Keywords

1 Introduction

The Selective Laser Melting (SLM) is one of the additive manufacturing (AM) process that allows the manufacturing of three-dimensional complex functional structures of metals or alloys using high power laser [1,2,3]. The recoater spreads a layer of metal powder uniformly across the build platform. A high power laser beam scans the individual cross-sections of the parts to be built using a galvo scanner with an f-theta lens. In the scanning regions, the metal powder melts, consolidates, and fuses the scanned layer to the previous layer already built. The whole process is repeated until the part is fabricated. The L-PBF process typically carried out in an inert atmosphere to prevent oxidation of the material and chemical reactions between the molten material and the environment.

Due to the large numbers of controllable and uncontrollable parameters, it is challenging to control the process for better repeatability and reliability. For example, Galy et al. summarized all the possible defects arising in the Al alloy due to a variation in process parameters [4]. Due to the layer-wise principle of the process, these defects are complicated to detect after the part is completed. An expensive technique such as computed tomography is performed to ensure the internal quality of the final part. Since CT is expensive and time-consuming, it is not viable to use it for every part. Therefore, the need for an on-line monitoring system is inevitable. Berumen et al. developed the in-situ monitoring system consists of a CMOS camera and a photodiode, which captures the radiation in the wavelength range of 700–940 nm [5]. Similarly, Lott et al. [6] developed a coaxial setup based on the illumination of the laser beam to study the melt pool dynamics during printing. Doubenskaia et al. [7, 8] developed a monitoring system consisting of two photodiodes with a detection range of 900–1700 nm. Nowadays, commercial systems are also equipped with monitoring systems such as Melt Pool Monitoring (MPM) from SLM Solutions, EOSTATE from EOS GmbH, InfiniAM from Renishaw.

Due to the enormous size of the in-situ data acquired and the difficulties encountered in analyzing it, processing the data to qualify the final part is a difficult task. Also, manually detecting the drift or defect is time-consuming and laborious work. Therefore, there is a need for applying Machine Learning (ML) techniques for processing the data to detect drift. Thus, in recent years many researchers have used machine learning approaches on large datasets for decision making [9, 10]. Grasso et al. [11] demonstrated the use of the T-mode Principal Component Analysis (PCA) for image data obtained by OlympusTM I-speed 3 camera mounted outside the build chamber at an angle of 400 w.r.t build plate to define a spatial statistical descriptor to detect local over-heating phenomena along the scan path by analyzing the intensity profile of each pixel. Ikenna et al. [12] used a semi-supervised machine learning approach for automatic fault detection in enormous data obtained using photodiodes. They have used the ultimate tensile strength as their descriptor to label their tensile parts as acceptable or faulty to train a semi-supervised learning algorithm. Scime et al. [13] used one-megapixel Photron FASTCAM Mini AX200 high-speed camera in the visible range to capture melt pool images. The machine learning technique, called Scale Invariant Feature Transforms (SIFT), was used to extract features from the captured melt pool signatures. The bag-of-words (or Keypoints) were used to obtain a scale-agnostic description of melt pool morphology. The melt pool morphology represents the shape of the melt pool, spatter, and vapor plume. With the ML approach, the author classified the individual melt pool morphologies into four categories, such as desirable, balling, under-melting, and key-hole porosities. Supervised classification ML technique called Support vector machine (SVM) was used to train and test the model. In another study, Scime et al. [14] used a computer vision algorithm to predict the percentage of types of defects present in part based on the powder bed images. Scime et al. [14] extracted the regions of the images and classifies such as anomaly free, recoater hopping, recoater streaking, debris, superelevation, part failure, and incomplete spreading. These categories were used as the fingerprints for the computer vision algorithm.

In this study, a co-axial melt pool monitoring (MPM) system installed on a commercial SLM 280HL machine was used to acquire the in-situ data. It shall be noted that no additional modifications were made to the hardware available. A balanced labeled dataset (equal number of “drift” and “no-drift” data points) is prepared for training a supervised Support Vector Machine (SVM) classifier to detect the drift in the parts. For testing the trained algorithm, specific geometries that incorporate the overeating and lack of fusion drift were printed by varying the volumetric energy density.

