Additive Manufacturing 21 (2018) 598–604

Contents lists available at ScienceDirect

Additive Manufacturing
journal homepage: www.elsevier.com/locate/addma

Acoustic emission for in situ quality monitoring in additive

manufacturing using spectral convolutional neural networks
S.A. Shevchik, C. Kenel, C. Leinenbach, K. Wasmer ∗
Empa, Swiss Federal Laboratories for Materials Science & Technology, Laboratory for Advanced Materials Processing, Feuerwerkerstrasse 39, CH-3602
Thun, Switzerland

a r t i c l e i n f o a b s t r a c t

Article history: Additive manufacturing, also known as 3D printing, is a new technology that obliterates the geomet-
Received 22 March 2017 rical limits of the produced workpieces and promises low running costs as compared to traditional
Received in revised form 2 October 2017 manufacturing methods. Hence, additive manufacturing technology has high expectations in industry.
Accepted 30 November 2017
Unfortunately, the lack of a proper quality monitoring prohibits the penetration of this technology into
Available online 6 December 2017
an extensive practice. This work investigates the feasibility of using acoustic emission for quality mon-
itoring and combines a sensitive acoustic emission sensor with machine learning. The acoustic signals
Keywords:
were recorded using a fiber Bragg grating sensor during the powder bed additive manufacturing process
In situ quality monitoring
Acoustic emission
in a commercially available selective laser melting machine. The process parameters were intentionally
Spectral neural networks tuned to invoke different processing regimes that lead to the formation of different types and concen-
Wavelet transform trations of pores (1.42 ± 0.85 %, 0.3 ± 0.18 % and 0.07 ± 0.02 %) inside the workpiece. According to this
Additive manufacturing poor, medium and high part qualities were defined. The acoustic signals collected during processing were
Fiber Bragg gratings grouped accordingly and divided into two separate datasets; one for the training and one for the testing.
The acoustic features were the relative energies of the narrow frequency bands of the wavelet packet
transform, extracted from all the signals. The classifier, based on spectral convolutional neural network,
was trained to differentiate the acoustic features of dissimilar quality. The confidence in classifications
varies between 83 and 89 %. In view of the narrow range of porosity, the results can be considered as
promising and they showed the feasibility of the quality monitoring using acoustic emission with the
sub-layer spatial resolution.
© 2017 Elsevier B.V. All rights reserved.

1. Introduction very limited (e.g. austenitic stainless steel, eutectic Al-Si alloys).
This is the result of the extremely complex physical phenomena
Selective laser melting (SLM) is a powder bed AM technology, during the SLM process. It implies the very rapid consolidation of
which allows building components with complex 3D geometries the base material powder in a small material volume (∼0.001 mm3 )
from an alloy powder layer by layer. Other terms that describe the using a focused laser followed by cyclic heating and cooling at
same technology are Laser Cusing, Direct Metal Laser Melting, or high rates (103 –107 K/s) [11]. The processing parameters like laser
Laser Metal Fusing. The technology has been successfully applied energy, scan velocity, hatch distance, powder layer thickness, scan
for rapid prototyping of unique workpieces with highly complex strategy, etc. have to be carefully adjusted to the alloy of interest.
geometries that are impossible to produce using traditional form- An improper set of parameters can lead to pronounced porosity,
ing methods [1]. It also has clear advantages in small series and cracking and/or accumulation of residual stress inside the work-
individualized products for medical [2,3,4], turbine [5,6], aerospace piece [12], resulting finally in poor mechanical properties [13,14].
[7,8], robotics [9], automotive and machine tool industries [10]. The process repeatability is thus still limited, preventing the tech-
Even though a lot of progress has been made on behalf of nology from being used in a much wider range. A possible solution
the machine manufacturers to improve the process efficiency, the to overcome this situation is the development of an in situ and real-
number of materials that can be reliably processed is currently still time monitoring of the workpiece quality and to better control the
machine operation [12].
At present, the industrial standard for monitoring the workpiece
quality with regard to porosity or cracking is X-ray tomography
∗ Corresponding author.
[15]. The tests are carried out post mortem, when the machine time
E-mail address: [email protected] (K. Wasmer).

https://doi.org/10.1016/j.addma.2017.11.012
2214-8604/© 2017 Elsevier B.V. All rights reserved.
 S.A. Shevchik et al. / Additive Manufacturing 21 (2018) 598–604 599

