1. Introduction
Due to advances in the Additive Manufacturing (AM) process, an increase in new manufacturable products and designs have entered the market at a rapid rate. AM has revolutionized manufacturing because of its ability to streamline production parts from computerized models to market, reduce waste, utilize a variety of materials for production, and print harsh complexity demands. The layer-by-layer process of producing 3D objects through computer-aided design (CAD) was developed in the 1980s and called ‘rapid prototyping’ [
1]. It was originally intended as a means to create a specimen quickly as a prototype or basis model upon which further models would eventually develop towards a final product [
2].
AM is a method currently best suited as a rapid prototyping technique due to engulfing issues of quality consistency, thermal stresses, and internal defects and distortions which cannot be seen from part surfaces [
3]. Growth and demand in the AM area suggest a need to transverse AM into a mass production technique while maintaining its versatile production capabilities. Compared to traditional manufacturing processes (subtractive manufacturing, injection molding, etc.), the AM research area is still relatively new and in need of in-depth investigation. The AM process possesses many advantages over traditional forms of manufacturing and can be a viable solution when conventional methods of manufacturing are not approachable. One major advantage AM has over traditional manufacturing methods is the use of designing and producing highly complex parts. As the complexity of parts increases, the price of manufacturing those parts with conventional methods increases in a similar fashion [
4] however, this is not the case with AM. The cost to produce highly complex parts is drastically reduced when compared to conventional methods, such as AM produced jet engine manifolds [
5]. In addition to reduced cost, AM processes have the potential to become fully automated. Research has demonstrated a strong desire to develop fully automated metal AM techniques with rapid modeling capabilities to reduce the potential of defects and ensure quality standards of parts is achieved [
6,
7]. In [
8], Panchagnula and Simhambhatla highlight the ability of the metal AM process to produce complex designs including sudden overhang features (features that are perpendicular to the deposition direction or nearly horizontal features) without support material. The technique consists of re-orienting the workpiece and/or the deposition head upon every instance using higher order kinematics. The sudden overhang is identified from a CAD model and an orthogonal tool path is generated. While the variety of manufacturable products and designs has increased, Cyber Manufacturing (CM) systems have become a necessity for the success of Industry 4.0. The fourth industrial revolution merges both physical and digital technological concepts including analytics, robotics, additive manufacturing, artificial intelligence, advanced materials, natural language processing, high-performance computing, cognitive technologies, and augmented reality [
9,
10]. As a result, AM has developed into a manufacturing technology fueling Industry 4.0 and in [
11] a comparison of several popular processes is reviewed in detail. This interest has sparked a tremendous amount of research efforts to improve and develop new methods for monitoring and quality control of workpieces during AM building [
12,
13,
14,
15,
16,
17,
18].
Manufacturing methods for AM vary in advantages and disadvantages which we direct the reader to [
11] for a thorough analysis of the techniques. Directed-energy deposition (DED) is a technique where thermal energy is used to melt material during deposition [
11]. During the DED operation, metal powder (or wire) is injected from the feedstock which is melted via a laser or electron beam and deposited on top of a substrate [
19]. This particular method produces high-quality parts and possesses a high degree of controllable grain structure. The traditional form of the DED process is used with a wire feedstock drawing inspiration from the traditional welding process. The other technique that utilizes powder flow from the feedstock is called Laser Engineered Net Shaping (LENS) [
20]. In our study, we utilize Laser-based Directed Energy Deposition (LB-DED) with a powder flow method for the production of metallic workpieces.
Performance of parts is highly contingent upon which AM process is selected for production. The AM process is critical for desired mechanical properties in layer-produced parts, as seen in the study by [
21], tensile and impact strengths are strongly dependent on the layer building direction. Recent studies are focused more on nondestructive internal dynamic defect monitoring (or quality monitoring) of metallic parts [
19,
22]. One major issue with part defects is the presence of undesired cracks and low-density occurrences such as porosity. Researchers demonstrate a heavy interest in methods to identify and classify these malfunctions to avoid part defects [
23,
24,
25,
26,
27]. Acceptable low-density occurrences such as porosity numbers and sizes vary based on the application. Most applications aim to limit the size and number of low-density occurrences in parts to produce fully dense specimens. Other literature shows that low to no density occurrences and cracks that are accumulated throughout the build process negatively hinders parts and causes defects [
28,
29]. With this theme in mind, our focus was to create a process in which we can monitor and possibly control low-density occurrences such as porosity sizes and numbers.