2 Definitions

  • Drift: Non-uniformity in the melt pool signatures results in the “hotspots,” which are the areas where the intensity of the signal is higher compared to the rest of the layer. These hotspots are an indication of drift and the areas of highest probability to generate real defects in part. Recently, Mohr et al. linked hotspots in the MPM layer to the final porosity in part using computer tomography [15]. Therefore, in our study, the whole layer is termed as “Drift” if there is a presence of hotspot confirmed via the MPM visualization tool.

  • No-Drift: The layer with no significant hotspots are termed as no-drift layer.

  • Labeled Data: The layer for which the labels such as “drift” and “no-drift” are known.

  • Unlabeled Data: The layer for which the labels, i.e., “drift” and “no-drift,” are not known.

  • Balanced Dataset: The dataset contains an equal number of data points for each label.

3 Machine Learning Theory

In recent years, data-driven methods have been widely used to monitor and improve the overall performance of the AM process. The challenges, such as post processability of the enormous size data and automatic drift detection associated with the in-situ monitoring systems, can be overcome by using the ML approach. Moreover, ML can monitor the process and provide feedback in real-time. ML approaches can broadly be classified into two categories, called “Supervised” and “Unsupervised” approaches. The supervised approaches work for classification and regression problems, whereas unsupervised approaches work for clustering problems. The main difference between both approaches mainly depends on the type of data required as input for training. In supervised, one has to provide a labeled dataset in which each data point belongs to a specific class as categorized by the label associated with it. In contrast, unsupervised learning is used when we do not have the labels for the dataset or, in other words, when we do not have ground-truth knowledge about the dataset. For the AM process, it is challenging and expensive to obtain ground truth dataset for training. Therefore, a small dataset is prepared for training the algorithm. In this study, a supervised learning model called “Support vector machine” (SVM), which works sufficiently good for small datasets, is used.

SVM is a classification algorithm that divides the dataset into multiple classes. In this study, the in-situ data is treated as a binary classification problem, i.e., only two classes (“drift” and “no-drift”) are identified, so the SVM classifier perfectly fits for this case. During training the SVM classifier, the decision function amounts to identifying a suitable reproducible hyperplane that maximizes the distance (also called “margin”) between the support vectors of both class labels (Fig. 1). SVM classifier can work for both linear or non-linear classification problems. Further, the linear SVM problems range in their complexity depending upon their number of features selected. For example, in two feature dimensions, the hyperplane corresponds to a line, whereas in the case of three features, the hyperplane is a two-dimensional plane. Regardless of the SVM’s complexity (i.e., dimensionality), classification problems are often linear, which means that hyperplane used is a straight line, not a curved line. If the features selected for the SVM classifier are linearly separable, then we can draw a straight hyperplane to separate the two labels of the class of interest. Usually, there are two types of margins: hard margins and soft margins. With a hard margin, we restrict the classifier to make any error while training. Although the hard margin is computationally less expensive, but not always the linear separability of the features is so easy. Therefore, by allowing the classifier to misclassify, the greater generalizability to a new data can be obtained by a larger margin. This misclassification can be obtained by the so-called “Soft margin,” which relies on the variable ξ. The values of ξ lies in the range 0 ≤ ξ ≤ 1. The non-zero values of ξ, in turn, can allow classification error that can result when outliers in the training data lead to the hyperplane making mistakes, i.e., misclassification (Fig. 2). This hard margin is a particular case of soft margin when the slack variable (ξ) is zero. A soft-margin constant denoted by C is introduced to incur a penalty on ξ. The parameter C helps to choose the trade-off between the tanning errors and complexity and reduces the chances of overfitting, which means fine-tuning the classifier for maximum accuracy in the training dataset. The decision boundary of a non-linear classifier depends on the data in a non-linear way. In these cases, the kernel method is used to transform the support vectors into a high dimensional space. A detailed explanation for SVM can be studied in the literature [16, 17].

Fig. 1.
figure 1

Schematic of the hyperplane that maximally separates the support vectors corresponding to each of the two classes [16].

Full size image
Fig. 2.
figure 2

(a) Hard margin on linearly separable dataset where no training errors are allowed. (b) the soft margin where two training errors are permitted to make data non-linearly separable [18].

Full size image

Typically, there are three stages in the SVM classifier: (a) feature selection and preparing feature dataset, (b) training the classifier and testing, and (c) checking the accuracy of the classifier. All three stages for our case are discussed in the next sections.