and materials are already spent. This method is also known for in this work as the spatial approach does not provide the time order-
being very expensive and time-consuming. Consequently, online ing [29], which can be valuable information for the localization of
control is of a great demand and several approaches are seen the individual defects in the future. More information about the
to solve this problem. Temperature measurements of the pro- SCNN can be found in the two comprehensive reviews [29,30].
cessed zone with 1D or 2D pyrometers are used in multiple In this study, the AE signals were collected during the SLM of a
works [12,16,17]. The main objective is to keep the temperature stainless steel using a fiber Bragg grating. Different laser processing
parameters within a certain range that guarantees the mechan- parameters resulting in different part qualities with regard to the
ical properties of the workpiece to stay in an acceptable range. amount of porosity were selected. The AE signals were evaluated
This approach uses built-in models for data analysis and provides and the classification of the process in terms of “high”, “medium”
a high spatial resolution quality control. Unfortunately, in real-life and “poor” quality was tested using both, conventional CNNs and
conditions, the multiple non-uniformities of the laser-matter inter- SCNNs.
actions are responsible for a divergence between the models and
the real-life situation resulting in performance inaccuracies. Image
processing is another approach that is already introduced in a num- 2. Experimental setup and acoustic datasets
ber of the available AM machines [12,18,19]. Quality monitoring is
carried out with the visual inspection of each layer and an image 2.1. Data acquisition setup
processing routine searches for defects. The technical realization of
such system is performed using matrix photodetectors in the visible The SLM process of CL20ES stainless steel (1.4404/316L) powder
spectral range. Again, one major weakness of this system is that the (particle size: 10–45 ␮m) was performed on an industrial machine
quality is monitored after an entire layer is produced. Besides, this Concept M2 (Concept Laser GmbH, Germany). The machine is
method has certain limitations with regard to the spatial resolution equipped with a fiber laser operating in continuous mode at a wave-
and requires expensive optical systems to detect small defects. length of 1071 nm with a beam quality M2 = 1.02 and a spot size
The present work is a feasibility study that focuses on an alterna- of 90 ␮m. The Concept M2 was additionally equipped with one
tive approach, in which acoustic emission (AE) is taken for quality fiber Bragg grating (FBG) sensor to detect the airborne AE signals,
monitoring of the AM process. In particular, the aim was to iden- generated during the AM process.
tify the laser processing regimes that potentially induce different The FBG sensor is an interferometric structure, imprinted inside
types (e.g. due to incomplete particle melting or due to evaporation) the core of the optical fiber with unique spectral characteristics
and concentrations of porosity. The attractiveness of AE signals as of reflectivity. The acoustic waves created during the AM process
compared to other methods lies in the high sensitivity of existing result in periodical extension/compression of the optical fiber core
AE sensors and the relatively cheap hardware. The processing of AE and, consequently, the FBG structure. These momentary deforma-
signals is fast as it uses 1D data as compared to imaging (2D data) or tions affect the reflectivity properties of the FBG that follow the
tomography (3D data). Finally, AE sensors provide a high temporal behavior of the upcoming pressure waves. This behavior results
resolution allowing to precisely localize the defects. The applica- in the intensity of the reflected light that encodes the momentary
bility of AE to detect defects during AM was already shown by Wu deformation states of the fiber core. These sensors are character-
et al. [20] for polymer materials, although no in situ and real-time ized by a linear response in a broad frequency range of 0–60 GHz
monitoring was proposed. Fiber Bragg gratings (FBG) are often used (as compared to the frequency range of approximately 20 KHz
for detecting AE signals because of their high sensitivity in a wide −1.2 GHz with piezo technology). The basic models of FBG are
acoustic spectral range [21]. The combination of high sensitive AE cheaper than piezo sensors and are commercially available in
sensors and machine learning (ML) has been already successfully different configurations. The extreme sensitivity of FBGs allows
applied to a number of applications, such as tribology [22,23] and detecting even the thermal fluctuations of the environment (that is
fracture mechanics [24]. All these works are characterized by com- out of the sensitivity range of piezo technology) with sub nanosec-
plexity in acquired data structure and noisy environments, thus ond time resolution. All this make those sensors very flexible for
making this combination promising for AM, as well. acoustic applications and therefore they were chosen in this work.
In recent years, convolution neural networks (CNN) has become More details about FBGs can be found in [31,32].
a popular technique for acoustic tasks [25] showing high efficiency The FBG was placed inside the process chamber at a distance
and possibilities to suppress the stationary noises. Nevertheless, of 20 cm away from the processed zone. This simple setup was
the conventional CNNs have certain limits when analyzing non- selected in order to demonstrate the feasibility of using AE signals
regular data. Spectral convolutional neural networks (SCNN) are for in situ process monitoring under realistic manufacturing con-
a recent extent of conventional CNN with improved efficiency in ditions and without modifying the conception of the machine. The
classification/regression tasks that are operating on irregular data FBG sensor was pre-strained in a special plastic holder and installed
grids [26,27,28]. High performance of SCNN is achieved combining on the lateral wall inside the working chamber of the machine as
the best from: i) deep learning, inherited from conventional CNNs, shown in Fig. 1a. Its connection to the read out system was provided
and ii) spectral graph theory [26]. In the latter case, the local spatial via an optical feedthrough mounted on a plate that hermetically
structure of the input data is recovered using a weighted graph G. closed the working chamber. The employed FBG sensor possessed
The learning of the local irregularities in SCNN is carried out using 50 % reflectivity at 1547 nm wavelength with a reflectivity band
the convolution operator, which is realized by applying G upon of ±1 nm. The read out system included a tunable laser source
the input signals f. In conventional CNN, the spatial convolution with an output light power of 4 mW and a band of single mode of
using regular kernels fails when operating on the same type of data ± 0.05 nm (Yenista Inc., USA). The FBG read out was carried out by
[26,27,28]. Currently, the SCNN methods are mainly reported for pumping the fiber with a continuous narrow band irradiation at a
image processing [26,27,28]. However, the representation of acous- wavelength of 1546.8 nm and the reflected signals were digitized
tic signals in the time-frequency domains using wavelets allows by a high speed photodiode. A general schematic of the FBG system
applying already developed 2D image processing methods, making is shown in Fig. 1b. The received signals were recorded through an
SCNN a ready “on-shelf” solution for acoustic analysis. acoustic data acquisition system connected to a personal computer
The modelling of the data irregularities in SCNN is possible using with dedicated software from Vallen (Vallen Gmbh., Germany). The
spatial and spectral approaches, and both exhibit the same classifi- sampling rate of the originally digitized signals was set to 10 MHz.
cation efficiency [29]. However, the spectral approach was chosen However, to fit the dynamic range and reduce the computational
600 S.A. Shevchik et al. / Additive Manufacturing 21 (2018) 598–604

