions. Finally, Sect. 7 summarizes the main conclusions of the study.

Material extrusion
In MEX, a feedstock is extruded and deposited in beads by the relative movement between a nozzle and a substrate. During extrusion, the material is in a semi-solid state and solidifies when it reaches its final position and shape [19,35]. Various sub-categories are grouped under the MEX process category. They differ in the type of extruder (plunger, gear, or screw), form of feedstock (filaments, rods, or pellets) [36], and kinematic design (Cartesian, polar, delta, or robot arm) [37].
The advantages of MEX are the simplicity of the process, relatively low costs [9] and a large variety of feedstock materials [38]. In addition to standard plastics, fiber-reinforced polymers can also be processed [39]. Furthermore, it is possible to produce parts from concrete [6], metals, ceramics, and multiple materials [36]. Because of the high material deposition rates that can be achieved [40], special MEX systems can be used for large-format additive manufacturing (build volumes of over 1 m 3 ) [41]. MEX can compete with conventional manufacturing processes in terms of cost per unit for small and medium batch sizes [42]. An example of an application in this batch size range is polymer components for the aircraft industry [43].
Numerous influencing variables (e.g., process parameters and material properties) affect the mechanical and geometric properties as well as the surface characteristics of the parts produced by MEX [39,44,45]. Depending on the application, the requirements for the parts differ. Therefore, only certain quality characteristics related to the respective requirements are the target of process monitoring. Examples of quality characteristics are the geometric dimensions and density of parts [46]. Owing to the complex interactions among different influencing variables, various process faults that can negatively affect the quality of parts may occur. A selection of typical part defects is listed in Table 1.

Materials and methods
This study can be considered as a state-of-the-art review based on the classification of different review types by Grant and Booth [56]. The focus is on the presentation of the current status as well as the identification of research gaps. During the literature search step, as many thematically congruent publications as possible are identified using a systematic and reproducible search methodology. There is no evaluation and selection of publications based on the relevance of the study results and the quality of the study design. An aggregative approach is used to synthesize the identified sources by collecting and interpreting empirical data. In addition, a primary purpose is to provide an understanding of relevant research directions and topics [57].
The process of literature search shown in Fig. 1 included, as a first step, a literature search of nine different popular databases in February 2020. Each database was searched multiple times. The searches corresponded to the keyword ("fused deposition modeling" OR "fused deposition modelling" OR "fused filament fabrication" OR "material extrusion" OR "fused layer modeling" OR "filament freeform fabrication") AND ("process" OR "quality" OR "defect" OR "error" OR "fault" OR "condition") AND ("assurance" OR "control" OR "detection" OR "inspection" OR "measurement" OR "metrology" OR "monitoring" OR "sensor"). Single search operations contained only one term for naming the additive manufacturing process (first operand for the Boolean AND operators). Therefore, six individual searches were performed to query the keyword completely. In each database, the entire record was searched, but the number of exported hits was limited to 500 per single search operation. If the database supported a limitation of the search to titles, abstracts, and keywords of the publications, an additional search in these categories was performed without limitation on the number of exported hits.
After removing the duplicates with the aid of the literature management software Citavi (Swiss Academic Software GmbH), the dataset contained 9176 entries. To analyze relevant sources only, inclusion and exclusion criteria were