4 Materials and Methods

4.1 Material and Machine

The gas atomized AlSi7Mg0.6 spherical powder supplied by SLM Solutions was used to print specimens with varying process parameters. The particle size distribution was 20–63 µm as specified by the supplier. The apparent density of the powder was 1.53 g/cm3, and the chemical composition of the as-received powder is tabulated in Table 1.

Table 1. Elemental composition of as received AlSi7Mg0.6 powder (All the values are given in wt%).
Full size table

The commercial SLM 280HL (SLM Solutions Group AG, Lübeck, Germany) equipped with 700 W twin continuous wave (CW) ytterbium fiber lasers with an emitting wavelength of 1070 nm and a spot diameter of 80–115 μm was used for printing. The build envelope volume is 280 × 280 × 365 mm3, and the build chamber was maintained in the Ar gas environment with an oxygen level below 0.1%. The Al base plate was preheated to a temperature of 150 °C prior to printing to reduce the effect of residual stresses in part.

The SLM 280HL machine is equipped with an in-situ monitoring device called ‘Melt Pool Monitoring,’ which consists of two on-axis photodiodes. The specifications and working principles are discussed in Sect. 4.3.

4.2 Part Geometry and Process Parameters

To prepare a training dataset for the SVM classifier model, a balanced dataset comprising an equal number of drift and no-drift layers is necessary. To obtain the balanced dataset, an artificially drift is introduced in the samples. Therefore, the unique geometrical specimens were printed, as shown in Fig. 3. The process parameters tabulated in Table 2 were varied to obtain varied volumetric energy density in the range of 40–73 J/mm3 for each shown geometry. The overhang samples (Fig. 3a, and 3b) with an overhang of 8 mm was printed without any support structure and stripes rotation of 67° for the overhang layer. No down-skin parameters were used for the overhang layer. For lack of fusion samples (Fig. 3d), an internal cuboid type groove with dimensions 10 × 8 × 0.09 mm3 was printed. The thickness of the groove set to 0.09 mm, i.e., three times the layer thickness of 30 μ. The printed parts were cut along the build direction (z-direction), and 3 μ polished for microscopic analysis. The optical microscope supplied by Leica systems was used for analyzing the porosity.

Fig. 3.
figure 3

Sketch of the specimens (a) cubic overhang (size 10 × 10 × 10 mm), (b) cylindrical overhang (diameter: 10 mm and height 10 mm), and (d) specimen with inner groove.

Full size image
Table 2. Varied process parameters for printing.
Full size table

4.3 Melt Pool Monitoring System

The co-axial melt pool monitoring (MPM) system installed on the commercial machine SLM 280HL was used to collect thermal emissions from the melt pool formed due to laser-powder interaction. The melt pool systems are co-axial systems, which means it is in the alignment of the laser path and collects the real-time emissions from the laser path at acquisition frequency of 100 kHz. The MPM module consists of two photodiodes with different sensitive areas. The spectral range of the photodiodes cannot be revealed due to confidentiality issues. However, both the photodiodes capture the thermal emission in the near-infrared region. The schematic diagram of the MPM system is shown in Fig. 4. Only the emissions traveling perpendicular to the build platform are taken into account. The thermal radiations follow the same path as of laser and directed into the MPM module with the help of a semi-transparent mirror, which does not allow laser wavelength to pass. The signal is split into two different spectral ranges and captured by the installed photodiodes, respectively. The received signal is forwarded to associated ADCs (Analog to digital convertor) and provided in an FPGA (field-programmable gate array) by the individual photodiodes. The captured thermal emissions from photodiodes 1 and 2 are stored along with the x/y-coordinates (16-bit). The values are stored in parallel with the laser on/off signal from FPGA to PC in every 10 µs. All the data is stored for every layer in a data file, which can be accessed as 2D representation in MPM software provided by the SLM Solutions [19]. The new file is created automatically for each layer after the complete exposure. For this work, no additional modifications are made to the installed hardware. Also, only the data from ADC1 (Photodiode 1) are further post-processed. The reason to select the ADC1 is due to its specific detection range.

Fig. 4.
figure 4

Schematic of melt pool monitoring system installed on SLM 280HL [19].