Fig. 1. a) view of the FBG location inside the AM chamber with the optical feedthrough on the chamber panel (left) and the FBG read out system (right); b) scheme of the
FBG read out system.

load further, those signals were down-sampled to 1 MHz. The data (with z being the build direction) using image thresholding and area
acquisition system was synchronized with the laser of the Concept determination with Fiji (ZITAT FIJI) [34]. Fig. 2b–d are representa-
M2 machine, providing a precise registration of the AE signals from tive cross-section images obtained from the regions processed with
each manufactured layer. This procedure allowed neglecting the different scanning velocities. The optically measured porosities
noise from the intermediate manufacturing processes such as the were 0.3 ± 0.18 % (Fig. 2b, medium quality, 132 J/mm3 , 300 mm/s),
deposition of a new powder layer with the moving blade. Simul- 0.07 ± 0.02 % (Fig. 2c, high quality, 79 J/mm3 , 500 mm/s), and
taneously, several recordings were made with a higher resolution 1.42 ± 0.85 % (Fig. 2d, poor quality, 50 J/mm3 , 800 mm/s). The
at sampling rate of 1 GHz using an oscilloscope Lecroy HDO4104 largest energy input (132 J/mm3 ) creates porosity typically asso-
(Teledyne Lecroy, USA). Those signals were used to evaluate pos- ciated with the formation of deep keyhole channels which lead to
sible changes in the AE signal and noise parameters but were not tubular defects with a diameter of around 30 ␮m (Fig. 2b). With a
included into the training and tests data sets. decreased energy input to 79 J/mm3 , these defects disappear and
only a very few defects resulting from lack of fusion and remain-
2.2. Data collection and quality categorization ing gas porosity are observed (Fig. 2c). At the lowest energy input
(50 J/mm3 ), defects associated with insufficient bonding or lack of
The acoustic data collection was carried out throughout the fusion becomes the predominant defect type in the form of large
manufacturing process of the entire workpiece. Its geometry was cavernous voids ranging in size from 10 to 100 ␮m (Fig. 2d). For
defined by a 10 × 10 × 20 mm3 cuboid. The processing parame- more details about the physics of pores formation, the reader can
ters were adapted to create a workpiece consisting of sections be referred to [33].
with various quality levels. The parameters of the laser power P, To summarize, the changes in the laser scanning parameters
hatch distance h, and powder layer thickness t were kept constant resulted in different types, sizes and concentrations of defects. The
at 125 W, 0.105 mm and 0.03 mm, respectively. Three scanning selected scanning velocities were chosen to recreate specific and
velocities v were used to create samples with incomplete particle undesirable defects during an AM manufacturing process.
melting at low energy density (50 J/mm3 , 800 mm/s), optimal pro-
cess condition at a medium energy density (79 J/mm3 , 500 mm/s) 3. Data processing
and increased porosity due to local evaporation at high energy
input (132 J/mm3 , 300 mm/s). The energy density was determined 3.1. Features extraction
according to Eq. (1) [33]:
P In the present work, the relative energies of the narrow fre-
E= (1) quency bands were taken as the input features for the SCNN
v·h·t
classifier. The frequency bands were extracted using a standard
A meander scanning strategy oriented parallel to the part edges
wavelet packet transform (WPT) [35]. WPT is an extension of the
was used. Every layer was scanned in a 5 × 5 mm chessboard pat-
traditional wavelet transform that can be represented as a pass of
tern with a 1 × 1 mm shift between the subsequent layers. The
the signal f through a set of filters [35]:
experiment was performed in a N2 atmosphere with an O2 content
controlled to be below 1 % throughout the process.
 √
ϕj (n) = h0 (k) M␸(Mn − k), k ⊂ Z (2)
The produced sample is shown in Fig. 2a. The sample regions
n
corresponding to different laser processing regimes can be identi-
fied from the color differences. Those are related to the local energy
 √
ji (n) = hm−1 (k) M (Mn − k), k ⊂ Z (3)
input and ultimately to the porosity concentrations. The local
n
porosity was characterized by visual inspection of cross-sections
prepared from the sample. Light microscopy images correspond- where ho is a low pass and hm high pass filters, ϕ() and () are the
ing to regions with different sample quality are presented in Fig. 2. scale and wavelet functions, respectively, j is a scale, and n is the
Those were obtained with a Zeiss Axioinvert 100 microscope at current sampling point of the digitized signal f. The parameter m is
a magnification of 20 x on polished cross-sections in the x-z plane the total number of filter channels and in traditional WPT is equal to
 S.A. Shevchik et al. / Additive Manufacturing 21 (2018) 598–604 601

Fig. 2. a) SLM test workpiece produced with three energy densities where 50 J/mm3 are bright regions, 79 J/mm3 are brawn regions and 132 J/mm3 are dark/blueish regions;
b) – d) Typical light microscope cross-section images of regions produced with b) 132 J/mm3 (medium quality), c) 79 J/mm3 (high quality) and d) 50 J/mm3 (poor quality).
The insets show typical defects observed such as b) tubular, c) small lack-of-fusion and d) large lack-of-fusion defects. The z- and y-directions correspond to a).

