Quality Control of Metal Additive Manufacturing

Abstract

Metal Additive Manufacturing (AM) is an emerging technology for rapid prototype manufacturing, and the structural integrity of printed structures is extremely important and should meet the specifications and high standards of the above industries. In several metal AM techniques, residual stresses and micro-cracks that occur during the manufacturing procedure can result in irreversible damage and structural failure of the object after its manufacturing. Thus effective quality control of AM is highly required. Most Non-Destructive Testing (NDT) techniques (X-Ray, Computed Tomography, Thermography) are ineffective in detecting residual stresses. Bulk, cost, and resolution are limitations of such technologies. These methods are time consuming both for data acquisition and data analysis and have not yet been successfully integrated into AM technology. However two sets of NDT techniques: Electromagnetic Acoustic Transducers (EMAT) and Eddy Current (EC) Testing, can be applied for residual stress detection for AM techniques. Therefore a crucial and novel extension system incorporation of big data collection from sensors of the both techniques and analysis through machine learning (ML) can estimate the likelihood of the AM techniques to introduce anomalies into the printed structures, which can be used as an on-line monitoring and detection system to control the quality of AM.

1. Introduction

Additive manufacturing (AM) is a process to build complex 3D parts from computer-aided design (CAD) models through layer-by-layer or drop-by- drop deposition of materials [1]. It is an emerging technology due to its capability to build complex-shaped products with less tooling and production time [2, 3]. AM provides significant advantages over traditional subtractive (machining) and formative (casting, moulding) manufacturing processes, such as reducing material waste, eliminating specialised tooling cost, and enabling the creation of intricate and free-form geometries. Lower prices of AM technologies make it more accessible to industries [4]. Reducing costs of material, higher level of automatization with less intervention of humans and increasing of manufacturing quality are also impulses to expand AM in further sectors of industry. AM industries have grown rapidly since 2000 [5], and have shown almost six times the growth during the 2000s as compared with the growth during the 1990s [5]. It was reported that in 2013 AM system sales revenue of industrial products, consumer products, automotive, medical, aerospace and military are 19%, 18%, 17%, 14%, 12% and 5% separately. Automotive, medical, aerospace and military, which require high precision and reliability, lead 48% in total [6]. It has been estimated that the global market for AM processes and services will reach around $50 billion by the year of 2031 [7]. Nowadays, industries invest 10 times more on end-part production than on prototyping [8]. Therefore, final and functional part productions and relevant researches are growing faster than the general market.

Although AM technology has been dramatically developed and it brings high feasible applications into real industry, there are still many obstacles in the adoption of AM for reliable production, such as part quality inconsistency, repeatability, and absence of material process standard [9]. For example, process variations and uncertain factors significantly impact the microstructure and mechanical properties of AM builds, which will further lead to internal defects deteriorating the build hardness, strength, and residual stress [10]. Process monitoring is a big challenge in AM as different factors affect the monitoring from materials to geometries to hardware and software limitations [11]. AM even with consistent process parameters is affected by little variations [12] in air flow, melting pool, etc. Due to its high-quality variability particularly for critical industries such as automotive and aerospace, high precision and mechanical properties certification is required [5].

For quality control improvement, numerous researchers have adopted different analyses focused on quality-approached technologies. For example, the effect of printing parameters on precision and internal cavity of fabricated parts using fused deposition modelling 3D printer was studied in [13]. Another research [14] investigated the effects of layer printing delay on physical and mechanical properties. Yang et al. [15] identified that process monitoring with closed-loop process control allows to achieve consistent quality of manufactured parts. Due to the complexity of AM processes, more and more complex real-time monitoring technologies are investigated. A non-exhaustive list includes Thermal camera, high-speed optical camera, photodetector, pyrometer, acoustic emission. Advanced sensing leads to the generation of big and complex data. The full exploitation of the data is critical to understanding the quality variability during the printing. Advanced signal processing/deep learning frameworks are required to achieve near-real-time defect detection. As a result, there are increasing interests and rapid development of sensor-based models for the characterisation and estimation of defects in the past few years [10]. Paper [15] reviewed several experimental configurations and adopted the vision-based and thermal sensing metrology approaches for the in situ process monitoring of typical material variation and failure modes. A spectral-graph approach was proposed to study the photo-detector sensor signature for the identification of defects caused by material cross-contamination in LPBF AM process [16].