3
defined and applied to the dataset. The inclusion criteria were: • one of the sub-categories of MEX is treated; • central aim is in-situ process monitoring for quality assurance (assessing the status of 3D printer components or parts in production); • contribution is original research (peer-reviewed), dissertation or active patent.
The exclusion criteria were: • process monitoring is included but not for the purpose of quality assurance (e.g., sensor system to validate a simulation of the MEX process); • not in English or German; • older than 2013.
A total of 221 elements comprise the dataset for the review. The approach to analyze the identified publications, as well as the paper's corresponding sections, is presented in Fig. 2. 4 How much activity is involved in the field of process monitoring?
Some monitoring systems have been published several times and sometimes, several systems have been described in one publication. The grouping of sources according to project affiliation indicated that the dataset involved 145 different MEX monitoring systems. The criteria for grouping the sources according to project affiliation were research group membership and sensor technology.
For further characterization of the dataset, Fig. 4 illustrates which levels of functionalities of a process monitoring system have been addressed by the publications and in which development stage they are. The sensor system (F1) is a pure hardware setup. In the level that builds on it, data are processed and extracted (F2), e.g., for visualization. The third functionality level describes the automated data evaluation (F3) for the detection of anomalies. A closed-loop control (F4) represents the maximum possible functionality level of a monitoring system. Note that these categories progress in a typical order (F1→F2→F3→F4), where the latter categories necessitate accomplishment of the prior categories. Publications are placed in the highest category that their content represents. The stage of development is described with the following classifications: patent (P), preliminary studies (D1), and realized solution (D2). Figure 4 shows that the current focus of research is in F3 since the maximum number of D1 and D2 occurs on this level of functionality. However, the conspicuously high number of patents in F4 indicates that an economic benefit is seen particularly for this level of functionality. In the long term, therefore, further research activity can be expected in this area.

What methods and technologies are
used for the process monitoring of which quality characteristics?

Sensor technology groups and inspected elements
Various sensor technologies are used for process monitoring. Figure 5 displays the percentage shares of sensor technology groups in the total number of sensors used. The grouping is based on the measured physical quantities. The respective share of each sensor technology that is used simultaneously with another is represented by the "sensor fusion" section of the bar. Furthermore, all sensor technologies that have a share of less than 2% in the "one sensor technology" section and cannot be assigned to the other groups are collected under "other." Figure 6 depicts a statistical analysis of which elements of the additive manufacturing process are directly monitored by which sensor technology groups. On one hand, it is possible to monitor the components of the MEX machine that have an influence on the part quality. According to the main functional components of the MEX machine [19, 35, 45], the following are distinguished: • extrusion head (EH), including the extrusion nozzle and feedstock delivery mechanism; • feeding system (FS), for feedstock transport to the extrusion head; • build chamber (BC), including the housing and frame; • build platform (BP); and • axis system (AS), including the motors.
On the other hand, the part can be directly monitored. The following are distinguished depending on the area of monitoring: • entire part (P); • layers (L), equivalent to the build surfaces in the majority of cases; and • sidewalls of part (S). Figure 6 shows that the measurement of vibration, acoustic and electrical signals, as well as force and pressure, is primarily used to monitor the components of the MEX machine. The part is inspected primarily using vision technologies. The focus is on monitoring the extrusion head and individual layers.
The following subsections describe the identified publications sorted by sensor technology groups and project affiliations. The general functional principles are introduced, and selected monitoring systems are explained precisely. For detailed descriptions of the treated sensor types and their general advantages and disadvantages, the reader can refer to Vora and Sanyal [28].