two, while the maximum number of scales, in this case, is equal to riorating the classification efficiency. Several time spans for RW
log2 m. The application of Eqs (2) and (3) result in the extraction of were investigated and the corresponding details are discussed in
low and high frequency bands of the digitized signal f, where each the results and discussion section.
frequency band is represented by the coefficients d and is localized
in both, time and frequency domains. The multiscale analysis using
WPT is carried out by applying Eqs (2) and (3) at different scales
3.2. Spectral convolutional neural networks (SCNN)
j to both, the low and high frequency bands, extracted at the pre-
vious scale j-1. The energies of the narrow frequency bands were
In the SCNN, the convolution operation is carried out on a graph.
employed and computed as:
In this work, the spectral approach developed by Mathieu et al.
 
2 2
[26] was used. In this approach, the data structure is investigated
Ej,m = |dj,m (t)| dt = |dj,m | (4) using the Laplacian L on the graph G. The weighted graph G for the
k set of elements E = {ei , i = 1,. . .,n} is defined as G = {E, W}, where
Following Eq. (4), the relative energies are the normalized ver- W = {wij , i,j = 1,. . .,n} is a symmetric, positively defined matrix, in
sion of the sub-band energies so that Eq. (4) becomes: which each element wij characterizes the similarities between
a pair of elements {ei , ej }. In this study, the similarity was
Ej,m  
norm j,m = (5) computed using a heat kernel: wij = exp  ei − ej 2 /t [29]. The
Ej un-normalized Laplacian of G was employed and defined  asL = D
− W, whereas the elements of the diagonal matrix D = dii are
 
M j

where Ej = |Ej,m |2 is the total energy accumulated in all fre- the sums of the raw elements of W: dii = wiy , y,i = 1,...,n . In the
k=0 k
quency bands at all scales. The result of Eqs (2)–(5) is a sparse SCNN, the features from the input dataset are considered as ele-
wavelet spectrogram. An example is presented in Fig. 3, where ments collection E and the Laplacian graph can be computed for the
Fig. 3a is a collected AE signal and Fig. 3b the corresponding spec- entire input space as described above. The dimensionality of the
trogram, built from the relative energies of the narrow frequency input space n in E is defined by the number of the narrow frequency
bands, localized in time-frequency domain. Such spectrograms bands extracted as described in Section 3.1. The frequency domain
were the input for the SCNN and the details are discussed in Section for E is defined by a singular value decomposition of L, the product
3.2. Several wavelet families of Daubechies, Symlets and Coiflets of which is a matrix U that includes a set of eigenvectors [36]. For
were investigated with regard to their applicably to the acquired this case, the direct Fourier transform for G is defined as: Efft = UE,
AE data. The best choice was provided by Daubechies wavelet with while its invers is E = UT Efft [36]. The product of the convolution
ten vanishing moments which showed the minimum approxima- operation in SCNN is a set of feature maps, which are computed
tion errors on the given AE signals. For this reason, it was further using the Fast Fourier Transform (FFT) representation of the input
used for feature extraction. dataset Efft as C = UEfft k, where k are the spectral multipliers [26].
The analysis of the acquired AE signals was performed using The spectral multipliers in the SCNN are analogues to the weights
a running window (RW) that scanned the signal and which is in the kernel receptive fields in conventional CNNs that provide
schematically marked in red in Fig. 3a. The choice of the time the spatial localization of the elements E in the spectral domain.
span for the RW is a compromise between the spatial resolution The training of the SCNNs aims to tune the multipliers k to fit the
in defects detection and classification accuracy. On the one hand, output response of the SCNN to a target function. To do so, the
a short running window increases the spatial resolution in detec- gradients for the back propagation for a given set E are defined
tion of possible defect areas within each layer. On the other hand, by:  E = UUT  Ck, and the gradients for the updated spectral
very short time spans are more affected by noise and are thus dete- coefficients k are defined as:  k = UT  C UE [26,37,38].
602 S.A. Shevchik et al. / Additive Manufacturing 21 (2018) 598–604

Fig. 3. a) An example of a fragment of an AE signal that corresponds to a high quality layer produced with optimal process condition at medium energy density (79 J/mm3 ,
500 mm/s); b) the spectrogram corresponding to the relative energies of the narrow frequency bands, localized in time-frequency domain. The red marker in (a) shows the
running window (RW). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

The architecture of the SCNN is inherent from the one of conven-

tional CNNs that combines the convolution layers with the pooling
ones. As mentioned before, the product of the data pass through
the spectral convolution layer is a set of features maps. The follow-
ing pooling layer aggregates the information resulting in the sparse
representation of the data from the input feature maps.
The set of the pooling and the convolution layers provide a self-
feature extraction. The output of this structure is further forwarded
to fully connected layers, where the classification of the features
takes place and the results are observed in the output layer. More
details on the SCNN through FFT can be found in [26]. Fig. 4. Projection of the acoustic features into a 3D feature space using principle
component analysis (PCA). The visualized dataset includes thirty features from each
The training and testing of the SCNN were carried out in
quality level.
Microsoft Visual Studio C# environment, for which an original code
was designed. A single CPU with i5 processor was used.
Table 1
Classification tests accuracy results for the SCNN (conventional CNN).