Full size image

Volume Energy Density Sensitivity Analysis

The sensitivity of the MPM system for the volume energy density was studied. The cylindrical specimens with the sandwich structure were printed, where the bottom and upper part of the cylinder was printed with the optimized process parameters. In contrast, the middle part was subjected to varied volumetric energy density. The geometry of the samples is shown in the right corner of the graph depicted in Fig. 5. It shows the mean thermal counts for every layer, and it can be observed the photodiode is sensitive to change in volume energy density in the samples, which ranges from 40–73 J/mm3. The first 50 layers that account for a higher signal represents the support structure of the part. Therefore, it is necessary to remove support structure data from the data preparation before processing. The layers from 300 to 580 indicate the variation in the input volumetric energy region. Sample 3 has the lowest energy density of 40.48 J/mm3, which contributes lower mean thermal emissions, whereas sample 7 has the highest volume energy density of 73.26 J/mm3, which shows higher mean thermal emissions. So it can be concluded that the MPM system signal is proportional to input volumetric energy density.

Fig. 5.
figure 5

The mean thermal emissions recorded by MPM systems for the induced drift detection in the specimens shown in right corner of graphs (dark red color: drift area, red area: optimized processed parameters).

Full size image

Down-Skin Sensitivity Analysis

To study the sensitivity of the MPM system for down-skin, overhang samples was printed, as shown in Fig. 3a and Fig. 3b. It was observed that the photodiode signal shows a gradual increase in the thermal emissions as the laser exposure moves from the printed part to the overhang part (Fig. 6). This is in conjunction with the theory that the powder has lower thermal conductivity compared to the consolidated part. The presence of voids in loose powder makes it an inferior heat conductor compared to the printed part. Thus, the consolidated part acts as a major heat sink in the SLM process. Therefore when the region above the powder exposed, the melt pool thermal emissions are higher compared to the melt pool region on the printed part. The same phenomenon could also occur during a lack of fusion defects. As sometimes, due to bad powder spreading, there is non-uniformity in the powder bed, and some regions are not covered with the powder uniformly for a few layers.

Fig. 6.
figure 6

Scan vector wise MPM signal for downskin layer for (a) cylindrical sample, (b) cuboid sample.

Full size image

Nevertheless, in the next passes of recoating, the regions with lack of powder are uniformly covered, but now the laser prints thicker layer compared to other regions with uniform powder spreading. This phenomenon can lead to a lack of fusion defects as laser energy is not enough to fully melt the powder layer. So it can be concluded that the MPM systems are also sensitive to lack of fusion defect.

4.4 Training Data Preparation

Recently, the Machine learning approach has gained much attention in AM due to its capabilities to resolve the issue of big data and easy post processability of the in-situ monitoring data. To perform supervised machine learning, the need for labeled data is vital. However, it is challenging to obtain ground truth certified data for the AM parts as the drift in the process can be linked to many parameters [20]. Usually, to obtain a labeled dataset commonly referred to as certified dataset, the expensive techniques such as computer tomography of the part is done. For this study, an artificially drift (overheating and lack of fusion) in the parts with special geometrical features and by varying process parameters was created. The careful selection of the layers from the build parts (81 parts) and labeling it as ‘Drift’ and ‘No-drift’ was performed by analyzing the layers in the MPM software provided by the SLM solutions and also by statistical analysis. It was noticed that the layers which having the hotspots show a right shift in the histogram compared to the layers without any hotspots. This relation led us to decide the mean and median as our features for every layer. A balanced labeled dataset of 600 data points, which comprises an equal number of drift and no drift layers so that the biasing of the SVM model can be avoided was prepared. Before preparation, the SVM classifier has to be optimized for best-fit parameters, as discussed in the next section.

4.5 Bayesian Optimization and Training of SVM Classifier

The best-fitting parameters are to be selected for the SVM classifier to increase the success rate of the classifier. To find the best-suited hyperparameters for the model, a bayesian optimization algorithm was used. A hyperparameter is an internal parameter of an algorithm that needs to be optimized. For example, in our case (SVM model), the box constraint, kernel-function, and kernel-scale are the hyperparameters. These parameters can significantly influence the performance of the algorithm. Thus, optimization of the hyperparameters is advisable. However, optimization is difficult and time-consuming. Therefore, Bayesian optimization is well suited for classification and regression algorithms in machine learning. The Bayesian optimization algorithm minimizes the objective function f(x) for x in a bounded domain. The f(x) can be scholastic or deterministic, which means it can return different results for the same point x. The overall working principle of Bayesian optimization can be found elsewhere [21, 22].

