Page 317 - IJB-8-4

Bonatti, et al.
2. ML techniques in additive manufacturing the DL model, reaching a high classification accuracy
(around 98%) . In another work, Tonnaer et al. tackled
[31]
ML refers to a group of AI techniques that enable the problem of detecting anomalies on the surface of
automatic learning from data to make decisions or FDM-printed parts using a semi-supervised approach.
predictions without being explicitly programmed to do The authors employed a variational autoencoder, a
so. Traditional ML algorithms, such as SVM, random type of deep architecture which, when trained with
forest, and logistic regression, are limited by the high images from non-faulty parts, can learn their probability
dimensionality of input data (e.g., images), the handling
of series (e.g., time series and videos), and the large distribution. When fed with images from error containing
amount of data to process [18-20] . parts, the model can then assign a value to those images
representing the probability that they come from the same
To overcome these limitations, researchers have
moved toward DL and deep neural networks (DNNs), distribution of non-faulty images. As a result, by setting
a proper threshold, the model can effectively distinguish
which have demonstrated to scale much better with the [32]
increase of data size . In general, DL algorithms can be between good and erroneous prints . The work by
[21]
classified into three main groups depending on the type Zhang et al. highlighted how multiple information
of data available. In supervised learning, the network is coming from different sensors can be integrated together
used as a binary or multiclass classifier using labeled data for process monitoring. In particular, the authors
instances. Semi-supervised models use a small amount of proposed a DL model to understand the tensile properties
labeled data together with a larger amount of unlabeled of FDM-printed parts. They added a set of sensors to a
data, while in unsupervised learning, no labeled data commercial printer, including infrared sensors to measure
are available . At its core, DNNs use a complex the temperature of the printed layer and accelerometers to
[22]
composition of linear and non-linear functions to learn quantify vibrations. The layer-wise output of these sensors
an expressive representation of the data. The term “deep” was fed to a long short-term memory (LSTM) network,
refers to stacking multiple layers (i.e., a set of neurons) whose output, together with other relevant information
to obtain more complex function approximators . The (e.g., extruder temperature and printing speed) was used
[23]
[33]
type of layer to be used in a neural network depends on to predict the tensile strength of the printed part .
the type of data and processing of interest. For example, In the field of laser-based technologies, Zhang et al.
the basic architecture of a CNN, designed to operate proposed a method for in-process monitoring of direct
on data in array format (e.g., a stack of three 2D arrays laser deposition. Briefly, a high-resolution camera was
corresponding to the pixel values of a color image), is used to record the formation of a single line of material.
given by the repetition of convolutional and pooling After stitching together multiple successive images, the
layers. Each convolutional layer implements multiple collected data were used to train and evaluate a CNN
filters, whose weights are optimized during training to model to both detect the presence of porosity (with a 91%
extract relevant features from their inputs (e.g., edges in accuracy) and predict the overall volume porosity (with
[34]
the input image for the top layers of the CNN). On the a root mean squared error of 1.3%) . Finally, Wang
other hand, pooling layers are used to downsample the et al. implemented a closed-loop control to stabilize
output of the convolutional layers and so summarize the the printing process in inkjet printing. In particular, the
features learned from them . authors used a vision-based system to monitor the droplet
[23]
The properties of DNNs make them good candidates formation process. Feature extraction was performed on
for multiple applications in the field of AM, including: (i) the images by: (i) cropping around a region of interest
designing new composite materials and topologies (e.g., (ROI), (ii) binarizing the image, and (iii) extracting
given a target Young’s modulus, optimize the structure connected components information, such as area and
to achieve it considering the constraints of AM), (ii) position. These features were then fed to a neural network
optimizing the process parameters, and (iii) monitoring to detect abnormal droplets and compensate them using a
the printing process to detect defects (e.g., cracks, feedback loop with proportional integral derivative (PID)
[35]
delamination, and porosity) [24,25] . Regarding these last system controlling the operational voltage .
two points, several works have focused on applying DL
to fused deposition modeling (FDM) [26,27] , inkjet , and 3. Materials and methods
[28]
powder sintering/melting processes [29,30] . 3.1. Material preparation
For example, Jin et al. used a CNN to implement a
real-time feedback loop on solid (no infill) FDM-printed Poloxamer 407 (commercial name Pluronic F-127) was
layer. The authors identified three main classes (i.e., used for all printing experiments. A solution of Pluronic
under-extrusion, over-extrusion, and good quality) and acid F-127 (Sigma-Aldrich) at 25% w/v was prepared by
built a dataset of images using a top view of the printing gradually dissolving the Pluronic powder in deionized water
process. The collected data were used to test and validate at 90°C through magnetic stirring. After complete dissolution

International Journal of Bioprinting (2022)–Volume 8, Issue 4 309