Research on Process Control of Laser-Based Direct Energy Deposition Based on Real-Time Monitoring of Molten Pool

Abstract

In the process of laser-based direct energy deposition (DED-LB), the quality of the deposited layer will be affected by the process parameters and the external environment, and there are problems such as poor stability and low accuracy. A molten pool monitoring method based on coaxial vision is proposed. Firstly, the molten pool image is captured by a coaxial CCD camera, and the geometric features of the molten pool are accurately extracted by image processing techniques such as grayscale, median filtering noise reduction, and K-means clustering combined with threshold segmentation. The molten pool width is accurately extracted by the Canny operator combined with the minimum boundary rectangle method, and it is used as the feedback of weld pool control. The influence of process parameters on the molten pool was further analyzed. The results show that with an increase in laser power, the width and area of the molten pool increase monotonously, but exceeding the material limit will cause distortion. Increasing the scanning speed will reduce the size of the molten pool. By comparing the molten pool under constant power mode and width control mode, it is found that in width control mode, the melt pool width fluctuates less, and the machining accuracy is improved, validating the effectiveness of the real-time control system.

1. Introduction

Laser-based direct energy deposition (DED-LB) is an advanced additive manufacturing technology that uses a laser as an energy source to molten and deposit materials layer by layer, creating three-dimensional structures. This technology has shown significant application potential in high-end manufacturing fields such as aerospace, automobile manufacturing, and medical devices [1,2,3]. Its advantages include the ability to manufacture lightweight, high-performance integrated structures, flexibility in material feeding methods, high forming efficiency, broad forming freedom, and the feasibility of multi-process integration. However, the DED-LB process is a complex metallurgical process involving the coupling of multiple parameters. The slight changes in key parameters such as laser power, scanning speed, powder feeding amount, and defocusing amount [4,5,6] may affect the morphological characteristics of the molten pool. In addition, DED-LB technology also involves the coupling effect of multiple physical fields, such as thermodynamics and optics, which may lead to the non-uniform heat accumulation of heat in the deposition layer, which in turn leads to the stability of the molten pool, such as the formation of cracks, pores, and other defects [7,8,9]. Therefore, the geometric shape of the molten pool has a decisive impact on deposition quality, and real-time, accurate monitoring of the molten pool state is an important means to improve molten pool stability and dimensional accuracy, ensuring the quality of the final formed part.

Driven by the continuous advancement of computer vision technology, vision-based molten pool monitoring methods have gradually become a research hotspot [10,11,12]. These methods mainly rely on image capture technology to analyze the shape, size, and dynamic changes in the molten pool through image processing and pattern recognition techniques, thereby achieving real-time monitoring of the molten pool state. Currently, many scholars start from the process conditions of molten pool monitoring and are committed to developing more camera-based sensor systems [13]. Maffia et al. [14] developed a coaxial multi-sensor monitoring system that can simultaneously monitor the height, area, and temperature of the molten pool. By extracting concurrent signals from the workpiece area in real-time, the comprehensive state of deposition can be inferred. On this basis, Liu et al. [15] further realized the monitoring of multiple key parameters in the high-power laser welding process using six advanced sensors. These sensors capture detailed information about metal vapors, spatter, keyholes, and molten pools, providing comprehensive monitoring of the welding process. The auxiliary light source also plays an important role in the monitoring of the molten pool. Caprio et al. [16] can directly characterize the oscillating motion of the molten material through the off-axis configuration of the lighting lamp, thereby improving the clarity of the molten pool image. In experiments, we often find that when the camera and the auxiliary light source are installed on the rear side of the molten pool, the captured image is the clearest. However, due to the principle of spatial imaging, there will be a certain shape deviation in the state of the molten pool in the image when the camera side axis is used. In order to eliminate this kind of off-axis positioning error and restore the true shape of the molten pool, Gu Z et al. [17] proposed a three-dimensional information acquisition method for molten pools based on binocular vision and developed an image-matching algorithm. This algorithm uses the coordinates and color information of extracted feature points, significantly improving the accuracy of molten pool three-dimensional reconstruction.

In the process of DED-LB, the extraction of molten pool characteristics is facing great challenges due to the interference of environmental factors such as strong light radiation, molten droplet flow, and powder splashing. Therefore, image processing technology plays a vital role in this field. Sampson et al. [18] separated the molten pool edge from the substrate by utilizing the difference in light emissivity between the substrate and the molten pool edge, obtained its contour through the Canny operator, and then used morphological processing to obtain a clear internal structure of the molten pool. Recently, edge extraction algorithms based on active contour models have also been widely used in the process of DED-LB. Wang et al. [19] designed a multi-scale feature fusion semantic segmentation network, Res-Seg, using a residual network, which demonstrated excellent accuracy and robustness in actual welding environments. Vito Errico et al. [20] used a coaxial monitoring system to test the effect of the active contour, threshold segmentation, and Canny edge extraction algorithm of the region. The experimental results show that the efficiency and stability of the active contour of the region for image processing are more effective and stable than other algorithms. Cai et al. [21] proposed an image processing method based on U-net. By semantic segmentation of the welding process monitoring image, the multiple interference is effectively eliminated, and a more accurate molten pool contour is extracted. It is worth noting that the active contour model algorithm is easily affected by the local optimal solution when using the global information to extract the edge of the target object, and the selection of the initial contour position has an important influence on the accuracy of edge extraction.

