Online monitoring of direct laser metal deposition process by means of infrared thermography

Abstract

Direct laser metal deposition (LMD–DED) is an additive manufacturing (AM) process that is used to build up and repair high-quality metal components. It works by overlapping layers of powder material and melting them with a laser. To get a stable process without defects and to reach, at the same time, high mechanical properties, a robust assessment and control of the process parameters, and above all of their combination, is required. The ideal goal is to assure the online control, to stop or correct the process in case of unexpected anomalies. In this work, a robust online monitoring of the laser metal deposition (LMD–DED) process based on the use of infrared thermography was developed and proposed. After choosing the suitable process parameters, a customized design of experiments (DOE) was set, and the statistical analysis of different thermal features was carried out to develop the most robust models that correlate them with the input process parameters (laser power, scanning speed, and powder flow rate). The proposed procedure was based on the extraction of different thermal features from suited regions of interest (ROI), performing statistical analyses by means of analysis of variance (ANOVA) and building regression models to correlate the process parameters with the thermal behavior. The obtained results demonstrated the possibility to control the process by means of the chosen thermal features, independent of the position of the ROI. Moreover, the possibility to use the models to detect typical AM defects, and anomalies, online directly during the process, has been proved and verified by destructive macrographs carried out on the manufactured coupons.

1 Introduction

Additive manufacturing (AM) offers a range of novel and interesting applications in different fields to produce metal components with a complex geometry and high material mechanical properties [1]. However, the proper setting of process parameters and their combination is an important point to deal with since it affects the structural integrity of the produced parts [1].

One of the most used AM processes is the laser metal deposition (LMD–DED), that commonly provides industrial and commercial solutions for repairing high-value expensive parts and components [1,2,3,4,5]. The LMD–DED process is used for applications in the oil and gas, automotive and aerospace industries, especially in the case of the production and repair of high-quality claddings and coatings [6, 7]. Basically, during the process, a laser beam is used to form a melt pool on the surface and the metal powder is then injected using a gas stream. Main process parameters are the laser power, the scanning speed, the energy density, and the powder flow rate. Depending on these parameters and their combinations, typical defects can occur such as porosities, keyholes, lack of fusion due to trapped gas or falling of material dust, and cracks [7,8,9,10,11]. Moreover, residual stresses, and inhomogeneous microstructures, might compromise the final mechanical properties and the structural integrity of the produced parts [6,7,8,9].

The most common approaches for controlling the process concern a proper monitoring mainly based on visual control systems (CMOS-, CCD-), or infrared thermal cameras [12,13,14,15,16,17,18,19,20,21,22,23] as well as pyrometers and structured light scanning that have also been used to provide information around the melt pool [24, 25]. Also, non-optical methods such as acoustical sensors and thermocouples have also been used for monitoring and control the AM process [8]. Another approach concerns the possibility to control the produced components through an offline mode by means of non-destructive and destructive tests [19, 26]. These latter can be applied to verify the presence of defects after the process and, eventually, can be used in association or to validate the implemented online solutions.

In this complex context, thermography could be a valuable tool both for the online monitoring and the offline non-destructive evaluation of defects, as already demonstrated in different recent works [21,22,23,24,25,26,27,28,29]. With respect to the use of simple pyrometers, that limit the measurement in a restricted area, the process monitoring with focal plane array (FPA) sensors allows for acquiring a full-thermal field, making the monitoring strategy particularly effective.

Although the number of research works on thermal methods used for the LMD–DED process monitoring has grown in recent years [5,6,7, 11,12,13, 18,19,20,21,22,23], there are still a few of them that focus on procedures developed to correlate directly thermal and process parameters with statistical models, that are also able to give information about the presence of any anomalies.

Many papers try to investigate the LMD–DED process, starting from the analysis of the melt pool. In particular, Liu et al. [11], performed a real-time temperature detection of Inconel 718 deposition in the LENS process by means of a high-resolution IR camera. In their work, the effects of deposition parameters on the melt pool temperature and cooling rate have been examined. They demonstrated that it is not easy to separate the various effects of the process parameters, especially if the monitoring regards the melt pool area.

There are some works based on the analyses of thermal gradients and specific algorithms of post-processing to study the influence of process parameters on the final material microstructure. Among them, in the work of Campanelli et al. [15], a thermal analysis was carried out to understand the relationship between the key process parameters, the thermal fields produced during the treatment, and the effect on the characteristics of the cladding for nickel-based superalloy. However, the analysis regards the monitoring of a single-track deposition, hardly adaptable to multiple passes and several layers as it occurs for complex components in industrial applications. Moreover, no monitoring models were provided.