2D vision
In Table 2, the projects identified within the field of 2D vision are listed, along with their associated references. The projects were sorted based on the following priority: (1) used sensors (column "Sensors"), (2) inspected elements (column "Ele"), (3) project level of functionality (column "Fun"), and (4) stage of development (column "Dev"). The column "Data handling" provides a brief description of the methods  used for sensor data processing. "Quality characteristics" are the features checked by the monitoring system. If the publications on a project do not contain certain information, this is indicated in the corresponding cell with the phrase "not applicable" ("n.a."). The generic term 2D vision is used in this paper to describe all sensor technologies that acquire two-dimensional images of an object in the visible wavelength range. Seven of the 23 patents identified in this work exclusively addressed 2D vision [111,117,[136][137][138]163]. Therefore, the potential of the sensor technology for MEX process monitoring is considered high by the industry.
The 2D vision technology is often used for the sequential inspection of layers. One technical variant includes mounting the sensor on the extrusion head [58, 65-67, 89, 104-106, 111, 112, 159-162]. For example, Liu et al. [160,161] investigated overfill and underfill defects using two digital microscopes, which were attached to the extrusion head to continuously analyze the layer surface in a small area next to the nozzle (Fig. 7). For the extraction of features, a texture analysis method in which the layer surface was described with a gray-level co-occurrence matrix was used. Subsequently, the layer surface was divided into five classes using the k-nearest neighbors algorithm. The material flow rate and speed of the cooling fan on the extrusion head were adjusted using a proportional-integral-derivative (PID) controller according to the classification to increase the layer quality.
In addition to projects that include mounting vision sensors on the extrusion head, another relevant approach is the stationary mounting of the camera with a view on the build platform. In this scenario, the entire layer is captured in one image acquisition [100,101,[104][105][106][107][108][109][110][129][130][131][132][133][134]. In one of the projects [131][132][133], statistical process control is used to evaluate the quality of the layer contours. Significant changes in the process caused by the exceedance of tolerance limits were displayed on quality control charts. In contrast, Delli et al. [134] compared images of a defect-free part with the actual manufactured part and used both a simple threshold method and a support vector machine to classify the part into one of two categories: good or bad. Aside from process monitoring of individual layers, 2D vision sensors may also be used for the exclusive inspection of the sidewalls of parts. In this technical variant, the camera axis is often perpendicular to the normal vector of the build platform. Baumann et al. [113] used this approach to detect deformations on printed objects, detachments from the build platform, and lack of material flow. Because the 3D printer is a desktop device with an open housing, the camera can be placed in front of the 3D printer to capture images of one side of the part.
The use of a camera to inspect sidewalls in large-format additive manufacturing was investigated by MacDonald et al. [59]. Fourier analysis was used to determine the variation in layer heights from the image data. Due to the large size of the beads, they can be easily distinguished from one another with an algorithm. Especially in large-format MEX with pellet feedstock, the extrusion process is highly sensitive to parameter variations. The authors demonstrated that the resulting slumping of beads or small irregularities protruding from the sidewalls could be detected with the monitoring system.
In a series of publications, Straub [119-126] presented a sensor system consisting of five cameras arranged around the build platform. For data acquisition, the printing process is stopped, and the build platform is moved to a predefined position. Besides the use of multiple cameras, mobile solutions to move the camera around the object to be printed have been proposed in further studies [88,90,117]; thus, the sidewalls of the part can be fully captured. Figure 8 shows this as an example with a camera attached to the extrusion head of a robot MEX system using a special mount.
In addition to the inspection of manufactured parts, some systems also use 2D vision to monitor the mechanical components of a 3D printer. Greeff et al. [157,158] utilized a digital microscope to inspect the filament delivery mechanism in an extrusion head. The speed and width of the filament were measured to calculate the volume flow. Moreover, the speed of the feeding gear was determined and compared with that of the filament to calculate slippage effects.

Temperature monitoring
Since materials are melted because of heat during MEX, the acquisition of temperature data is a practical method for evaluating the condition of the manufacturing process. Table 3 summarizes the corresponding publications. Temperature sensors for measuring and controlling the temperature of the build platform, extruder, and ambient air in the build chamber are conventionally installed in many MEX systems [178]. However, aside from sensors that are in contact with the measured surface, a large portion of the identified publications involve temperature determination via thermography. Thermography is an imaging technique used to display the surface temperature of objects. The intensity of the infrared radiation serves as a measure of the temperature.
Thermal cameras are often used to determine the temperature of the layers. Borish et al. [60] developed a method for calculating the average temperature of a layer in large-format MEX. They paused the printing process until the temperature decreases below a certain value. When this condition is attained, the next layer can be processed. The thermal camera is attached to a movable arm that is pneumatically driven. The study shows that temperature measurements are particularly relevant for large-format additive manufacturing since in rapid printing processes cooling times are sometimes insufficient and parts collapse under their own weight.
Monitoring the sidewall of a part with a thermal camera, Ferraris et al. [171] determined a correlation between the characteristic temperature curves and the size of the bonding surfaces between adjacent beads. Using a similar hardware setup, the tensile strength of samples was predicted in a work by Bartolai et al. [173,174].