Depending on the target material and involved energy source, an AM system (phase 4 of Figure 1a) entails different types of processes. One of the most popular in industry is powder bed fusion (PBF) which typically uses a high source of thermal energy to melt powder-based materials (such as metal) into desired structures and shapes [18]. For applications targeting the production of metal components, the most diffused PBF technique exploits a laser as the heat source to either melt or sinter the metal powder together (see a schematic representation in Figure 1b). In several metal AM techniques, e.g. selective laser melting (SLM), electron beam additive manufacturing (EBAM) and wire arc additive manufacturing (WAAM), residual stresses and micro-cracks that occur during the manufacturing procedure can result in irreversible damage and structural failure of the object after its manufacturing. Repetitive faults which occur during manufacturing due to incorrect estimation of appropriate operating conditions of the printer should be eliminated, as any waste is undesirable and costly for a company. Thus effective QC of metal AM is highly required. Currently there are many different monitoring techniques in detecting residual stresses and defects for QC of metal AM, for example, thermography, X-ray computed tomography (CT scan), eddy current (EC) and electromagnetic acoustic transducers (EMAT). Therefore, this chapter discussed and investigated different techniques for the QC of metal AM technology separately. Sections 2 and 3 reviewed temperature monitoring and CT scan techniques while Sections 4 and 5 investigated EC testing and EMAT testing methods.

Figure 1.

Additive manufacturing in industry. (a) Main phases of a typical process. (b) Scheme of laser powder bed fusion [17].

2. Temperature monitoring

Kim et al. [19] studied mechanical property improvement of Ti-6Al-4 V alloy part printed by EBM. Increasing bed temperature or post-processing the printed parts (e.g. peening) allows improving the mechanical properties. Although useful, this approach requires modification of the process or adding another process which increases the cost. The complexity of the geometry makes it also not easily generalisable.

It is much more agile to dynamically optimise the process parameters thanks to quality control sensors [19]. A video microscopy system was used to observe sintering and flow behaviour in real-time, which can evaluate different sintering characteristics of the materials [20]. An infrared light thermal sensor was developed to control laser power for better uniform sintering performance [21]. Paper [22] reviewed literatures on the thermal modelling method in SLS and SLM techniques. It was summarised that uniform temperature distribution of fields during printing processes leads to better quality; there is a need to provide information from monitoring temperature of the melt pool to be able to control process parameters for part quality. As a temperature monitoring system, pyrometers and thermocouples were used for monitoring its temperature [22].

Another way of monitoring temperature is through emitted thermal radiation (pyrometry).

This is achieved using either photodiodes or CMOS/CCD digital cameras which convert radiation and light into electric signal. Schürmann et al. [23] developed a novel laser cladding head with photodiodes and digital camera (CCD) to enable real time condition monitoring. A high resolution CCD camera was used for surface error detection through image processing, as shown in Figure 2 [24]. Research [25] analysed images of melt pool size to adjust laser output power by using a CMOS camera with an additional illumination source required for high scanning velocities, resolution, as well as photodiodes. A thermal model was developed that the temperature evolution and sintering formation can be simulated by a 3D FEA to predict thermal properties (i.e. thermal conductivity and specific heat) [26].

Figure 2.

(a) Setup of the CCD camera system in front of machine window [24] and (b) setup of CMOS camera system with illumination source and photodiode for high scanning velocities and resolution [25].

Unlike pyrometers, thermocouples are a contact measuring technique thus inconvenient for additive manufacturing where the part is changing over time. Nonetheless, thermocouples were used to monitor the temperature on the powder bed [27]. In addition, research [28] used thermocouple attached to the bottom of the base plate with a strain gauge to record residual stresses, and paper [29] also used thermocouples positioned under the bed’s surface to monitor energy absorption and powder effective conductivity for better understanding of heat transfer in metal powder during laser processing of the powder bed. A thermocouple control system was invented to improve uniform distribution of temperature on powder bed during part build. The thermocouple was attached inside of the powder bed along with IR sensor and communicates with temperature transmitter through circuitry in real-time [30].