Some works are focused on investigating the performance of IR sensors. The work of Altenburg et al. [18] based on the comparison between MWIR and NIR sensors for the duplex stainless steel AISI 2205 concluded that the NIR and MWIR cameras are more suitable to monitor temporal and spatial temperature changes than the visual camera. This study clearly demonstrated that the surface emissivity strongly depends on the material state (solid, liquid, oxidized, roughness), the wavelength, the temperature, and the observation angle. As the emissivity depends on many variables, it might be very time consuming to experimentally determine all emissivity curves for all the considered variables. In [18], only a temperature calibration was performed using the known solidification temperature of the materials, but as declared by the authors, it allows for obtaining only a temperature estimation that is not suitable for numerical simulations and/or in predictions about material parameters.

Considering all difficulties related to the emissivity assessment, the definition of a model that does not rely on it would allow a more stable process monitoring. Anyway, a robust statistical approach is needed to minimize the uncertainty.

The aim of this work is to provide an effective procedure for control the LMD–DED process based on infrared thermography. The thermal behavior of coupons has been investigated and a new approach based on extracting several thermal features significantly correlated with the process parameters is proposed.

A microbolometer sensor integrated with the laser head has been used to monitor the production of different coupons made of Inconel 718. Following a design of experiments (DOE), the influence of the process parameters as the laser power, the scanning speed, and the powder flow rate has been evaluated and correlated with the thermal behavior, analyzing the data with a statistical approach based of the analysis of variance (ANOVA) [30, 31].

Different thermal features have been extracted monitoring the thermal behavior within regions of interests (ROIs), box and profiles, whose position has been changed several times, with the aim to find robust regression models able to describe the process parameters variation within the investigated range. The influence of process parameters combinations was also evaluated, studying partial sub-plans with the same procedures and methodology. Positions around and far away the melt pool area have been evaluated, returning statistically more robust results by monitoring the thermal features in the solid area of the material.

With the main aim to determine the best ROIs positions for thermal monitoring, regression polynomial models were built and optimized checking different statistical indexes, as the square correlation coefficients (R²), the root mean square error (RMSE) and the signal to noise ratio (SNR).

Furthermore, as one of the main results of this work, the proposed approach allows for assessing the presence of defects and anomalies, online, during the process. The defect evaluation has been assessed considering the thermal behavior of successive layers and finally performing destructive controls and macrographs as direct confirmation.

Finally, another motivation for this work is represented by the possibility to use a low-cost solution in terms of equipment and data processing. In fact, the proposed approach uses a microbolometer IR camera, less expensive than the one normally proposed for monitoring this kind of process (SWIR camera with both high resolution and frame rate [18]), and a small number of data (ROI and/or profile) to extract the thermal features.

2 Materials and methods

2.1 LMD–DED set-up and experimental plan

The LMD–DED set-up used to carry out the experimental campaign is a commercial laser metal deposition machine, up to a laser power of 4 kW and a wavelength of 1070 nm (Laser source YLS 4000 IPG Photonics Ytterbium Laser System and laser head KUKA). The laser head can be moved in x-, y- and z-directions within a circular volume space of 1200 mm × 1200 mm × 800 mm (Fig. 1a).

Fig. 1

a Photograph of the laser head, nozzle, and entire deposition platform; b example of realized coupon and its geometry

The material used for this experimental campaign is a nickel-based superalloy Inconel718. The production regarded 26 coupons of predefined geometry with an L-shape, as showed in Fig. 1b, of dimensions 150 mm × 6 mm × 150 mm (Fig. 1b). The chosen geometry is the one commonly used by the company to characterize a new material from a mechanical point of view; when the process is complete, tensile tests have been performed for each coupon and set of process parameters to determine the mechanical properties of the material as build in both directions, as showed with a sketch in Fig. 1b, saving material thanks to the designed geometry. Moreover, several hardness tests have been performed. The tensile and hardness results will be shown in future publications. The scanning strategy foreseen six side-by-side passes and consecutive layers, for a total of about 150 layers, half long, and half short, as shown in Fig. 1b.

The experimental campaign regarded the analysis of three different process parameters and their main combinations, as laser power (P), scanning speed (v), and powder flow rate (PFR), following a design of experiments (DOE) of the type 3³, with 3 main levels and a central point as reported in Table 1. As it can be observed from Table 1, Level 2 is the central point, instead the extremes are indicated as Level 1 and Level 3 with the indication of the ranges.