In-situ monitoring for workpiece quality is a popular topic in the AM community. There have been studies that aim to measure printer parameters in real-time to quantify changes in the printing conditions [
30]. Neural Networks (NN) are being used for a variety of in situ applications. For example, data-driven Artificial Neural Networks (ANN) have been used in an attempt to predict characteristics of porosities in a part based on features during the build [
31]. The study conducted in [
31] demonstrates a framework of how NN’s can be used for the evaluation of low-density occurrences from real-time datasets. Faulty parameters lead to the possibility of pronounced porosities, cracks/corrosion, and internal residual stresses in the workpiece [
32,
33]. Research demonstrates a great desire to monitor porosity in real-time because a specimen with pronounced low densities (porosity) will often exhibit poor mechanical performance [
34,
35]. As a result, research has trended towards using acoustic emission for identifying cracks/corrosion in real-time [
33,
36,
37]. Since acoustic emission data is rich, these investigations use Deep Learning (DL) techniques for prediction.
There is a scarcity of literature focused on classifying layer quality with a real-time evaluation model based on layer features. In addition, we believe understanding the issue as a time-series component could lead to an acceptable method of determining how or if a poorly produced layer in the incipient stage of a print has drastic effects later on in a print causing multiple poor layers resulting in workpiece defects. Where [
21] focused on mechanical properties to provide insight into manufacturing principles of the AM process, this study is focused on creating a benchmark and providing insight into the features that most affect internal quality at the layer level of the LB-DED process. Based on an adjustable threshold, we have developed a method that classifies layers as acceptable or unacceptable based on low-density occurrences such as porosity size and quantity. We seek to examine and illuminate this complex problem and provide insight into classifying layers based on only print features during the build process.
The novelty of this investigation is two-fold. First, we utilize a synchronously collected in situ dataset with a low sampling rate in conjunction with an ex-situ evaluation based on CT scans of parts to create a labeling dataset. Second, a DL algorithm classifies inter-layer quality through the build process of a part based on in situ input data and ex-situ labels. In this study, a CNN is trained offline on a sample deemed ‘high-quality’ (sample 2) and tested on the same sample. Next, using the mappings learned from the process parameters of building a high-quality part, the model tests on a sample deemed ‘low-quality’ (sample 1). This process is repeated for sample 1. This method also consists of (a) collecting a dataset of structural specimens with defects (unacceptable discontinuities), (b) extracting features for use in the algorithmic approach, (c) developing an advanced DL technique highly sensitive to process parameters in the build process to identify and classify defects during layer construction of specimens, and (d) testing layer classification capabilities of the advanced DL model on sequential data from two samples. This investigation demonstrates promising results for the possibility of implementing a data-driven classification CNN model for the use of DL during the LB-DED process to ensure defect-free parts are produced.
The following literature presented primarily focuses on DL techniques for classification during AM prints based on avoiding porosity and/or cracks/corrosion in a real-time data sphere. Therefore, several subsections divide the following topics: (1) layer-wise feature monitoring, (2) classification of porosity methods, and (3) DL techniques for quality monitoring of AM workpieces.
1.1. Layer Level Feature Classification
Literature has investigated layer-by-layer evaluation for material extrusion methods [
38,
39] and DED applications [
40]. In [
38], Wang et al. focus on residual strain issues that occur during the extrusion process. In the investigation, researchers measure the solidification-induced residual strain distribution via a novel optical backscatter reflectometry (OBR) based fiber-optic sensing system was embedded to measure distributed residual strains. Jin, et al. create a self-monitoring system with real-time camera images to evaluate interlayer imperfections during the build process for a fused deposition modeling (FDM) print [
39]. Researchers trained a CNN that was capable of predicting delamination and warping issues in real-time with an accuracy of 97.5%.