Besides, research was focused on the IR thermography based monitoring and control system incorporated in the EBM system. Research [31] demonstrated the feasibility of using a near IR thermal camera for temperature measurement in hatch melting, preheating and contour melting events during the EBS process. A galvanometric scanner system was used for temperature distribution monitoring of melt pool of Ti6Al4V alloy in the SLM system [32]. Paper [33] developed continuous data capturing method using the IR camera to demonstrate feasible work to detect porosities inside materials and understand thermal phenomena such as it happens when beam and powder interact with each other. A temperature feedback control system was developed by positioning photodiodes and CMOS digital camera on laser beam to stabilise melt pool temperature distribution in SLM system [34]. In 2013, Mireles developed automatic IR camera monitoring system for defect detection allowing process stoppage when porosity exceeds a threshold. In addition, the work allowed to stablise temperature and optimise the process automatically using image processing algorithms [35]. In 2015, Mireles developed an in-line inspection system using IR thermography applied to an EBM machine. This allowed measuring defects geometry and locating it. A re-melt allows porosity defects repair as shown in Figure 3 [36]. Compared to mechanical processes such as hot isostatic pressing, the in-situ monitoring and repair allows saving total manufacturing time and avoiding microstructure alteration.

Figure 3.

Correction of un-melted powder through layer re-melt shown by (a) defects detection and localisation, (b) defect labelling, (c) repair after re-melting process [36].

3. CT scan

Micro X-ray computed tomography (CT) is an advanced measuring technique that can be used for characterising AM components [37, 38]. Examples of an XCT slice of an aluminium sand cast engine cylinder head and of porosities are shown in [37]. Metrological CT systems allow accurate dimensional measurements [39] as well as internal and surface defect detection sizing and localisation (e.g. voids and inclusions) [40]. CT is an excellent technique allowing topographical measurements at micro-scale achieving 100% volume coverage [41, 42]. Due to the cost, radiation hazard and bulk and lengthy scanning time it is not a monitoring technique but more applicable as a post printing quality control technique (Figure 4).

Figure 4.

Examples of an XCT slice of an aluminium sand cast engine cylinder head and of porosities [37].

Zanini et al. [43] has carried out experimental investigations on different SLM parts made of Ti6Al4V to address the accuracy of CT-based evaluations of AM. Based on samples with calibrated defects, it was proven that CT porosity measurements evaluate internal porosity with measurement errors below 5 μm for diameter and below 5% for volume measurements. However CT scan is not suited for residual stress detection and is time consuming both for data acquisition and data analysis due to slice by slice visual interpretation requirements as shown in Figure 5.

4. Eddy current monitoring

Eddy current non-destructive testing (ECNDT) as shown in Figure 5 is currently used for surface defects detection in metals. It is efficient and reliable. The ECNDT research is focusing on improving eddy current transducers (ECT) and the development of new methods for transducer signals processing [44, 45]. Previous work in this field has shown that metal AM parts can be tested with commercially available eddy current testing equipment (Figure 5) [46].

Figure 5.

Eddy current non-destructive testing [47].

Paper [48] has further discussed the applicability of EC testing with magnetoresistive sensors for laser PBF parts using giant magnetoresistance arrays in combination with a single wire excitation coil. To evaluate the influence of the powder used in the manufacturing process on EC testing and vice versa, a laser PBF mock-up made from stainless steel powder (316L) is used with artificial surface defects down to 100 μm. This laser PBF specimen was then examined using eddy current testing and the underlying principles.

4.1 Case study I

The first case study was based on an EC dataset collected from a sample with defects through in-line scanning tests in laboratory. The in-line scanning data was pre-processed and manually labelled, then two different ML algorithms (time series data processing and image data processing) for anomaly detection were investigated for the dataset.

4.1.1 Pre-processing

As each raw data file saved a lot of in-line scanning data as shown in Figure 6, a program of using a sliding window to automatically split the raw data into multiple sub datasets was developed to pre- process the raw data. Then each subset was labelled manually while examples of each labelled class are shown in Figure 7. The number of each classes was summarised in Table 1. Finally the labelled data was prepared for the following ML model training and testing.

Figure 6.

Data pre-processing.

Figure 7.

Examples of labelled data. (a) 0.2 mm, (b) 0.5 mm, (c) 1 mm, (d) edge and (e) normal.

Class	Number
0.2 mm	956
0.5 mm	979
1.0 mm	985
Edge	2118
Normal test	7930

Table 1.

Data number of in-line scanning dataset.

