AM is increasingly being used to develop new products in a variety of industries such as aerospace, biomedical implants, and automotive. Among all the AM technologies, Laser Powder Bed Fusion (L-PBF) and specifically Selective Laser Melting (SLM) has been regarded as the most promising process for fabricating metal components (Froes et al.,
2019). In SLM a thin layer, corresponding to a slice of a 3D CAD model (Kumar,
2020), is spread over the working platform (or a substrate) using a blade. The laser scans the powder bed according to the shape defined in the CAD file. After each layer has been scanned, the powder bed is moved down by one layer thickness, followed by an automatic leveling mechanism that dispenses a new layer of powder. The laser then melts a new cross-section. The process is repeated to form the desired solid metal part (Kruth et al.,
2015). Generally, Nd: YAG-fiber laser is used for melting the powder. The working area is enclosed and either filled with an inert gas for protecting molten metal from reacting with the air (Kumar et al.,
2019).
However, critical events can occur during the layerwise process which can affect the fabrication and consequently lead to defects such as internal porosity (Sanaei et al.,
2019; Brennan et al. (
2021); Hamidi Nasab et al.,
2020), cracking (Carter et al.,
2014; Wang et al.,
2018), formation of the material balling on the part surface (Galy et al.,
2018; Hamidi Nasab et al.,
2020; McCann et al.,
2021; Sanaei & Fatemi,
2021) and high residual stresses (Aboulkhair et al.,
2019). These defects introduce common quality issues such as layer misalignment, dimensional errors, and distortions (Aboulkhair et al.,
2019; Lee et al.,
2021; Martin et al.,
2019; Sanaei & Fatemi,
2021; Seifi et al.,
2017). The reproducibility, the precision, and the mechanical properties (Gong,
2013; Jaber et al.,
2020; Zhao et al.,
2018) of the finished product can be compromised due to the abovementioned defects (Dowling et al.,
2020; Smith et al.,
2016; Yadroitsev et al.,
2021). Reliable and robust monitoring tools are necessary for quick detecting defects and reducing the time and costs associated to post-process quality inspections.
Monitoring methods
Despite the technological advancements in the last 20 years, the SLM suffers from poor repeatability. The development of sensors led to a significant increase of data that an operator is not capable of manually screen. The analysis of data coming from SLM fabrication was significantly improved in the recent years. This is demonstrated by several papers published in the area of in situ monitoring and control of AM processes (Everton et al.,
2016).
Different methods are used to evaluate the physical characteristics of the fabricated parts. It is necessary to gain a comprehensive representation of errors distribution and their features in an AM component, to investigate the trends that these distributions follow. These data are used to improve process optimization techniques, post-processing treatments and performance prediction (Sanaei et al.,
2019). It is critical to discover flaws as early in the manufacturing process as feasible to improve product quality and reduce the risk of failure caused by defects. In theory, this could enable corrective actions during the process to reduce part failure and to minimize additional post processing operations necessary to refine the fabricated components (Koester et al.,
2016). To obtain optimal parts, in-situ approaches are necessary for understanding the causes of flaws, recognizing defects, and their spatial distribution within the components (Perram et al.,
2017). This form of monitoring is an early step toward closed-loop control of the process, in which in-situ data is used to modify processing parameters to avoid or rectify problems as the following layers are processed (Croset et al.,
2021). At present, most of the works are related to off-line monitoring performed in either a destructive or nondestructive manner (du Plessis et al.,
2018; Ziółkowski et al.,
2014). Non-destructive characterization of SLM fabricated parts, using Scanning Electron Microscopy (SEM), X-ray Computed Tomography (XCT), ultrasonic, electromagnetic, eddy current, and thermography are studied in many works (Seifi et al.,
2016,
2017; Maire & Withers,
2013; Croset et al.,
2021; du Plessis et al.,
2018,
2020; Yadroitsev et al.,
2021; Repossini et al.,
2017; Grasso & Colosimo,
2017; Taheri et al.,
2017; Sharratt,
2015; Lu & Wong,
2017). Quality monitoring has frequently been carried out alongside conventional testing methods enabling detection of abnormalities in advance and aids in the rapid decision-making process for quality concerns (du Plessis et al.,
2020).