The cross-validation of the optimized SVM classifier was performed on the training dataset. The whole dataset was partitioned into 70% and 30% sub-datasets to cross-validate and check the performance of different hyperparameters of the SVM model. The performance and accuracy % of the different hyperparameters are tabulated in Table 3. It can be seen that the fine gaussian SVM has the highest accuracy with a cross-validation success rate of 95.5%. So, for our dataset, fine gaussian hyperparameter to train our SVM classifier was used.

Table 3. Bayesian optimization results of SVM classifier for different hyperparameters.
Full size table

Finally, the SVM classifier was trained on the whole dataset, as shown in Fig. 7a and predicted labels are shown in Fig. 7b. Further, testing of the SVM classifier is performed using the different datasets (which is not used during training) as discussed in the results section.

Fig. 7.
figure 7

(a) The un-clustered training dataset, and (b) labels predicted by the trained fine gaussian SVM classifier model.

Full size image

5 Results and Discussion

To check the performance of the classification at different thresholds, the Receiver Operating Characteristic- Area under the curve (ROC-AUC) was analyzed. ROC represents the threshold probability of the classifier, and AUC indicates the separability of the classifier, i.e., how well the classifier can distinguish between different classes. Higher the AUC, better is the model separability. ROC curve is plotted between TPR (True positive rate) vs FPR (False positive rate). For the model, the ROC curve was plotted, as shown in Fig. 8. The threshold probability is 95.5% (marked by an asterisk (*)), and the AUC for the model is 98.53%. In other words, the proposed model is 98.53% capable of allocating each data point with the correct label.

Fig. 8.
figure 8

ROC-AUC curve for the trained SVM classifier on 600 training data.

Full size image

5.1 Overheating Drift

The trained SVM classifier was tested for the overheating samples. The feature dataset (mean and median) was prepared for two different overhang samples (Fig. 3a and 3b), which failed due to the overheating of the layers. The specific samples were used for the case study as the location of the overheating drift is known and can visually be verified. The prepared feature dataset was passed to a trained SVM classifier, and labels were predicted, as shown in Fig. 9. For the cubic overhang sample, the layer numbered from 360 to 378 was labeled as “drift,” which can also be visually verified (printed part is shown in the top left corner of Fig. 9a). The failure of the particular layers was due to poor heat transfer in the absence of proper support structure for the overhang part. Similarly, the labels were predicted for the cylindrical overhang part (Fig. 9b) and are in conjunction with the visual inspection of the part.

Fig. 9.
figure 9

SVM model predicted labels for (a) cubic overhang specimen, and (b) cylindrical specimen.

Full size image

5.2 Lack of Fusion Drift

Lack of fusion is the most common defect in the SLMed parts, which results in internal porosity in the final part. The detection of internal porosity is a challenging task and often requires expensive techniques such as computer tomography. We prepared samples that will lead to internal porosity in the final part, and the location of the porosity is known to check the trained SVM classifier for lack of fusion defect. The feature matric (mean and median) for the lack of fusion part, as shown in Fig. 3d is prepared and passed to the trained SVM classifier. The SVM classifier predicts the labels for each layer, as shown in Fig. 10a. The three layers numbered from 190 to 192 were predicted as “drift,” which correspond to the location of porosity in the sample. The same was also verified by the optical micrograph, as shown in Fig. 10b. Therefore, it can be concluded that the trained classifier works well for the presented case studies. However, it shall be noted that the effectiveness of the model for other types of drifts shall be studied in the future.

Fig. 10.
figure 10

(a) Predicted labels for the lack of fusion defect, (b) optical micrograph of the defect.

Full size image

6 Conclusion

An attempt to classify the in-situ monitoring data into drift and no-drift was made using a supervised machine learning model (SVM classifier). For this study, the sensitivity of the installed monitoring system was studied for artificial defects in the built-up specimen. The varied process parameters were purposely chosen to obtain a variety of drift and no drift data points for training the SVM classifier. The in-situ monitoring data is treated as a two-class problem (“drift” and “no-drift”), and the SVM classifier used to predict the drift layers in the parts studied. The feasibility of the trained algorithm to detect the drift in the process was verified by comparing it with the Melt pool monitoring system. The post processability of the enormous sized in-situ data can be significantly improved with the help of the ML models. Also, the use of ML models can be a revolutionary step in improving and monitoring the SLM process in real-time.

The future aim of this study is to test the classifier on real case complex parts and verify it with the micro-computer tomography dataset. Also, to study the reliability of the ML model for other materials.