Vibration monitoring
Vibration can be measured at many of the mechanical components of the 3D printer (Table 4). A key issue is the monitoring of extrusion head vibrations. Tlegenov et al. [181,182] attached an accelerometer to an extruder to determine the effective nozzle diameter, which was used as a measure for nozzle clogging conditions. They observed that the amplitude of the vibration increased nonlinearly with decreasing effective nozzle diameter. The results of an analytical model for the theoretical determination of the amplitude exhibited good agreement with those of the experiments using both Bowden and direct extruders. In another research work [185] sensors were attached to both the extrusion head and build platform. This enabled the detection of part deformations and defective extruder conditions. The detection of defects in mechanical components of the MEX machine was solely investigated by Yen and Chuang [87].

3D vision
The advantage of 3D vision compared to 2D vision is that height information can be captured. Table 5 indicates that nearly all of the publications address the monitoring of individual layers, in which comparison with different types of digital reference information was used for error detection. If structured light or stereoscopic imaging systems are used, the sensors are rigidly aligned to the build platform [75,76,100,[186][187][188][189][190][191]193]. Holzmond and Li [193] for example, used two five-megapixel cameras to create a stereoscopic imaging system. The viewing axes of the cameras were aligned perpendicular to the layers. To capture images of the layers, the extrusion head was moved out of the viewing axis by making it print a waste part parallel to the target part. After each layer, the extrusion head moved to the waste  part, creating a time window for image acquisition. A reference point cloud was generated from the G-code, which could be compared with the captured point cloud to detect defects. The approach was limited in that the system could only inspect materials with naturally textured surfaces. In contrast, laser triangulation sensors record single height profiles. Therefore, a relative movement between the inspection object and sensor should be attained to generate a 3D point cloud from a large number of height profiles. Hence, the laser triangulation system is attached to the extrusion head of the MEX machine and can be moved over the layer surface [62,64,72,73,[195][196][197].  [192] Camera, illumination L Comparison with reference, artificial intelligence control Various defects F4 P [193] 2 cameras, illumination L Comparison with G-code Geometric deviations, holes, blobs F3 D2 [75,76] 3 × 2 cameras, illumination S Comparison with STL file Geometric deviations F3 D2 [194] 3D camera P Comparison with reference Geometric deviations F4 P [195] Laser triangulation L Comparison with CAD model, measurement of defects Underfill, overfill F3 D2 [196] Laser triangulation L Visualizing sensor data Bead shape F4 D1 [62] Laser triangulation L Comparison with G-code Underfill, overfill F4 D2 [72,73] Laser triangulation L Comparison with nominal layer height, reslicing Layer height, bead width F4 D2 [64] Laser triangulation L Comparison with reference, generating modified path Spatial bead position F4 D2 [197] 2 laser triangulation L 2D comparison with G-code Geometric deviations, voids F3 D2 [198] n.a L Comparison with reference Geometric deviations F3 P