Table 1 Scheme related to the analyzed DOE and complete subplan with main parameters, laser power (P), scanning speed (v), and powder flow rate (PFR)

Unfortunately, only 17 different coupons of the 27 forecasted in the full factorial plan of the type 3³ have been manufactured because some combinations of parameters did not make the process stable. The analysis was also performed on some sub-plans to analyze a complete experimental plan (3²—power fixed at Level 2) and also to determine the influence of the process parameters combinations reported in Table 1, in particular, the energy density (ED) and the powder flow length (PFL), defined as follows:

$$\begin{array}{*{20}c} {{\text{ED}} = \frac{P}{vD} \left[ {{\raise0.7ex\hbox{$J$} \!\mathord{\left/ {\vphantom {J {{\text{mm}}^{{2}} }}}\right.\kern-0pt} \!\lower0.7ex\hbox{${{\text{mm}}^{{2}} }$}}} \right]} \\ \end{array} ,$$

(1)

$$\begin{array}{*{20}c} {{\text{PFL}} = \frac{{{\text{PFR}}}}{v} \left[ {{\raise0.7ex\hbox{$g$} \!\mathord{\left/ {\vphantom {g {{\text{mm}}}}}\right.\kern-0pt} \!\lower0.7ex\hbox{${{\text{mm}}}$}}} \right]} \\ \end{array} ,$$

(2)

with D spot diameter, function of the adopted process parameters, held constant around an averaging value of 2.24 mm. To evaluate both the process and thermal measurements replicability, some coupons were produced as replication, considering their variability in the data analysis. The value of each Level has not been reported because of a confidential agreement with the company that provided the coupons. However, the indication of the investigated range for each process parameter, together with the difference with respect to the central value can be considered sufficient for the aims of the work.

2.2 Thermographic set-up: preliminary evaluations and final choice

The adopted set-up is reported in Fig. 2. The two thermal sensors (FLIR A655sc indicated as 2 and 3 in Fig. 2) are microbolometer sensors in the range 7.5–14 μm, both used within the highest temperature range 200–3000 °C as reported in Table 2. In this work, only the thermal data related to the thermal sensor integral with the laser scanning head have been analyzed.

Fig. 2

Experimental thermographic set-up: (1) thermal source (laser head), (2–3) microbolometer thermal sensors FLIR A655sc and (4) platform for the deposition

Table 2 Acquisition parameters adopted during the monitoring of the LMD–DED process

The first part of the experimental campaign has been dedicated to the definition of the best set-up concerning the mutual position among the laser, printing platform and thermal camera. In this regard, the main aims were to reach both a good spatial resolution and guarantee repeatable thermal measurements. In Fig. 3a, b, c and d are reported some set-up configurations evaluated for preliminary thermal measurements. In particular, Fig. 3a shows a thermal map in which the thermal sensor has been placed in a fixed position beyond the printing platform. This position does not allow to reach a high geometrical resolution (higher than 0.5 mm/pixel), and, in any case, it continuously changes according to the printing position. Figure 3b and c are related to other positions of the thermal camera integral with the laser head that have been discarded as the IR cameras are too far from the melted area and some parts of the coupons are hidden during the printing. Finally, in Fig. 3d, the final choice is reported, corresponding to the set-up and parameters reported for the IR sensor integral with the laser head. Figure 3c and d refers to the temperature calibration range 100–650 °C for which it is possible to observe several saturated pixels around the melt pool area. For these reasons, the range of 100–650 °C has not been considered for further analysis as it does not allow to capture the thermal behavior in the melt pool.

Fig. 3

Choice of the best position for thermographic acquisitions and parameters: a unique platform and camera in a fixed position, mm/pixel too low, b camera integral with the laser head, temperature calibration range 300–2000 °C, hidden view, c camera integral with the laser head—the view is too far from the melted area and with the temperature calibration range (100–650 °C), d camera integral with the laser head, temperature calibration range 300–2000 °C, final choice

3 Procedure for data analysis

The scheme of the adopted methodology is reported in Fig. 4. The raw thermal data were acquired in terms of apparent temperatures, and, after the data processing, the thermal features were extracted and correlated with the process parameters. Three different subsequent layers and all the six passes have been selected for each specimen and then analyzed (Fig. 4-1). All sequences were forced to the same number of frames, by sub-sampling the original ones and considering about 4500 data for each coupon to simplify the statistical analyses and have the same number of elements for each sequence and set of parameters.

Fig. 4

Schematic and synthetic flow of the procedure of data analysis

Two different strategies were investigated using both a box and a profile as a ROI to extract the thermal features over time (Fig. 4-2). To investigate the robustness of the chosen features, both the position and the dimension of the ROI was changed. The starting position of both box and profile includes part of the melt pool, while the final position is at a certain distance from the melt pool (Fig. 4-3).

A simple scheme of the used ROIs is reported in Fig. 5. In particular, following a series of preliminary results, a box of dimensions 13 × 35 pixels was selected and then the position has been changed along X and Y directions, 15 pixels along X and 5 pixels along Y, respectively, considering a step equal to one pixel. The starting position of the ROI (the coordinates and the initial frame of the thermal sequence) has been identified for each sequence in a semi-automatic way; in particular, for frame identification, an area around the laser head was considered together with its mean temperature value, considering as a threshold value the background apparent temperature when the laser is off equal to approximately 370 °C (Fig. 4-3). The spatial information has been obtained starting from the identified first frame and localizing the maximum temperature value and its coordinates (Fig. 4-3). In this way, for each sequence, the initial position was always obtained just beyond the laser print, as shown in Fig. 5.