Other studies have focused on layer evaluation techniques involving image-based detection systems. Gobert et al. investigated layer-wise image-based evaluation for defect detection during metallic powder bed fusion with supervised ML [
41]. The technique included using a high-resolution digital single-lens reflex camera to capture images at each layer during a build process. Using a Support Vector Machine (SVM), features are extracted from the imagery dataset and classified as either a ‘flaw’ or ‘normal’ class. These two classes are used to represent either an “undesirable interruption in the typical structure of the material or a nominal build condition” (respectively) [
41]. The authors utilized ex-situ CT scans to identify incomplete fusion, porosity, cracks, or inclusions and successfully built a 3D high-resolution dataset of defects. The SVM was capable of detecting defects in situ with greater than 80% accuracy. Another study employed a similar tactic of utilizing high-resolution optical imagery to detect defects in situ during a metal AM build process [
42]. In this investigation Abdelrahman, et al. focuses on Lack-of-fusion defects from parts built using the PBF approach. In this approach, the algorithm is designed to create a 3D volumetric image of defects that is then compared to a computer-aided design (CAD) model to determine sensitivity and specificity metrics. In total, 28 intentional defects were placed throughout the part varying in size and shape. Sensitivity was calculated as 0.915 and specificity calculated as 0.840. In [
43], Davtalab et al. developed a CNN that was capable of detecting deformations in large-scale AM concrete construction. The technique uses a segmentation process by applying a binary mask to RGB input images to extract concrete layers. Defects are detected with 97.5% accuracy.
1.2. Porosity Monitoring
Most research agrees that porosity can be divided into three categories which are (1) keyhole, (2) gas pores, and (3) lack of fusion [
31,
44]. A keyhole porosity occurs when there is an excess in the energy density in the melt pool, typically as a result of high LP and low scanning velocity, as a consequence the melt pool becomes deep with a depression zone of vaporized material [
45]. Gas porosities are created as a result of the entrapment of vapors within the melt pool [
46]. If a substrate or preceding layer experiences insufficient penetration from the melt pool of a succeeding layer, Lack of fusion may be produced [
47,
48].
1.2.1. Post-Processing Evaluation
In this section, we define X-ray computed tomography (XCT) and Ultrasonic testing methods of usage pertaining to AM. XCT is defined as a method of establishing 3D representations of objects through obtaining X-ray images centered around an axis of rotation and developing 3D models through mathematical image reconstruction algorithms [
49,
50,
51]. Between 2010–2014, XCT became a well-established method for measuring porosity by comparing empty voxels [
52]. It has since developed into a common nondestructive testing method [
53,
54,
55,
56]. In [
56], an ML classifying technique is used for evaluating porosity segmentation and a benchmark is established by examining disparities in porosity segmentation within XCT scans.
Ultrasonic testing (UT) has been used for assessing damage in metal parts and manufacturing processes [
57,
58]. Specifically, DL is used in [
59] for ultrasonic testing for porosity evaluation of additively manufactured components. During UT application, ultrasonic waves are emitted into a workpiece [
58]. The ultrasonic waves interact with any flaws in the workpiece and detect and evaluate defects to ensure the integrity of the workpiece [
58].
A unique image-based post-processing technique was used in [
60], where authors collected a large database of image-based defects. These images were classified into ’good quality’, ’crack’, ’lack of fusion’, and ’gas porosity’ to create a set of labels that were then input to a CNN. Cui, et al. reported that the CNN was able to classify the types of defects with upwards of 92.1% accuracy with 8.01 milliseconds of computer time recognition.
1.2.2. In-Situ Based Methods
Research has also focused on image-based methods for in situ monitoring during metal AM applications [
61]. Zhao et al. created a real-time approach to defect detection through the use of high-speed synchrotron hard X-ray imaging [
62]. During the study, insight is provided into the formation of defects by tracking the motion of powder particles and observing melt pool as a function of laser heating in an attempt to observe the structure underneath the surface. Authors demonstrate that the approach indeed is capable of observing the melt pool development in real-time. The approach consists of using a pseudo pink beam that impinges upon the powder bed directly above the sample while the X-ray penetrates the sample from “the side”. Then, imaging and diffraction detectors collect the distance of each component and is labeled. One particularly insightful finding is the author’s observed the processes of the molten metal spreading during melt pool creation due to continuous laser heating. In doing so, the cavity depth and strong oscillation behavior is observed that occurs under the surface in real-time [
62]. In [
33], research is conducted through the use of acoustic emission (AE) for in situ quality monitoring of pores in the building of a workpiece. Unlike the XCT method for detecting and characterizing pores, AE is a technique performed as the build is progressing. AE is an advantageous method due to the acoustic signatures which possess potential robustness and richness of data pertaining to the material properties, process conditions, outliers in the data, and defects [
36]. Research has been conducted in an attempt to improve the strength of AE during monitoring applications [
63]. In the study, Zhensheng, et al., aim to improve upon weak AE amplitude, especially when the signals are diluted due to environmental noise. In [
64], a nondestructive technique is used for acoustic measurement during quality monitoring of machine and material state AM process.