Generally, the amount of data necessary for understanding the repeatability of the SLM process avoids traditional manual analysis and modeling. Artificial intelligence is a solution to overcome the challenge in handling data and it is nowadays used to identify pattern and irregularities with limited process knowledge (Razvi et al.,
2019). Cross-sectioning coupled with SEM is a common method for SLM monitoring. Rahman et al., (
2022) proposed a deep learning-based filler detection system using Mask Region-based Convolutional Neural Network (CNN) architecture to extract the filler morphology (size distribution, orientation distribution, and spatial homogeneity) from SEM images. Another study (Rahman et al.,
2021) proposed five distinct approaches for automatically extracting straight fibers from SEM pictures to address major problems, morphological fiber extraction and overlapping or cross-linking issues. SLM surface roughness was evaluated in (Akhil et al.,
2020) by deriving image texture parameters from surface SEM pictures using first-order and second-order statistical techniques; prediction models were developed using various Machine Learning (ML) algorithms. The SEM approaches are neither in-situ nor real-time. Furthermore, SEM analysis necessitates the preparation of metallographic samples, which is an inherently destructive technology, limiting it to an offline process study tool (Collins et al.,
2017). Traditional methods of cross-section analysis or bulk density may provide quantitative assessment of the geographic distribution and shape of AM inherent flaws problematic. Because AM components are somewhat expensive, nondestructive methods of defect assessment like the Archimedes method, gas pycnometry, thermal imaging and X-ray micro-computer tomography are quite appealing (Sanaei et al.,
2019). Baumgartl et al. (
2020) used thermographic imaging and thermal mapping in a deep learning model for monitoring powder bed anomalies. Mohr et al. (
2020) used thermography and optical tomography for defect detection in comparison with Computed Tomography (CT). Guerra et al. (
2022) used High Resolution-Optical Tomography (HR-OT) for detecting geometric distortions specifically on the overhang as critical area of defect formation. All these approaches showed promising results in defect detection. As mentioned in (du Plessis et al.,
2018), one of the drawbacks of XCT is the limited resolution for large objects. Depending on parameters used, X-ray penetration problems can result in image quality issues and hence decreased data quality (du Plessis et al.,
2018). In general, reliable defects detection by these techniques are determined by the size, geometry, location, and morphology of the defect, as well as the complexity, density, and surface finish of the part (Yadroitsev et al.,
2021). In addition, these techniques are expensive. On the other hand, during the fabrication phase, process parameters such as shielding gas flow and laser power might change, affecting the melting process. Variations in these parameters can result in a lack of fusion-based porosity, even if they occur in just a few layers of the part. Depending on the size of the item, it might be difficult to detect this type of process failure using typical inspection techniques (Froes et al.,
2019).