Fig. 5

a Scheme of data extraction considering the case related to the analysis of a box (0 is the starting point—column 1, row 1) and b a profile, the color-bar represents the apparent temperature pixel by pixel

A similar procedure was also adopted for the profile, changing the dimension five times along the Y direction and the position along the X direction 35 times, considering a step equal to one pixel. In this way, all the thermal features were extracted 175 times.

The extracted thermal features are here listed: mean, standard deviation, 98th percentile, 2nd percentile, kurtosis, skewness, and mode of the apparent temperature (Fig. 4-4). All these features were extracted from the chosen ROIs, frame by frame. In the case of skewness and kurtosis, the feature for a box refers to the dimension along Y because, along this direction, it is possible to recognize a characteristic distribution of temperatures along a certain profile due to powder distribution during the deposition, as described in [22, 23]. In Fig. 6 the characteristic thermal profile is reported referred to a generic thermal map and a profile taken just downstream of the melt pool.

Fig. 6

a Generic thermal map and typical distribution of the powder during the deposition with consequent thermal trend b for a generic profile downstream of the melt pool

As already said, to have a good compromise among the amount of data to be analyzed and a statistical significance of the procedure, the thermal data set has been built up considering the thermal sequences of three layers and all the six passes for each coupon. In this way, a total of 18 laser passes have been selected (36 if the replication is considered).

As an example of the results, the trend of each thermal feature as a function of the number of frames is reported in Fig. 7 for a specific coupon, for a box (ROI) considered in a certain position (last one—column 15, row 5). In the same Fig. 7, the effect of the laser repositioning among passes is evident.

Fig. 7

Thermal features representative of the apparent temperature distribution within a box taken in correspondence of the last investigated position considering a particular coupon and so a set of process parameters (coupon 12—low energy density P—level 2, V—level 3, PFR—level 1)

Then, to compress the information related to each thermal feature, the mean value within each pass has been considered, obviously not considering the return paths of the laser. So, for each layer, 6 different values were collected. This operation was repeated for 3 layers (6 in case of replications) and for each ROI position, obtaining as synthesis a box plot, that represents the thermal behavior associated with a particular set of process parameters, for each extracted thermal feature.

The described procedure concerned the extraction of data from both boxes and profiles. Moreover, the vertical profile has been used for obtaining an effective thermal map, representative of the thermal behavior for an entire deposition layer. As an example, in Fig. 8 is depicted the map obtained by placing the profiles side by side for a given coupon and layer.

Fig. 8

Synthetic thermal map that represents a layer of a printed coupon, placing a vertical profile, frame by frame (coupon 12—low energy density P—level 2, V—level 3, PFR—level 1)

To assess the influence of the ROIs position on the thermal features, the described procedure was repeated changing the position of the ROIs, and performing a statistical analysis by means of ANOVA, considering, in particular, the result related to the p value, with a significance level of 0.05 (Fig. 4—points 6 and 7).

The statistical analysis regarded the complete experimental plan and so the influence of the main process parameters (power, scanning speed, powder flow rate) on the thermal response, and also some sub-plans to evaluate the significance of combinations of some parameters, such as the energy density. All the performed analyses are summarized in Table 3, reporting the factors, the reference levels, and the considered thermal responses. The experimental plans analyzed in this work are shown in detail in Table 1.

Table 3 ANOVA analysis carried out for each ROI considering the starting DOE and some sub-plans and some details related to the analyses

It was not possible to evaluate the effect of the interactions among the different process parameters with the complete plan because of missing data, as explained before. So, the assessment of the interaction effects, were partially analyzed by means of the sub-plans.

After verifying the statistical significance of the chosen thermal features, a polynomial regression has been performed to model the thermal features as a function of the process parameters (Figs. 4-8).

The models were then used and as a baseline of standard behavior of sound process so to be able to identify anomalous process variations or localized defects.

Second-degree polynomial models with first-order interactions were considered, where the variables represent the investigated process parameters:

$$\begin{array}{*{20}c} {{\text{Thermal feature = Ax}}^{{2}} {\text{ + By}}^{{2}} {\text{ + Cz}}^{{2}} {\text{ + Dx + Ey + Fz + Gxy + Hxz + Iyz + L}}} \\ \end{array} ,$$

(3)

$$\begin{array}{*{20}c} {x{\text{ = power }}\left( {P - {\text{kW}}} \right){, }\,y{\text{ = scanning speed }}\left( {v - \frac{{{\text{mm}}}}{{{\text{min}}}}} \right){, }z{\text{ = powder flow rate }}\left( {{\text{PFR}} - \frac{{\text{g}}}{{{\text{hours}}}}} \right)} .\\ \end{array}$$

(4)

No higher-order polynomial degrees than two were considered because of the limited degrees of freedom available for the analysis.