4.1.2 ML algorithms

Based on the above labelled data, two different ML algorithms (time series raw data processing and image data processing) were investigated separately and each ML algorithm was implemented for two objectives: anomaly detection and defect type classification. The anomaly detection is able to classify two classes: Normal class which comprises ‘edge’ and ‘normal’ data and abnormal class including ‘0.2 mm’, ‘0.5 mm’ and ‘1.0 mm’ defects. The defect type classification aims to identify all the different classes in Table 1.

4.1.2.1 Time series raw data processing

As the EC data is time series data, it was feed directly into training and testing the ML algorithms based on XGBOOST model. Results of the confusion matrix, accuracy, precision, recall and f1 score for anomaly detection and defect type classification are shown in Figure 8. In the images, scale number means the number of images and the number in each class means the image number been classified as the class. For example, in Figure 8a, 575 and 1994 images were correctly predicted as “Abnormal” and “Normal” separately while only 21 and 4 images were predicted incorrectly for the two classes. The results indicate good performance of the both models due to 99% accuracy for anomaly detection and 98% accuracy for defect type classification.

Figure 8.

ML results based on time series data format. (a) Anomaly detection, (b) defect type classification.

4.1.2.2 Image data processing

The EC time series data was converted into images so that convolution neural network (CNN) as a subset of ML, which has the advantage of image classification can be investigated for the two objectives. Resnet CNN model was trained and tested based on the image data. Results of the confusion matrix, accuracy, precision, recall and f1 score for anomaly detection and defect type classification are shown in Figure 9. For example, in Figure 9a, 288 and 1000 images were correctly predicted as “Abnormal” and “Normal” separately while only 5 and 6 images were predicted incorrectly for the two classes. Both results demonstrate very good performance of the models resulting from 99% accuracy for both objectives.

Figure 9.

ML results based on image data format. (a) Anomaly detection, (b) defect type classification.

Although both of the ML results based on the two data formats has excellent performance (above 98% accuracy), Resnet model based on image data format has slightly better performance due to 99% accuracy for both objectives (anomaly detection and defect type classification).

4.2 Case study II

The second case study was based on an EC dataset collected from a sample with defects through 2D scanning tests in laboratory. The 2D scanning EC data was pre-processed and manually labelled before feed into ML algorithm for anomaly detection.

4.2.1 Pre-processing

The 2D scanning data is a large dataset, which was based on tests on a 100 mm by 100 mm plate sample with different testing directions as shown in Figure 10a. Each 2D scan has multiple in-line scanning as shown in Figure 10b, for example, the green line means the in-line scanning at the height of 82 while the red line means the in-line scanning at the height of 52. A program of using a sliding window to automatically split the raw data into multiple sub datasets was developed to pre-process the 2D scanning data. Then each subset was converted into images before labelled manually while examples of each labelled class are shown in Figure 11. The number of each classes was summarised in Table 2. Finally the labelled data was prepared for the following ML model training and testing.

Figure 10.

2D scanning data. (a) Testing in different directions, (b) each 2D scan has multiple in-line scanning.

Figure 11.

Examples of labelled image data.

Class	Number
Abnormal	2237
Normal	53,653

Table 2.

Number of 2D scanning dataset.

4.2.2 ML algorithms

Based on the above labelled data, ML algorithm (Resnet model) based on image data was applied for anomaly detection. The anomaly detection aims to classify two classes: Normal class which comprises ‘edge’ and ‘normal’ data and abnormal class including different defects. Resnet model was trained and tested based on the image data. Results of the confusion matrix, accuracy, precision, recall and f1 score for anomaly detection are shown in Figure 12, for example, there were 309 and 8012 images correctly predicted as “Abnormal” and “Normal” separately while only 28 and 37 images were predicted incorrectly as the both classes. These results indicate great performance of the model due to 99% accuracy.

Figure 12.

ML results for anomaly detection.

4.3 Summary

Above two case studies based on EC testing which collected two datasets with different formats: in- line scanning and 2D scanning, have been investigated. With respect to the in-line scanning data, two different ML algorithms (XGBOOST model for time series data processing and Resnet model for image data processing) have been applied for anomaly detection. Results indicate both ML models have good performance due to 99% accuracy. With respect to the 2D scanning data, the raw data was converted into images before fed into ML training and testing for anomaly detection, which results of accuracy above 90% demonstrate the good performance of the ML models. Therefore both high accuracy anomaly detection results indicate that EC testing can be applied for residual stress detection to control the quality of metal AM.