Ground truth Poor Medium High

quality quality quality
4. Results and discussion Test category

4.1. Collected datasets Poor quality (1.42 ± 0.85 %, 800 mm/s) 89 (62) 4 (19) 7 (19)
Medium quality (0.3 ± 0.18 %, 300 mm/s) 5 (25) 85 (53) 10 (22)
High quality (0.07 ± 0.02 %, 500 mm/s) 8 (20) 9 (17) 83 (63)
The collected AE signals were divided into three categories
according to the manufacturing quality of the workpiece layers
described in Section 2.2 and are shown in Fig. 2. A number of pat- (79 J/mm3 , high quality) and 800 mm/s (50 J/mm3 , poor quality),
terns, bounded by the RW, were collected from the AE recorded respectively.
signals to form two datasets: one for training and one for the test. The SCNN for classification counted four convolutional lay-
Each category (that corresponded to poor, medium and high qual- ers that alternated with four pooling. The layer number was
ity) in each dataset was equally represented by 300 patterns with experimentally estimated as the best compromise between the
no common RW between both datasets. This approach simulated computational complexity and performance efficiency. A decrease
real-life conditions where the trained system has to operate with of the number of the convolutional layers revealed a slight diminu-
new input data. The features from all RW were extracted using tion of the classification accuracy. In contrast, an increase of the
the maximum possible number of scales (Cf. Section 3.1 for the layers number did not affect the classification performance but
explanation of the maximum scale). Furthermore, these features raised the computational time during the SCNN training.
were fed to the SCNN classifier. The time span of RW was selected The non-regularities in the structure of the input acoustic fea-
after an exhaustive search in which the classification accuracy of tures can be depicted by observing the collected features in the
SCNN was evaluated after each retraining using RWs with different feature space [39]. In the present work, the coordinates of each fea-
time spans. The goal was to find the shortest time span that did ture are defined by the values of the relative energy of the frequency
not affect the classification accuracy. This was achieved at 160 ms. bands (Cf. Section 3.1). In Fig. 4, a sparse dataset with thirty features
Further decrease of the RW time span significantly increased the of each quality category is shown and each point represents a posi-
classification errors. Each RW was decomposed using WPT at ten tion of an individual RW in the feature space. The dimensions of the
decomposition levels, resulting in the extraction of 2046 narrow feature space are reduced using the principle component analysis
frequency bands with their corresponding values of the relative (PCA) projection [40] into a lower 3D dimensional space. The data
energies (See Section 3.1). Taking into consideration the laser track irregularity can be observed as a mixture of the features from all
width, the layer thickness, the laser scanning velocity (See Section categories.
2.1) and the time span of 160 ms for the RW, the spatial resolution The classification results from the SCNN and the conventional
for the defect detection can be estimated. The setup presented is CNN (in brackets) are shown in Table 1. In this table, the sam-
able to provide a quality monitoring of 4.4 mm2 / 17.6·10-3 mm3 , ple quality (poor, medium and high qualities) (in rows) versus the
7.2 mm2 / 28.8·10-3 mm3 and 11.6 mm2 / 46.5·10-3 mm3 of the pro- ground truth (in columns) are given. The classification accuracies
cessing area [mm2 ] and powder volume processed [mm3 ] of a layer in the table are defined as the number of true positives divided by
produced with 300 mm/s (132 J/mm3 , medium quality), 500 mm/s the total number of tests for each category. These values are given
 S.A. Shevchik et al. / Additive Manufacturing 21 (2018) 598–604 603