The goodness of obtained models has been evaluated considering well-known quality indexes (Fig. 4), as reported below:

$$\begin{array}{*{20}c} {R^{2} = 1 - \frac{{\mathop \sum \nolimits_{i = 1}^{n} \left( {\widehat{{y_{i} }} - y_{i} } \right)^{2} }}{{\mathop \sum \nolimits_{i = 1}^{n} (y_{i} - \overline{{y_{i} }} )}}; \,{\text{RMSE}} = \sqrt {\frac{1}{n}\mathop \sum \limits_{i = 1}^{n} \left( {y_{i} - \widehat{{y_{i} }}} \right)^{2} } ; \,{\text{SNR}} = 10 \times {\text{Log}}\frac{{\mathop \sum \nolimits_{i = 1}^{n} (\widehat{{y_{i} }} - \overline{{y_{i} }} )^{2} }}{{\mathop \sum \nolimits_{i = 1}^{n} \left( {y_{i} - \widehat{{y_{i} }}} \right)^{2} }} } \\ \end{array} ,$$

(5)

where $\widehat{{y}_{\mathrm{i}}}$ are the estimated values with the polynomial regression, ${\mathrm{y}}_{\mathrm{i}}$ are the experimental values, and $\overline{{y }_{\mathrm{i}}}$ is the mean, while ${R}^{2}$ is the coefficient of determination, $\mathrm{RMSE}$ represents the Root Mean Square Error and $\mathrm{SNR}$ the signal to noise ratio. Obviously, the aim was to maximize the ${R}^{2}$ and the $\mathrm{SNR}$, and to minimize instead the $\mathrm{RMSE}$. Another quality index used to evaluate the goodness of the regression model was the variance associated to each thermal feature and set of process parameters; in this sense, an analysis to minimize the variance associated to each case has been carried out, considering the variability of each box containing the thermal data associated with a particular thermal feature.

4 Experimental results and discussion

4.1 Analysis of the DOE: ANOVA and regression models

A typical issue that needs to be investigated is the sensitivity of the extracted feature to a small variation of the position of the measurement point/line/area. In this regard, to investigate at the same time, the behavior of each feature, ROI and its position and small variations of the ROI position, a statistical approach was used.

As example, the obtained results for a specific coupon (coupon 12—low energy density P—level 2, V—level 3, PFR—level 1), are reported in Figs. 9 and 10. In particular, the results for each considered thermal feature, are reported as a function of the box position and in terms of mean (Fig. 9) and variance values (Fig. 10). On the x-axis, the ROI position along the columns (indicated as X in Fig. 5a) is reported, while the position along the rows (indicated as Y in Fig. 5a) has been reported by means of five different colors. Referring to the Fig. 5a, the solid line box represents the starting position of the ROI and an increase of the × value describe a movement of the box toward the left side. Similar consideration can be done considering as ROI a profile as schematically represented in Fig. 5b.

Fig. 9

Average values for each thermal feature considering the analysis of different box positions and the case of a specific coupon printing with a precise set of process parameters (coupon 12—low energy density P—level 2, V—level 3, PFR—level 1)

Fig. 10

Variance of each thermal feature considering the analysis of different box positions and the case of a specific coupon printing with a precise set of process parameters (coupon 12—low energy density P—level 2, V—level 3, PFR—level 1)

As it can be noticed, the box includes part of the melt pool at least for the first positions along x direction. In Figs. 9 and 10 the dotted lines represent the boxes for which the melt pool is partly included.

It is possible to notice that the choice of the position along the y direction has a great influence on the stability of the thermal data (change of the mean value of each single thermal features). Moreover, a stable behavior for most of the considered thermal features can be observed when ROI positions are far away from the melt pool zone.

To investigate the effect of the ROI position in a statistical way, a two-ways ANOVA analysis was carried out considering different box positions for all the coupons, to demonstrate that the data do not change significantly (p value ≥ 0.05) when the ROI position change.

In particular, a first analysis regarded all the investigated box positions, 5 positions along y direction (rows), and 15 positions along × direction (columns)—Table 4, Fig. 11a. For most of the coupons, there is a statistical influence of the box position when all the columns are considered but, the influence become not significant neglecting the first 5 columns (not considered in the analysis), as reported in Table 5 and Fig. 11b. However, for some coupons, in correspondence of a given set of process parameters there is still an influence in the columns position. In general, as shown in Fig. 11c, the farther the position from the melt pool, the less the influence of the box position along the column is on the results.