Consequently, several in-situ monitoring methods have been developed to examine specific process parameters and items like as melt pool and spatter behavior (Repossini et al.,
2017; Yakout et al.,
2021; Ye et al.,
2018; Zhang et al.,
2018), part distortion (Caltanissetta et al.,
2018; Li et al.,
2018), dimensional accuracy (Aminzadeh & Kurfess,
2015; Land et al.,
2015), powder recoating and powder bed surface (Aminzadeh & Kurfess,
2015; Craeghs et al.,
2011; Krauss et al.,
2014). These techniques have been reviewed in different studies (Everton et al.,
2016; Grasso & Colosimo,
2017; McCann et al.,
2021). Among them, in situ layerwise imaging techniques have been widely investigated in order to permit image-based layerwise anomaly identification for powder bed AM techniques. Scanned layer is usually monitored by integrating a Digital Single-Lens Reflex (DSLR) camera with SLM process to achieve the highest possible image quality (Nakamura,
2017). An external light module is common for improving the information obtained from the process (Gobert et al.,
2018). Authors in some studies used different lighting conditions and contrast to monitor internal defects (Abdelrahman et al.,
2017) or geometrical deviation (Foster et al.,
2020), because different lighting preferences are regarded as an important component of the imaging system in order to ease automated flaw detection. The positioning of the camera can be of two types: coaxial setup where the camera is connected to a dichroic mirror to collect images along the laser path; off-axis setup, where the camera is positioned outside the system window (Repossini et al.,
2017). Most of the literature works relate to this second type. The employment of ML is gaining more and more interest because it allows process predictions on the output without the need for explicit programming. In the SLM monitoring via Digital Image Processing (DIP) the supervised approach is the more common for layerwise monitoring (Imani et al.,
2018). This method is carried out by providing two sets of data, namely the training and testing ones. In the training phase the first set of data is labeled in order to provide an input–output pairing which allows ML algorithm to be trained and establish a set of metrics to predict values on new input data. The testing data are used in the validation phase aiming to determine the model accuracy (Wang et al.,
2020). Images are preprocessed through background removal, filtering and cropping to solve problems such as redundancy and noise (Scime & Beuth,
2019). The operations are undertaken by using a solid knowledge, skills and abilities necessary to identify images features and process signatures (Qi et al.,
2019). The most common metrics for the performance evaluation is the precision defined as the true to predicted instances ratio. It is worth to note that the supervised ML are well suited for huge amount of data, but it requires a big amount of in situ experiments which must be repeated as the processing conditions changes. The selected camera specifications are characterized by high quality and high resolution: in (Gobert et al.,
2018; Imani et al.,
2019; Snow et al.,
2021) the camera resolution was 36.3 Mpixels and in (Gaikwad et al.,
2019) was 24.2 Mpixels; moreover, the optical system is typically focused on the part under-fabrication area to enhance the spatial resolution. The use of the ML applied to these designed for the purpose systems allowed an accuracy ranging between 85% and 99%. Owing high resolution systems for many layers requires big data managing thus point clouds became widely used for computer vision and in AM monitoring. Liu et al. (
2021) extracted geometrical features from 3D printed part images to compute Sa roughness via point clouds. In Ye et al., (
2020) the relationship between the melt pool images acquired via an off-axial photodiode and the tensile test results was developed. Lin et al. (
2019) mapped AM images structuring the point clouds onto grids and comparing the defects with the CAD file. The parts quality produced by directed energy deposition were related to the structured-light scanning image by (Zhang et al.,
2020).
In the presented literature review the SLM powder bed monitoring requires high performance systems to detect detailed anomalies and reliably relate them to the part defects. However, warranty issues, manufacturer restrictions or local laws prevent the industry from modifying the original machine to develop a designed for the purpose acquisition setup. This work covers the identified industrial need to implement the DIP for layerwise monitoring without modifying the system and/or hampering the production activities at machine shop floor level. Typically, commercial SLM machines are equipped with an internal camera to manually check possible issues in the powders bed spreading and subsequential laser scanning. Unfortunately, they are not designed for in-depth evaluation and both the resolution, and the optical precision prevent specific investigations since they are addressed to data collection rather than data analysis (Grasso & Colosimo,
2017). Hence, the development of an automatic system for possible anomalies detection via machine built-in camera is a challenging activity. This work focuses on this feasibility and aims to provide a direct and economical way to quickly detect flaws in the final part associated with the powder bed spreading, thus saving time and costs associated with the post-process quality inspection. Furthermore, the proposed methodology is unsupervised, i.e. no preliminary expensive experimentation is necessary for the training stage. The goal is to recognize through DIP various defects including lack-of-fusion porosity, uneven top surfaces, and geometrical deformations. The build-in camera cannot be modified by means of position, field-of-view, resolution and lighting. For the purpose, particular care is paid to the camera calibration in order to maximize the analysis capabilities. Furthermore, quantitative 3D reconstruction of powder bed anomalies is provided and compared to the traditional expensive CT measurement.