in the diagonal cells of the table (dark grey cells). The classification ever, the authors decided to use airborne AE detection in order to
errors are computed as the number of the true negatives divided by determine the limits of the high sensible FBG as AE sensor. Never-
the total number of the tests for each category. These corresponding theless, the setup can be improved by taking two actions. First, we
values are filled in non-diagonal row cells. could use more sensitive acoustic sensors with selective filtering on
Based on the SCNN results, the classification confidence for the the hardware level. Second, we could attach the FBG to the build
different quality regimes varies in the range of 83 – 89 %. These plate. This would allow sensing pressure waves directly from the
results clearly show the potential of the proposed approach, in printed sample. In addition, two improvements are possible for the
particular when taking into account a non-optimized setup: the signal processing part. To start with, the involvement of deep SCNN
position of the FBG is by far not ideal, the acoustic emission (AE) with a greater number of layers gives additional advantages that
signals detected were airborne, and three parameter ranges result- were not exploited in the present study. The representation of data
ing in only slight differences in the sample porosity are investigated. through graphs allows realizing multiple mathematical operators,
Hence, we can conclude that the acoustic signal, recorded by the including wavelet and Fourier that can be additionally involved for
FBG, and processing with SCNN can be a versatile method for the noise reduction within the SCNN structure. Additionally, a multi-
in situ and real-time quality monitoring in AM. scale analysis of the input acoustic emission data can give some
The analysis of the classification errors structure can be evalu- benefits in detection of smaller size defects. All these actions are
ated from the non-diagonal rows in Table 1. For example, the AE test considered for further developments of an efficient and cost effec-
data from the high quality was classified with the lowest accuracy tive in situ and real-time quality control monitoring for additive
rate of 83 % and so it has the highest error rate. The classification manufacturing.
error is more or less equal between the medium and poor quali-
ties, 9 and 8 %, respectively. In contrast, the medium quality has a 5. Conclusions and future work
classification confidence of 85 % and the classification error is the
highest (10 %) for the high quality and the lowest (5 %) for the poor This work presents the results of a feasibility study for in situ
quality. The situation is completely invers for the poor quality with and real-time quality monitoring using acoustic emission (AE) and
a classification confidence of 89 % and the classification error is machine learning (ML). A fiber Bragg grating (FBG) sensor, mounted
the highest (7 %) for the high quality and the lowest (4 %) for the directly inside the process chamber 20 cm away from the process
medium quality. zone, provided a high sensitivity in signal detection. The airborne
The nature of the errors in Table 1 is in the non-distinct acoustic acoustic emission (AE) signals were detected and further grouped
patterns inside the signals from the different categories that affect in terms of AM processing quality. The Daubechies wavelet with
the classification performance. In order to understand the origin ten vanishing moments was used to decompose all signals and the
of the classification error, it is important to relate it to the pro- energies of the narrow frequency bands were taken as acoustic fea-
cess parameter that is the laser scanning velocity. As a reminder, tures. A spectral and conventional convolutional neural network
the slowest laser scanning velocity resulted in the medium qual- (SCNN and CNN) were involved to classify the features from differ-
ity (Fig. 2b, 300 mm/s, 0.3 ± 0.18 %, 132 J/mm3 ). The intermediate ent additive manufacturing qualities.
scan velocity led to high quality (Fig. 2c, 500 mm/s, 0.07 ± 0.02 %, The classification accuracies using SCNN are as high as 83, 85 and
79 J/mm3 ) whereas the highest scan velocity produced the poor 89 % for high, medium and poor workpiece qualities, respectively.
quality (Fig. 2d, 800 mm/s, 1.42 ± 0.85 %, 50 J/mm3 ). Taking into It was found that the lowest classification accuracy is in recogniz-
account this information when analyzing the results in Table 1 ing the high quality. This is due to the fact that the used scanning
described in the previous paragraph, the dependency of the clas- velocity for this category lies in between the scanning velocity for
sification errors with the laser scanning velocity becomes obvious. producing the two other qualities, resulting in an overlap of the AE
For a scan velocity of 300 mm/s (medium quality), the classifica- features. The use of SCNN showed higher classification accuracy as
tions error is the highest (10 %) at 500 mm/s (high quality) and compared to conventional CNN.
the lowest (5 %) at 800 mm/s (poor quality). The situation is com- Finally, today, AM process incorporates several independently
pletely invers for the poor quality where the classification error is developing techniques, such as stereo-lithography, laser sintering,
the highest (7 %) at 500 mm/s (high quality) and the lowest (4 %) at multi-jet printing, powder bed fusion and others. All those methods
300 mm/s (medium quality). Consequently, we can conclude that propose different technical solutions for AM realization, however,
the laser scanning velocity has an impact on the self-extraction of exploiting the same physical principles. This promises the appli-
the distinct features in the SCNN. For this reason, the AE signals may cability of our approach, although it was tested only on powder
include acoustic elements that are peculiar to the other two neigh- bed SLM. The possible transfer of the results presented to other
bored categories. The aforementioned classification error structure technologies is the subject of the future investigations.
depends on the scanning velocity, which affects the deposited light
energy, thus initiating different sintering events. A more detailed References
understanding of the underlying physics is needed to explain this
observation but this is subject of future investigation. This work [1] W.E. Frazier, Metal additive manufacturing: a review, J. Mater. Eng. Perform.
focused on a statistic analysis. 23 (2014) 1917, http://dx.doi.org/10.1007/s11665-014-0958-z.
[2] X.-B. Su, Y.-Q. Yang, P. Yu, J.-F. Sun, Development of porous medical implant
The results for conventional CNN are shown in brackets in scaffolds via laser additive manufacturing, Trans. Nonferrous Metals Soc.
Table 1. The classification accuracies observed are much lower than China 22 (Suppl. 1) (2012) 181–187, http://dx.doi.org/10.1016/S1003-
for SCNN confirming the efficiency of the non-regular convolution 6326(12)61706-3.
[3] S. Tunchel, A. Blay, R. Kolerman, E. Mijiritsky, J. AwadShibli, 3D
in this particular application. printing/additive manufacturing single titanium dental implants: a
Although the feasibility of the proposed approach has been prospective multicenter study with 3 years of follow-up, Int. J. Dent. 2016
demonstrated, it has some inherited disadvantages as compared (2016) 1–9, http://dx.doi.org/10.1155/2016/8590971, paper ID 8590971.
[4] M. Salmi, J. Tuomi, K.S. Paloheimo, R. Björkstrand, M. Paloheimo, J. Salo, R.
to other monitoring methods described in the introduction. One
Kontio, K. Mesimäki, A.A. Mäkitie, Patient-specific reconstruction with 3D
problem is the relatively weak acoustic emission signature from modeling and DMLS additive manufacturing, Rapid Prototyping J. (2012)
the additive manufacturing process which is mixed with the strong 209–214, http://dx.doi.org/10.1108/13552541211218126, Special issue:
noisy background. It is known that the high frequency of the cor- Medical advances in Additive Manufacturing.
[5] T.C. Allison, A.M. Rimpel, R. Pelton, J.J. Moore, J.C. Wilkes, K. Wygant,
responding acoustic content is rather difficult to detect in gas Manufacturing and testing experience with direct metal laser sintering for
environment, putting physical constraints on the method. How- closed centrifugal compressor impellers, in: Turbomachinery & 30th Pump
604 S.A. Shevchik et al. / Additive Manufacturing 21 (2018) 598–604