Fig. 11

Influence of the box position for the analysis related to the mean apparent temperature, considering 4 different cases, and the analysis of boxes at different positions and distances from the melt pool

There is still a certain influence of the box position regarding the row position (examples in Tables 4, 5 and 6). Also in this case, as for the analysis of the columns, the influence becomes not significant if the first and last row positions are not considered in the statistical analysis (data related to row 1 and row 5 as indicated in the previous graphs, see the results in Table 7, for example, and Fig. 11d). These results can be explained considering that the first and last rows probably include values that increase the variance of the considered thermal features including background or other passes, only for few pixels.

Tables 4, 5, 6 and 7 report the results for the coupon 12 as an example, and in the Fig. 11, the obtained p values are collected and synthesized for all the coupons (represented with numbers and dots with different colors in Fig. 11). All the results refer to the mean apparent temperature but, similar considerations can be done for the other thermal features.

Table 4 ANOVA analysis to establish the influence of the box position for the thermal features extraction: analysis considering all the positions (5 rows and 15 columns)—coupon 12

Table 5 ANOVA analysis to establish the influence of the box position for the thermal features extraction: analysis considering 5 rows and the columns from 6 to 15—coupon 12

Table 6 ANOVA analysis to establish the influence of the box position for the thermal features extraction: analysis considering 5 rows and the columns from 11 to 15—coupon 12

Table 7 ANOVA analysis to establish the influence of the box position for the thermal features extraction: analysis considering the rows from 2 to 4 and the columns from 11 to 15—coupon 12

Furthemore, all the previous results demonstrate that the iterations are significant. This means that the model can be considered purely additive, with rows and columns that can be considered as single factors, in one-way ANOVA analyses.

An ANOVA analysis was carried out to verify the correlation between the identified thermal features and the process parameters as a function of the ROIs position. Considering the complete experimental plan, the obtained results demonstrated the statistical significance of the process parameter on the thermal features (p value close to 0), for ROI positions far from the melt pool.

In Fig. 12 (apparent mean temperature vs laser power), it is interesting to notice that the ROI position is important in terms of capability of the thermal feature to describe the change of the laser power. In particular, it is important to remark that this process parameter can become not significant for ROIs very close to the melt pool.

Fig. 12

ANOVA for the apparent mean temperature considering the analysis related to the complete experimental DOE; statistical influence of the investigated process parameters when the box position for data extraction changes

Summarizing, once the box is set few pixels away from the melt pool, regardless of the row position, all considered process parameters influence the thermal features and a model can be defined. Similar results were obtained for all thermal features. It is important to remark that for box positions far away from the melt pool the p value is almost equal to 0.

As an example of the obtained results, in Table 8 the ANOVA analysis for the mean apparent temperature is shown for the last box position (column 15, row 5).

Table 8 ANOVA with the results related to the last investigated box position (column 15, row 5) for the analysis of the apparent mean temperature

To directly correlate the thermal features with the process parameters, different regression models were considered and compared. The choice of a regression instead of other possibilities such as, for example, surface response model, is due to the fact that for some of the chosen set of process parameters it was impossible to obtain a proper specimen and so the experimental plan is not complete. The proposed procedure can be improved including further tests for different set of parameters and adopting a more suitable regression model to include such new data. As polynomial degree, due to the number of degrees of freedom, the second degree was chosen for all the variables, also considering the interactions, as described before in the procedure.

In Fig. 13, the results related to the models and their quality indexes are reported, considering the different box positions. These results confirm, as previously shown with the ANOVA, that positions furthest from the melt pool make the extracted features more stable and less sensitive and so the results are less variable.

Fig. 13

Trends of the coefficient of determination ${R}^{2}$ for each thermal feature and investigated box position

The strongest correlation is obtained by analyzing the data of the last columns related to the apparent mean temperature (R² = 0.81, column 15–row 4). Even the analysis of the 98th percentile seems to provide good results in terms of correlation (R² = 0.76, column 15–row 1). Similar results were obtained also considering as a ROI the vertical thermal profile (maximum value 0.87, for mean apparent temperature in same position column 15–row 4).

Similar results and considerations can be obtained considering the trends related to the other quality indexes, as reported in Fig. 14, where also the RMSE and the SNR are reported to verify the correlations between the apparent mean temperature and the process parameters.

Fig. 14

Quality indexes trends for different box positions relative to the correlation “apparent mean temperature—process parameters”

As an example of the obtained results in term of regression models, the result related to the apparent mean temperature for the position that maximize the R² value (row 4–column 15 R² = 0.86) is reported in Fig. 15.