Acoustic emission monitoring
Acoustic emission monitoring can be used because various actuators and mechanical components of the 3D printer generate noise (Table 6). If anomalies occur, they will cause changes in the acoustic emissions. Many studies have used this sensor technology to monitor extrusion heads. For example, Wu et al. [203] attached an acoustic emission sensor to an extruder with vacuum grease. The mounting arrangement is depicted in Fig. 9. The state of the extruder was classified into the following using a hidden semi-Markov model: extruding without material, material loading/unloading, idle, and normal extruding. In validation experiments, a classification accuracy of more than 90% was achieved.
In another study [205,206], a sensor mounted on the build platform next to the part could detect detachment of the part from the build platform and deformations. The defective part came into contact with the nozzle, which resulted in altered acoustic emissions. Moreover, recording devices can be placed next to the 3D printer [207][208][209][210]. Using this setup, Chhetri et al. [207] reconstructed the geometry of layers based on the acoustic emissions of the axes and motors. By comparing the reconstructed geometry with the original G-code, they were able to identify cyberattacks. Evaluation experiments demonstrated that a modified geometry of a quadcopter baseplate was detectable. Table 7 lists all identified sources in the field of monitoring electrical quantities. The sensors used are often for monitoring motor currents. For example, the currents of the motors to push the filament through the extrusion head or to move the axes are measured. Nozzle blockages or incorrect axis movement cause changes in the motor current and can be evaluated. Kim et al. [211][212][213] observed that the motor current of an extruder is correlated with the level of extrusion pressure. The extrusion pressure depends on the size of the nozzle outlet and the distance between the nozzle and substrate. If the part is deformed and the distance to the nozzle outlet is reduced, or if a foreign object prevents the material from exiting, the pressure will increase and changes in the motor current will occur.

Force and pressure monitoring
Hitherto publications on force and pressure measurements focused on investigations of extrusion head elements (Table 8). Klar et al. [71] showed that the extrusion force in a piston-based extrusion device for processing ceramic, silicone, and acrylic pastes can be measured using a load cell. Force variations were directly related to the flow characteristics of the material. Other than the extrusion forces,  forces acting at the nozzle tip owing to the external effects of substrate defects can also be measured [219]. Furthermore, in the MEX of continuous fibers, fibers that are not fed at a sufficient rate by the delivery mechanism result in analyzable changes in forces. Exceedingly high forces, in turn, cause fiber pull-out and shearing [69].

Other sensor technologies
In this section, different sensor technologies with small numerical shares of publication in the literature are summarized (Table 9). In some publications, fiber Bragg grating sensors are presented as possible means of measuring strains. In such a system, the printing process is interrupted at a certain point and optical fibers are placed on the  (Fig. 10). If deformations of the part and consequently of the optical fiber occur, they can be detected and analyzed [221][222][223][224][225][226][227]. Since the placement of the optical fibers as well as the properties of the surrounding material have an impact on the accuracy of the measurements, Falcetelli et al. [246] discussed and investigated different fiber embedding strategies. Some research groups used ultrasonic sensors to analyze the part structures. Reflections of high-frequency pulses exerted onto the part were analyzed based on the duration until detection [229][230][231][232][233]. Another relevant approach is the use of encoders to determine the axis positions and to implement closed-loop control of the axis movement. This approach is considered state of the art within the NC machine industry [237]. It is also present in some MEX machines available for purchase [30].
The heterogeneity of monitoring systems prevented the further formation of clusters with similar functional principles. Therefore, the authors refer to individual publications for additional information.

Sensor fusion technologies
The fusion of data from multiple sensor technologies is a powerful method for monitoring a large number of features. Table 10 and Fig. 5 show that 2D vision and 3D vision are rarely used in combination with other sensor technologies. This is presumably due to the large information volume of the measurement data of the optical inspection systems. Additionally, optical measurement techniques are commonly used to inspect the quality characteristics of a part. In contrast, measurements that describe the condition of the 3D printer must be obtained via various routes to characterize the heterogeneous components of the machine.
An effective grouping of the identified monitoring systems is not possible. As an example, a monitoring system consisting of six thermocouples for temperature measurements at the extruder, at the build platform, and in ambient air is presented here. Furthermore, two sensors were used to measure the vibrations of the build platform and extrusion head. An infrared sensor measured the temperature of the build surface near the nozzle at the location at which the material was deposited. The authors explained that no additional benefit could be expected from using the thermocouples; therefore, only vibration and infrared sensors were used for process monitoring. The dimensional accuracy, surface roughness, and underfills could be determined [276][277][278]. The underfills were classified as "normal operation," "stringy extrusion," and "nozzle clogged." When producing a standard test artifact, the system achieved an accuracy of 97% for classifying into these three categories [276].
6 What are the research gaps?