Users Symposia, Pump & Turbo 2014, Houston, TX, USA, 2014, http://dx.doi. [23] S.A. Shevchik, F. Saeidi, B. Meylan, K. Wasmer, Prediction of failure in
org/10.21423/R1F64W. lubricated surfaces using acoustic time-frequency features and random forest
[6] L.N. Carter, M.M. Attallah, R.C. Reed, Laser powder bed fabrication of algorithm, IEEE Trans. Ind. Informat. 13 (2017) 1541–1553, http://dx.doi.org/
nickel-base superalloys: influence of parameters; characterisation, 10.1109/TII.2016.2635082.
quantification and mitigation of cracking, in: Superalloys 2012, John Wiley & [24] S.A. Shevchik, B. Meylan, A. Mosaddeghi, K. Wasmer, Acoustic
Sons, Inc., 2012, pp. 577–586, http://dx.doi.org/10.1002/9781118516430. characterization of solid materials pre-weakening using electric discharge,
ch64. Submitted IEEE Trans. Ind. Informat. (2018).
[7] J. Scott, N. Gupta, C. Weber, S. Newsome, T. Wohlers, T. Caffrey, Additive [25] D. Palaz, M.M. Doss, R. Collobert, Convolutional Neural Networks-based
manufacturing: status and opportunities, in: Report from Sci. and Techn. continuous speech recognition using raw speech signal, IEEE Int. Conf.
Policy Insitute (IDA), 2012, pp. 1–35, paper at: https://cgsr.llnl.gov/content/ Acoust., Speech and Signal Proc. (ICASSP) (2015) 19–24, http://dx.doi.org/10.
assets/docs/IDA AdditiveM3D 33012 Final.pdf. 1109/ICASSP.2015.7178781.
[8] D.A. Türk, R. Kussmaul, M. Zogg, C. Klahn, A.B. Spierings, H. Könen, P. Ermanni, [26] M. Mathieu, M. Henaff, Y. Lecun, Fast Training of Convolutional Networks
M. Meboldt, Additive manufacturing with composites for integrated aircraft Through FFTs, International Conference on Learning Representations
structures, in: Society for the Advancement of Material and Process (ICLR2014), CBLS (2014) 1–9 [http://openreview.net/document/aa6ab717-
Engineering (SAMPE), Long Beach CA, USA, 2016, http://dx.doi.org/10.3929/ ca19-47e1-a958-823b9a106ca9, https://arxiv.org/abs/1312.5851].
ethz-a-010691526. [27] V. Jean-Charles, V. Gripon, G. Mercier, Generalizing the Convolution Operator
[9] C. Semini, J. Goldsmith, D. Manfrediy, F. Calignanoy, E.P. Ambrosioy, J. to Extend CNNs to Irregular Domains, 2016, pp. 1–9, CoRR abs/1606.01166,
Pakkaneny, D.G. Caldwell, Additive manufacturing for agile legged robots https://arxiv.org/abs/1606.01166.
with hydraulic actuation, Int. Conf. Adv. Rob. ICAR (2015), http://dx.doi.org/ [28] J. Bruna, W. Zaremba, A. Szlam, Y. LeCun, Spectral Networks and Locally
10.1109/ICAR.2015.7251444. Connected Networks on Graphs, 2013, pp. 1–14, arXiv preprint, paper at:
[10] W.W. Wits, S.J. Weitkamp, J. van Es, Metal additive manufacturing of a https://arxiv.org/abs/1312.6203.
high-pressure micro-pump, Procedia CIRP 7 (0) (2013) 252–257, http://dx. [29] M. Defferrard, X. Bresson, P. Vandergheynst, Convolutional neural networks
doi.org/10.1016/j.procir.2013.05.043. on graphs with fast localized spectral filtering, Adv. Neur. Inf. Proc. Syst. 29
[11] L.C. Zhang, H. Attar, Selective laser melting of titanium alloys and titanium (2016) 1–9, https://arxiv.org/abs/1606.09375.
matrix composites for biomedical applications: a review, Adv. Eng. Mater. 18 [30] M. Edwards, X. Xie, Graph Based Convolutional Neural Networks, Computer
(4) (2016) 463–475, http://dx.doi.org/10.1002/adem.201500419. Vision and Pattern Recognition, 2016, https://arxiv.org/abs/1609.08965.
[12] K.S. Everton, M. Hirsch, P. Stravroulakis, R.K. Leach, A.T. Clare, Review of [31] M. Ramakrishnan, G. Rajan, Y. Semenova, G. Farrell, Overview of fiber optic
in-situ process monitoring and in-situ metrology for metal additive sensor technologies for strain/temperature sensing applications in composite
manufacturing, Mater. Des. (2016) 431–445, http://dx.doi.org/10.1016/j. materials, Sensors 16 (1) (2016) 1–27, http://dx.doi.org/10.3390/s16010099,
matdes.2016.01.099. 99.
[13] J.J. Lewandowski, M. Seifi, Metal additive manufacturing: a review of [32] R. Kashyap, Fiber Bragg Gratting, 2nd ed., Elsevier, 2010, https://www.