5. Electromagnetic acoustic transducers monitoring

EMATs are devices made up of coils fed by a large dynamic current (a pulse or a tone-burst are commonly used) and a magnet or electromagnet providing a static magnetic field [49, 50, 51, 52, 53] as shown in Figure 13. They are used on metallic specimens based on Lorenz force as shown in Figure 13. For ferromagnetic metals, both Lorenz force and a magneto-strictive transduction mechanism occur.

Although contactless, EMAT require extreme proximity to the tested specimen (mm to few mms) to generate energy. EMATs have good repeatability since there is no external couplant. Common modes for non-ferrous materials include shear waves [49] making EMATs an ideal candidate for studying in- situ the interaction of shear ultrasonic waves with AM components (Figure 13).

Figure 13.

EMAT operation inspection [54].

5.1 Case study I

A case study was based on an EMAT data through testing in a CNC machine with samples in laboratory. Then signal processing algorithms and ML algorithms were developed for different tasks: automatic signal processing algorithm was developed based on filtering and peak detection methods to calculate acoustic birefringence results while different ML models based on EMAT time series plotting data and EMAT short-time Fourier transform (STFT) image data were developed for good/substandard measurement detection.

5.1.1 Signal processing

EMAT data has different measurements before and after additive printing while it also has good and substandard measurements, as summarised in Table 3. Each EMAT data file has two channels records while the beginning of the waveform has large variation, as example plot shown in Figure 14.

File name	File number
Steel pipe after printing left of new material (type 1 measurement)	9
Steel pipe after printing on new material (type 2 measurement)	46
Steel pipe after printing on new material2 (type 3 measurement)	45
Steel pipe before printing—bad data	45
Steel pipe before printing full axial scan (type 4 measurement)	45
Steel pipe before printing—poor data	47

Table 3.

EMAT data files.

Figure 14.

EMAT data plot.

Based on the FFT analysis of the raw data as shown in Figure 15, it indicates that there are a lot of noise which frequency is higher than around 4/5 MHz. Thus automatic pre-processing program was developed through applying low pass filters (either 4 MHz filter or 5 MHz filter), as shown in Figure 16. This figure demonstrate that both of the filters have similar results as the red curve and black curve are close to each other. Then peak detection algorithm was developed to detect the peaks automatically (shown in Figure 17). The pattern of each peak difference between two channels for different measurements before and after additive printing are plotted in Figure 18. The pattern shows that all the measurements have the similar increasing rate of the peak difference. This is because of the increased time difference along the peaks resulting from different velocities of both waves after more reflections.

Figure 15.

EMAT data FFT analysis.

Figure 16.

EMAT data pre-processing through low path filters.

Figure 17.

Peak detection.

Figure 18.

Pattern of peak difference for different measurements.

After pre-processing the data, acoustic birefringence (AB) was calculated through following function:

$$\mathit{AB}=\frac{t_s-t_f}{\left(\frac{t_s+t_f}{2}\right)}$$E1

ts = arrival time of ‘slow angle’ wavetf = arrival time of ‘fast angle’ wave

The arrival time is the time at maximum point of each peak. The time difference is a direct indication of change in velocity (assuming sample is the same thickness across the probe area). It is assumed that a change in velocity is an indication of residual stress change.

AB results for all the different measurements before and after printing are shown in Figure 19, which indicates that AB results decrease monotonically alone the peaks. The boxplots of AB results of the four types of measurements before and after printing for all the six peaks are shown in Figure 20, which demonstrates that all six peaks have the similar pattern: AB results of type 2&3 measurements have the same level and lower than that of type 1&4 measurement. These results are consistent with the facts that type 2&3 measurements were tested on the similar new materials after printing, which are different with type 1&4 measurements as type 1 measurements were tested before printing and type 4 measurements were tested at the edge after printing new material. This indicates that AB results can indicate the different residual stresses in a certain level thus as an indicator of QC of metal AM.

Figure 19.

AB results for all the different measurements before and after printing.

Figure 20.

AB results of the four types of measurements before and after printing for all the six peaks.

5.1.2 ML algorithms

Based on the good and substandard EMAT data, the raw data was converted into images through two pre-processing methods, then different ML algorithms based on the two image datasets were applied for good/substandard classification. Table 4 lists the number of each class. The first pre-processing method is plotting the time series data directly in a figure as shown in Figure 21a while the second pre-processing method is convert the time series data into STFT images as shown in Figure 21b.