Key topics for sensor technology and data processing
In a workshop of the National Institute of Standards and Technology, USA, the measurement science roadmap for polymer-based additive manufacturing was elaborated. Said roadmap specifies developments concerning measurement science required for the industrialization of additive manufacturing. For process monitoring, four prioritized roadmap topics (RT) were identified [13]: • RT1: new in-situ imaging modalities • RT2: real-time process measurement at required spatial and temporal resolution • RT3: in-situ control and model integration • RT4: big data analytics A comparison of RT1 with the identified literary sources shows that the current research activity likewise focuses on the development of imaging modalities. From an industrial perspective, approaches that address the inspection of layers are particularly promising. Here, a single sensor module can  be utilized to inspect both the outer walls and inner structures of parts. Geometries and surface characteristics can be effectively inspected using 2D vision and 3D vision. Optical temperature measurements can be used to verify the thermal material properties. In addition to imaging techniques, monitoring of extrusion head conditions should be prioritized in future research because it is a key element of MEX systems. Measurements of current, vibrations, and acoustic signals are advantageous because the sensors can be installed with minimal effort. In contrast, force and pressure measurements require modifying the mechanical extrusion head components. However, this enables precise determination of the polymer melt conditions. Regardless of the sensor technology, there is a fundamental necessity for research on integrating sensors into industrial MEX systems. New and improved sensor concepts that are designed for high ambient temperatures and large build volumes are required. Furthermore, efficient sensor modules, which can be realized in MEX machines despite restrictions due to moving machine parts and frame structures, must be developed.
The large number of patents on closed-loop control in Fig. 4 indicates that this topic is considered to be fundamentally important in the industry. High-performance measurement technology (RT2) is a prerequisite for these control loops (RT3). For the resolution of acquired data and speed of data processing, satisfactory results have already been achieved for some specific measurement tasks. This is demonstrated by the first controlled systems that adjust process parameters in sufficiently short periods and with adequate accuracy [63,160,180]. However, these systems require much improvement. For example, sensor technologies for detecting small voids or part contours in large-area, highresolution layer images at high speeds are not yet available. Furthermore, classifying monitoring systems use only a few classes; therefore, they have low resolutions. Moreover, the current closed-loop control is based on simple causal relationships. Mathematical models that describe complex relationships between several process parameters, control variables, and part properties have not yet been sufficiently researched.
Large and complex datasets generated by different sensor technologies and assignable to the field of big data analytics (RT4) were not used in the identified publications. Therefore, datasets with heterogeneous sensor data from several varying print jobs must be generated in the future to train robust inspection algorithms. The analysis of the literature has confirmed the significance of this subject by demonstrating that, owing to the complexity of the inspection task, only multi-sensor approaches enable comprehensive monitoring of the MEX process.

Rarely examined quality characteristics
Aside from the specific wear-prone components of the 3D printer, all properties of the parts are, in principle, relevant to MEX monitoring. The requirements for a part can be divided into mechanical and geometrical requirements, surface requirements, and requirements for feedstock materials [280].
The focus of the current research is on part geometries and surface properties in terms of overfill and underfill. However, measurements of surface roughness were addressed by only two research projects [267,[276][277][278]. The measurement of mechanical properties is another important aspect that was investigated by merely two works as well: Bartolai et al. [173,174] and Zhang et al. [271,272] addressed the prediction of tensile strengths. Means of inspecting material characteristics were not considered in any publication. The monitoring of these quality characteristics, which the current research only addresses to a limited extent, represents a gap for future research.