mechanical properties, Annu. Rev. Mater. Res. 46 (2016) 151–186, http://dx. elsevier.com/books/fiber-bragg-gratings/kashyap/978-0-12-372579-0.
doi.org/10.1146/annurev-matsci-070115-032024. [33] S. Bland, N.T. Aboulkhair, Reducing porosity in additive manufacturing, Met.
[14] S.F. Shirazi, S. Gharehkhani, M. Mehrali, H. Yarmand, H.S.C. Metselaar, N.A. Powder Rep. 70 (2) (2014) 77–86, http://dx.doi.org/10.1016/j.mprp.2015.01.
Kadri, N.A.A. Osman, A review on powder-based additive manufacturing for 002.
tissue engineering: selective laser sintering and inkjet 3D printing, Sci. [34] J. Schindelin, I. Arganda-Carreras, E. Frise, V. Kaynig, M. Longair, T. Pietzsch, S.
Technol. Adv. Mater. 16 (3) (2015) 1–20, http://dx.doi.org/10.1088/1468- Preibisch, C. Rueden, S. Saalfeld, B. Schmid, J.-Y. Tinevez, D. James White, V.
o
6996/16/3/033502, paper N 033502. Hartenstein, K. Eliceiri, P. Tomancak, A. Cardona, Fiji: an open-source platform
[15] A. Thompson, I. Maskery, R.K. Leach, X-ray computed tomography for additive for biological-image analysis, Nat. Methods 9 (7) (2012) 676–682, http://dx.
manufacturing: a review, Meas. Sci. Technol. 27 (7) (2016) 17, http://dx.doi. doi.org/10.1038/nmeth.2019.
org/10.1088/0957-0233/27/7/072001. [35] M.V. Tazebay, A.N. Akansu, Adaptive subband transforms in time-frequency
[16] W. Devesse, D. de Baere, M. Hinderdael, P. Guillaume, Hardware in the loop excisers for DSSS communications systems, IEEE Trans. Signal Process. 43 (11)
control for additive manufacturing processes using temperature feedback, J. (1995) 2776–2782, http://dx.doi.org/10.1109/78.482125.
o
Laser Appl. 28 (2016), http://dx.doi.org/10.2351/1.4943911, paper N 022302. [36] D.I. Shuman, S.K. Narang, P. Frossard, A. Ortega, P. Vandergheynst, The
[17] M.H. Farshidianfar, A. Khajepour, A. Gerlich, Real-time control of emerging field of signal processing on graphs: extending high-dimensional
microstructure in laser additive manufacturing, Int. J. Adv. Manuf. Technol. 82 data analysis to networks and other irregular domains, IEEE Signal Process.
(5) (2016) 1173–1186, http://dx.doi.org/10.1007/s00170-015-7423-5. Mag. 30 (3) (2013) 83–98, http://dx.doi.org/10.1109/MSP.2012.2235192.
[18] R. Vaidya, S. Anand, Image processing assisted tools for pre- and [37] O. Rippel, J. Snoek, R.P. Adams, Spectral representations for convolutional
o
post-processing operations in additive manufacturing, Proc. Manuf. 5 (2016) neural networks, Mach. Learn. (stat.ML) (2015) 1–10, Paper N 1506.03767v1,
958–973, http://dx.doi.org/10.1016/j.promfg.2016.08.084. paper at: https://arxiv.org/abs/1506.03767.
[19] F. Alhwarin, A. Ferrein, A. Gebhardt, S. Kallweit, I. Scholl, O. Tedjasukmana, [38] C. Chevalier, I. Safro, Comparison of coarsening schemes for multilevel graph
Improving additive manufacturing by image processing and robotic milling, partitioning, in: Int. Conf. Learning and Intelligent Optimization (LION 2009),
IEEE International Conference on Automation Science and Engineering (CASE) Springer-Verlag, 2009, pp. 191–205, http://dx.doi.org/10.1007/978-3-642-
Aug 24–28 (2015) 924–929, http://dx.doi.org/10.1109/CoASE.2015.7294217. 11169-3 14.
[20] H. Wu, Z. Yu, Y. Wang, A new approach for online monitoring of additive [39] K.L. Priddy, P.E. Keller, Neural Networks, An Introduction, SPIE, TT68, 2005,
manufacturing based on acoustic emission, ASME 2016 11th International http://dx.doi.org/10.1117/3.633187.
Manufacturing Science and Engineering Conference, paper No. [40] H. Abdi, J.L. Williams, Principal component analysis, Wiley Interdiscip. Rev.
MSEC2016-8551 (2016) 8, http://dx.doi.org/10.1115/MSEC2016-8551, pp. Comput. Stat. 2 (4) (2010) 433–459, http://dx.doi.org/10.1002/wics.101.
V003T08A013.
[21] M. Takuma, S. Hisada, K. Saitoh, Acoustic emission measurement by fiber
bragg grating glued to cylindrical sensor holder, Adv. Mater. Sci. Eng. 2014
(2014) 12, http://dx.doi.org/10.1155/2014/274071, Article ID 274071.
[22] F. Saeidi, S.A. Shevchik, K. Wasmer, Automatic detection of scuffing using
acoustic emission, Tribol. Int. (2016) 112–117, http://dx.doi.org/10.1016/j.
triboint.2015.08.021.