Figure 21.

EMAT image data. (a) Time series plotting image, (b) STFT image.

STFT is a frequency-time transform of a time signal and an effective tool to analyse the non- stationary signals because of the avoiding of severe interference by the cross-terms [55]. The basic principle for STFT is as follows: divide these characteristic signal into small time intervals, use the Fourier transformation to analyse separately each time interval. The definition of STFT for signal s(t) is:

$$S_t\left(\omega \right)=\frac{1}{\sqrt{2\pi }}\int s\left(\tau \right)h\left(\tau -t\right)e^{\mathit{j\omega \tau }}\mathit{d\tau }$$E2

The two formats of datasets were used to train and test ML algorithms based on Resnet model. Results of the ML models show in Figure 22, for example, there were 23 and 15 images correctly predicted as “Good” and “Substandard” measurements separately while no image was predicted incorrectly. The results indicate excellent performance due to 100% accuracy for both of data formats (Table 4).

Figure 22.

ML results. (a) Based on time series plotting image, (b) based on STFT image.

Class	Number
Good data	217
Substandard data	134

Table 4.

Total good/substandard data.

5.2 Summary

Based on above case study of MEAT testing, automatic signal processing algorithm was developed based on filtering and peak detection methods to calculate acoustic birefringence, which results of type 2&3 measurements have the same level and lower than that of type 1&4 measurement. These results are consistent with the facts that type 2&3 measurements were tested on the similar new materials after printing, which are different with type 1&4 measurements as type 1 measurements were tested before printing and type 4 measurements were tested at the edge after printing new material. This indicates that AB results can indicate the different residual stresses in a certain level thus as an indicator of QC of metal AM.

Additionally, different ML models based on EMAT time series plotting data and EMAT STFT image data were applied for good/substandard measurement detection, which results demonstrate excellent performance due to 100% accuracy.

6. Conclusion

Additive manufacturing technology has progressed a lot allowing to achieve higher parts quality and improved consistency although a lot of challenges remain with regards to consistency of quality, sensitivity to tiny process tolerances. Many techniques have been developed to enhance the quality: the application of post-processing of 3D printed parts, the use of monitoring sensors for early defect detection and the use of monitoring sensors with feedback loop for continuous process optimisation. In this chapter, many sensing techniques have been investigated for process monitoring. Thermography and CT scan are limited by the resolution of images, they are bulky and costly, and not suited to residual stress detection. These methods are time consuming both for data acquisition and data analysis (CT requires slice by slice visual interpretation) and have not yet been successfully integrated into AM technology due to health and safety standards and its poor resolution at the edges of AM structures. However investigation of the applications through EMAT and EC testing techniques shows that both of the methods can be applied for residual stress detection for meatal AM techniques.

Based on the case study of two sets of EC data which have two formats: in-line scanning and 2D scanning, different ML algorithms were applied for anomaly detection. With respect to the in-line scanning data, results indicate both of the two different ML algorithms (time series data processing and image data processing) have excellent performance due to 99% accuracy. With respect to the 2D scanning data, ML models have good performance resulting from accuracy above 90%. Therefore both high accuracy anomaly detection results indicate that EC testing can be applied for residual stress detection to control the quality of metal AM.

Based on above case study of MEAT testing, automatic signal processing algorithm was developed based on filtering and peak detection methods to calculate acoustic birefringence, which results of type 2&3 measurements have the same level and lower than that of type 1&4 measurement. These results are consistent with the facts that type 2&3 measurements were tested on the similar new materials after printing, which are different with type 1&4 measurements as type 1 measurements were tested before printing and type 4 measurements were tested at the edge after printing new material. This indicates that AB results can indicate the different residual stresses in a certain level thus as an indicator of QC of metal AM. Additionally, different ML models based on EMAT time series plotting data and EMAT STFT image data were applied for good/substandard measurement detection, which results demonstrate excellent performance due to 100% accuracy.

Above all, a crucial and novel extension system incorporation of big data collection from sensors of the both EC and EMAT techniques and analysis through ML can estimate the likelihood of the metal AM techniques to introduce anomalies into the printed structures before the beginning of the manufacturing, thus the system can be used as an on-line monitoring and detection system to control the quality of metal AM.