Fig. 15

Regression models obtained for three power levels, for the position that optimize the R² value, row 4–column 15

As indicated in Eqs. 3 and 4, the independent variables are 3 and represented the main process parameters, while the dependent variable is the adopted thermal features. To represent the result as a surface in a 3D plot, after obtaining the equation model, the level of power was fixed for the three levels and the surface were represented as a function of the scanning speed and powder flow rate.

To monitor anomalies and provide an alarm during the process, the models that correspond to a minimum of variance of the thermal features were obtained. For these models, the other quality indexes were also calculated, together with confidence bounds. As an example of the obtained results, the one related to the mean apparent temperature is provided in Fig. 16a, while in Fig. 16b the same model is shown considering the central level of the laser power and the related confidence bounds.

Fig. 16

a Regression models related to the mean apparent temperature considering the minimum of variance for each coupon (three power levels) and b representation of the central power level with confidence bounds

For completeness, in Table 9 are reported the RMSE, SNR, and R² values for the regression model shown in Fig. 16 (power level 2). Choosing the model related to the minimum variance for each set of process parameters, a good compromise that also optimizes the value of the other quality indexes was obtained.

Table 9 Values related to the different quality indexes related to the regression model (power level 2) of the mean apparent temperature (Fig. 16) where the variance for each coupon reaches its minimum value

4.2 Analysis of a complete subplan: ANOVA results and regression models

Sub-plans can be extracted from the main one by fixing one of the process parameters. In this way, it was possible to consider a complete factorial plan with three levels and two repetitions and then consider the interactions between the process parameters. For example, the subplan related to the central level of the laser power is complete, considering all the 3 levels available for the statistical evaluation of the other two process parameters. Figure 17 shows the result related to the ANOVA analysis associated to this subplan, considering the interaction between the scanning speed and the powder flow rate and the apparent mean temperature as a thermal feature. Also in this case, small variations of the box position do not affect the statistical influence of the chosen thermal features, demonstrating the robustness of the proposed approach. Similar results can be obtained for the other thermal features and for the analysis of other sub-plans, where, alternatively, the scanning speed or the powder flow rate are fixed.

Fig. 17

ANOVA for the apparent mean temperature considering the analysis related to a particular subplan where the laser power is fixed; statistical influence of the investigated process parameters when the box position for data extraction changes

In Table 10, the ANOVA results that come from the previous analyses were explicitly stated for the last position of the ROI. Again, as said before, the results in Table 10 are referred to the mean apparent temperature.

Table 10 ANOVA with the results related to the last investigated box position (column 15, row 5) for the analysis of the apparent mean temperature and the subplan with a fixed value of the laser power

Moreover, also for the analyses of the sub-plans, it was possible to evaluate regression models. As an example, the position that maximizes the R² value was considered as a reference (Fig. 18a), and the related model is shown in Fig. 18b. As expected, better results, for example in terms of R², were obtained with respect to the full plan because of fixing a process parameter.

Fig. 18

R² trend (a) and regression model that maximize this quality index (b) for the thermal feature related to the apparent mean temperature when the level of power is fixed (analysis of a subplan)

4.3 Defects and anomalies identification through the online monitoring

As a main result of the thermal monitoring, it is possible to identify defects and anomalies with the analyses of the proposed thermal features. The analyses regarded the thermal features identified as the best ones to correlate process parameters and thermal behavior. Thanks to the use of a Focal Plane Array sensor (FPA) instead of a single pyrometer, it was simple to identify defects and anomalies during the online monitoring, considering only a thermal map and different instants of time during the acquisition (Fig. 19a, b and c). The reported anomaly concerns the obstruction of the nozzle during deposition and therefore the fall of residual melted powder on the component.

Fig. 19

Some significative instants of time (frames) that allows the identification of an anomaly during the deposition

A powerful tool that allows the anomaly identification is the map related to the thermal profiles taken at a certain distance from the melt pool and placed side by side. As an example of the obtained results, Fig. 20 shows the maps of the thermal profiles reconstructed placing side by side the thermal profile assessed for each frame. The result is a localized anomaly in the thermal map that produces an increase of the apparent temperature that also affects the passes of subsequent layers, as reported in Fig. 20b). It is interesting to notice that is possible to detect the presence of thermal anomalies in the thermal profile even before the defect appears as a localized zone. As reported in Fig. 20a, from frame 6000 onwards, anomalies of the thermal profile appear due to a reduced powder during the deposition of the first layer (the apparent temperature increases since the nozzle is clogged). Then, a defective area occurs around frame 4000 due to the entrapped powder that falls during the deposition of a subsequent layer, Fig. 20b. The thermal behavior related to these two layers is then reported in Figs. 22 and 23.