Variety and complexity of monitored parts
A challenge with MEX monitoring is the required flexibility [14]. Varying and often complex part geometries are manufactured in very small batches. Furthermore, many different materials can be processed. Therefore, the extent to which the flexibility of the MEX is reflected in the reviewed monitoring systems was investigated. The properties of parts manufactured in projects with the aim of process monitoring for quality assessment were analyzed considering the aspects listed below: • complexity of geometries (simple or complex), • number of different geometries, • materials used, and • number of different materials used.
The analysis showed that 19.3% of the projects contained an investigation of complex part geometries, 55.9% monitored simple geometries, and 24.8% did not specify the geometry. Simple geometries include, among others, cuboids, cylinders, or single material beads. In contrast, the complex geometries describe a prosthesis or valve housing, for example. For the number of different geometries per project, the authors observed that 47.6% of the projects investigated one geometry, 12.4% two geometries, and 7.6% three geometries. More than three geometries were analyzed in only 8.3% of the projects, while 24.1% did not specify the geometry.
40.7% of works did not specify the material. Polylactide (PLA) and acrylonitrile butadiene styrene (ABS) were used in 34.5% and 26.9% of the projects, respectively.
Composite materials were used in 6.2%, polycarbonate in 2.1%, and ceramic materials in 1.4% of the projects. Other materials had a proportion of < 1% each. In 74.4% of the projects that specify the material, only one type of material was investigated, while 19.8% of the projects used two, 4.7% three, and 1.2% four different materials. Projects that employed more than one material consistently produced different parts separately from just one material each. Only one publication [252] stated that the part was made from PLA and one additional support material.
The results show that projects with high complexity and variation in part geometries and materials are strongly underrepresented in the dataset. The analyzed monitoring systems tend to monitor manufacturing processes for simple geometries and small numbers of varying parts. Regarding the materials used, ABS and PLA dominate the research projects, the number of different materials per project is oftentimes low and multiple material parts are only considered to a minor extent. However, complex geometries and cost-intensive materials (e.g., metal-filled or fiber-reinforced plastics) are particularly suitable for process monitoring, because this is where the economic efficiency of the inspection system is most easily achieved. Therefore, there is considerable potential for further research regarding the monitoring of various complex parts.

Standardization
Owing to the novelty of the technology and the diversity of the topic, standardization in the field of additive manufacturing is still in its early stages. There are only a limited number of standards for the specification of part properties and non-destructive testing methods [14,28]. Analysis of the identified publications has also shown that no consistent definitions are used for quality characteristic names, feature specifications, and tolerance limits.
As a first step towards standardization, ISO/ASTM 52901 [281] basically describes how part characteristics, tolerances, and test methods are to be defined between the customer and the supplier. With regard to process monitoring, the decision of whether a process variation represents a defect or not is particularly crucial [282]. Future projects can use the draft standard ISO/ASTM DIS 52924 [46] to specify these tolerance limits, as the document defines the quality levels of MEX plastic parts in terms of relative part density, dimensional accuracy, and mechanical properties for an entire part. However, to analyze small defects with high spatial resolution, MEX-specific characteristics must be considered. For example, unsupported bridging results in changes in geometric tolerances.
For the description of part characteristics, general standards such as the geometrical product specification matrix model [283] are applicable. Here, surface imperfections in the layer structure can be characterized according to the ISO 8785 standard, which specifies the nomenclature and characteristics of these irregularities [284]. Furthermore, standards for conventional non-destructive testing methods can be adapted to the process monitoring of MEX [14].

Conclusion
Monitoring of MEX during the manufacturing process is crucial for the industrial use of this technology. The publication activity in this field is increasing. This clearly indicates that the subject is significant. The wide range of sensor technologies used and quality characteristics monitored demonstrate that the existing monitoring systems have been researched at many functional levels. However, for the widespread utilization of monitoring systems, further optimization is required.
The strength of this review is in its systematic approach to the literature search and the large dataset used. The state of knowledge is presented comprehensively, and research gaps are identified. Limitations exist because of the possibility that the literature evaluation and identification of future priorities are affected by the individual perspectives of the authors. For a highly differentiated analysis of the publications, future reviews may also include more systematic and detailed assessments of the results and quality of studies. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/.