Many of these methods are centered around anomaly detection frameworks based on a preset threshold. Popular-visual based methods for measuring porosities are infrared thermography and optical tomography. In [
65], researchers introduce a new contactless off-axis thermal detection monitoring arrangement. In the study, two cameras were used simultaneously for thermography and optical tomography.
1.3. DL Application in Quality Control of AM
The following highlights current research in the in situ AM community with a focus on quality monitoring through the use of DL techniques based on in situ models.
1.3.1. DL Definition
Artificial Neural Networks (ANN) are a type of Machine Learning (ML) technique that is modeled and inspired by the brain [
66]. These networks mimic the learning process through a technique called back-propagation that optimizes weights in the network [
67]. DL networks are an advanced version of ANN and rely on a feedforward network to approximate some function
[
68]. DL networks get their name from having a chain of functions which are algorithmically deeper than the standard ANN with a connected final layer called the output layer [
68]. In essence, DL networks are very large versions of ANN.
DL has become popular in part due to advancements in computational power and the addition of advanced activation functions [
69,
70,
71]. One such activation function is called Rectified Linear Unit or ReLU. The ReLU function is mathematically represented as:
where,
is the input of the activation function located at
. A particular type of DL that has successfully been employed for classification experiments is Convolutional Neural Networks (CNN). CNNs are used for a variety of applications including time-series (1D grid) and imagery (2D grid) [
68]. In this investigation, we treat the data as a time-series and utilize it for a 1D CNN algorithm. Other activation functions include the hyperbolic tangent (tanh) and sigmoid functions, the sigmoid is used in the CNN model during this investigation for the dense layer. The tanh and sigmoid functions are defined in Equations (
2) and (
3) (respectively) as [
72]:
One advantage to the tanh function is that its range outputs between
and 1 for real values of the input
x [
72]. The main advantage of the tanh function compared to the sigmoid function is that the hyperbolic has a steeper derivative however, neither of these functions work well with deeper networks for convolution [
73]. The sigmoid function is commonly used in the last or ‘fully connected’ layer in classification problems because it outputs 0 or 1.
1.3.2. DL for Quality Inspection
Quality monitoring has been utilized for identifying or detecting and classifying defects of AM workpieces. In [
74], a DL approach is used for identifying small process shifts. Recent literature demonstrates a desire for quality monitoring systems in AM processes through the use of DL techniques. Many of the methods recently published are focused on real-time defect detection via acoustic emission and some have begun to introduce Convolution Neural Networks (CNN) and many variations of the network to predict part defects [
28,
33,
36,
37,
75,
76]. Although there have been investigations focused on evaluation of the printing parameters of an AM building process. In [
77], Chaudhry and Soulaïmani showcase an ML framework that analyzes the sensitivity and uncertainty in Selective Laser Melting (SLM) build process, which allows the optimization of printing parameters.
In [
33], researchers utilize acoustic emission for in situ quality monitoring that combined with machine learning. In the study, acoustic features are extracted and classified as ‘poor’, ‘medium’, or ‘high’ porosity using a spectral CNN (SCNN). Researchers report a confidence interval of 83–89% for porosity classification. Taheri, et al. investigate the use of K-means clustering algorithm for classification of different process conditions during in situ AM quality monitoring [
36]. Classification through the use of K-means clustering ranged from 70 to 90%, depending on the process condition. In [
76], Yang, et al. establish a method for detecting part defects with a single short detector network (SSD). Mi, et al. investigate quality monitoring during a laser-based DED manufacturing process utilizing a CNN [
78]. During the study, Mi et al. train a Deep CNN to observe defects through imaging and showcase how the predictive model detects defects with up to 94.71% accuracy. Li et al. present a deep learning-based quality identification method that image-based detection with low-quality (noisy and blurred images) data as inputs [
79]. In the investigation, authors display a model capable of functioning with the semi-supervised, low-quality dataset. Some inputs are labeled and some are non-labeled to mimic real-world imagery. This approach separates itself from other approaches seen in the literature by exploring supervised and unsupervised data for in situ quality monitoring.
This study investigates the layer-by-layer classification technique using a DL method that treats the real-time data as a time series dataset. Predictions are made based on a data-driven technique focused mainly on the parameters within each layer that are highly responsible for layer quality. We have developed a DL method for detecting good and bad layers based on post-process XCT porosity analysis resulting in classifying layers as good or bad. This study aims to provide insight into the effects and influences of in situ sensed or measured features and AM parameters on bad layers and how they render a part low-quality during the DED process.