Fig. 20

Maps of the thermal profiles when an unexpected anomaly occurs during the process: powder trapped inside the nozzle falls out and changes the thermal behavior along the selected profile (coupon 1, laser power Level 2, scanning speed Level 1, PFR Level 1, higher energy density ED 51.3 J/mm²)

Figure 21 reports the maps of the thermal profile for the same layer shown in Fig. 20b, considering the profile at a different height, and, respectively, 20 pixels before and 20 pixels after the previous case. This result demonstrates that small variations of the profile position does not affect the obtained results, as the anomaly can be, in any case, detected.

Fig. 21

Maps of the thermal profiles when an unexpected anomaly occurs during the process: monitoring a thermal profile in a different position (± 40 pixels) it is always possible to detect the anomaly

It is also possible to identify the presence of the anomaly considering the trend over time of the thermal features, as shown in Fig. 22 for the mean apparent temperature. Moreover, the trend of the thermal feature in nominal conditions (without defects), acquired for a layer just before appearing the anomaly for the same set of process parameters and coupon, was added. As shown in Fig. 22, the presence of the anomaly produces an abrupt variation of the signal that exceeds the confidence bounds. This result demonstrates the possibility for a real-time control by simply setting an alarm during the thermal monitoring.

Fig. 22

Regression model related to the mean apparent temperature and to the analysis of data when an anomaly affects some passes and layers

It is possible to have similar information also considering the analysis of other thermal features, as reported in Fig. 23 in terms of standard deviation (a) and 98th percentile (b), for the same layers in presence of an anomaly. This result demonstrates that the simultaneously analysis of multiple thermal features can provide complementary or synergic information useful for the anomaly detection.

Fig. 23

Standard deviation (a) and 98th percentile trend during the monitoring of a box when an anomaly occurs during the layer deposition

To investigate the presence of any defects in the area in which a thermal anomaly was observed, a macrograph was performed using an acid etch capable of revealing the different passes during the deposition and the presence of defects. The macrograph confirms the presence of defects in the area interested by the thermal anomaly. Upon closer analysis, it can be seen that there are two lateral indications (black areas) joined together by a marked line which indicates a gluing in the interested area of about 2 mm (Fig. 24b).

Fig. 24

Macrograph which reveals the presence of defects in the area identified by an anomalous thermal behavior during the monitoring

It is possible to get similar considerations when the analysis regards other process parameters and anomalies, as demonstrated before in [22].

5 Conclusions and outlooks

In this work, the thermographic technique has been adopted for monitoring the LMD-DED process. In particular, a procedure for analyzing the thermal data has been proposed for extracting suitable features that can be correlated with the process parameters. Experimental tests were carried out to product 17 coupons made of Inconel 718, following a customized DOE and all the tests were monitored with an IR camera mounted on the laser head. A statistical approach and different analyses have been carried out to establish the influence of the process parameters on the thermal behavior by extracting several features coming from both a square ROI and a vertical profile. Moreover, the robustness of the procedure has been evaluated by performing a sensitivity analysis of the thermal features to the ROI and profile positions. Finally, regression models have been adopted to mathematically describe the correlation between the process parameters and the adopted thermal features.

The main results can be summarized as follows:

• A correlation among the different thermal features and process parameters was established and described, considering the influence of each single process parameter and relative combinations; in this sense, an empirical model based on the mean apparent temperature was proved to be capable of describing the correlation between thermal behavior and main process parameters (laser power, scanning speed and powder flow rate), valid within the investigated ranges.

• The variation of the main process parameters causes a change of the thermal behavior, described by the identified thermal features; all the analyzed thermal features are sensitive to the process parameters variation, with a p value almost equal to 0, regardless of the position chosen for their extraction.

• If the laser power is considered as a main process parameter and the mean apparent temperature as a thermal feature, more attention must be paid to the choice of ROI position to have a statistically significant influence on the variation of this parameter.

• The analysis of suitable sub-plans confirmed that the extracted thermal features are sensitive to process parameters interaction such as the product between the powder flow rate, the scanning speed and the energy density.

• The analysis of the influence of the ROI position has given better and stable results when the ROI are far away from the melt pool area (the extracted data do not change significantly with the ROI positions.

• Among the extracted thermal features, the mean apparent temperature was the one that showed the best correlation with the process parameters.

• The proposed procedure can identify anomalies/defects due to an imperfect combination of process parameters or random errors that can occur during the process. In this regard, the monitoring strategy based on the acquisition of a single profile has proved to be very effective, since it allows to collect the data in a very compact form and allow the monitoring model construction and anomaly identification, independently from its precise position. In this way, this strategy can be very useful for the online monitoring of large pieces and components, for which the adoption of a linear thermal sensor could be a good compromise in terms of effectiveness and data storage.

Feature works will be focus on the direct correlation between the proposed thermal features and the mechanical properties of material and its structure.