In view of the limitations of traditional image algorithms, some scholars have applied improved machine learning methods to molten pool image processing. Yang et al. [22] proposed a real-time recognition method based on improved DeepLabV3+, which can overcome the instability of molten pool images caused by strong arc light and extract the contour of the molten pool faster and more accurately. Cai et al. [23] proposed an adaptive fusion method for adjacent monitoring images, which eliminates the interference of the metal vapor plume in the image and highlights the keyhole and molten pool. Therefore, the monitoring method based on machine vision can improve the consistency of detection, detect the defects of the deposition layer in real-time, and realize process control [24,25,26,27].

This study proposes a real-time monitoring method for the geometric characteristics of the DED-LB molten pool based on coaxial vision. This method constructs a comprehensive image processing algorithm framework to achieve accurate extraction of the contour and area of the molten pool. Through the in-depth analysis of the molten pool area and width under different laser power and powder feeding speed conditions, the molten pool width is used as the feedback of the control system, and the laser power is dynamically adjusted through the closed-loop control mechanism. The experimental results verify the effectiveness of the system in achieving accurate control of the molten pool width.

2. Materials and Methods

2.1. Coating Preparation and Image Acquisition

According to the experimental requirements, the laser-based direct energy deposition experimental system is composed of a laser (RFL-C3300, Raycus, Wuhan, China), scraper powder feeder (HW-05SF, Huawei, Shenzhen, China), water-cooled machine (CWFL-4000, Teyu, Huizhou, China), control system (SIEMENS S7-1200, Siemens, Berlin, Germany), ABB robot (IRB-1600, ABB, Beijing, China), and other equipment, as shown in Figure 1. The parameters of the fiber laser are as follows: the rated power is 3300 W, the spot diameter is 2 mm, the laser focal length is 10 mm, and the defocus amount is 2 mm. The Siemens S7-1200 PLC is used as the central processor of the DED-LB control system to control and coordinate these devices. The PLC exchanges data and sends control commands through communication interfaces with each device, achieving precise control and adjustment of the DED-LB process. The system has performed well in precision control, deposition efficiency, and deposition quality under the verification of a large number of experiments, which provides a certain reference value for modern industrial production. The material used is AlCoCrFeNi2.1 eutectic high entropy alloy powder produced by Xuzhou Velari New Materials Technology, with powder particle sizes of 75~150 µm. The substrate is 45 steel with dimensions of 100 mm × 100 mm × 10 mm. Before the experiment, the substrate was polished to reduce surface roughness and increase surface activity, then cleaned of oil and dust, and finally placed on the experimental platform with high-purity argon gas used as a protective gas.

To accurately monitor the molten pool characteristics in a high-brightness environment, a monitoring system capable of real-time molten pool image acquisition and serial transmission of image data to the host computer for analysis was built. In the process of DED-LB, the molten pool is affected by the environment, showing high brightness, low contrast, large fluctuation amplitude, and noise. This requires a high-resolution camera to improve image quality and processing efficiency. Based on these requirements, the WP-US230W color camera from Huagu Power(WP-US230W, Huaguang, Shenzhen, China) was selected, with a maximum frame rate of 50 fps, meeting the demand for real-time molten pool image capture, and a resolution of 1920 × 1080, capable of capturing clear molten pool images. The image is acquired through the checkerboard calibration board, and then the image is corrected, and the actual pixel ratio is calculated to be 9.76 × 10⁻³ mm/pixel. Narrow-band filters, neutral attenuation sheets, and quartz glass protective sheets were used to reduce the light intensity in the molten pool area, filter out some stray light, and enhance image contrast. Figure 2 shows the installation diagram of the overall laser head and visual monitoring components.

2.2. Research on Visual Monitoring Algorithms

2.2.1. Molten Pool Image Preprocessing

To deeply explore the correlation between molten pool images and deposition layer forming quality, an image processing algorithm was established to accurately extract the geometric features of the molten pool. The image preprocessing steps mainly include collecting the molten pool image, selecting the region of interest, grayscale transformation, median filtering, and image segmentation. Firstly, the molten pool image is collected by a coaxial CCD camera, and the non-essential dark area in the image is removed by cutting. Then, the weighted average method is used to gray the color image because it can retain the edge information of the image to the maximum extent compared with the maximum and minimum methods [28]. This method converts a color image into a grayscale image by a weighted average of the RGB channels of each pixel, and its general expression is formula (1) [29]. The original image of the molten pool is grayed in Figure 3a, and the gray image is obtained as shown in Figure 3b.

$$Gray=[0.2990.5870.114]{[RGB]}^T$$

(1)

In order to reduce the influence of noise on the edge extraction of the weld pool during the image acquisition process, a 3 × 3 template is used for median filtering to remove a large amount of noise generated during the image acquisition process. The expression is as follows:

$$Y_{\mathrm{i}}=Med\{f_{i-v},\dots ,f_i,\dots ,f_{i+v}\}$$

(2)

In the formula, i ∈ Z, V = (m − 1)/2, Z represents the set of integers, m represents the number of sequences, f_i represents the center value of the window, and Y_i represents the median value after sorting.

Figure 4 shows the processed image where high-brightness noise points were effectively removed, resulting in a relatively smooth interior of the molten pool with well-preserved edges.

To accurately extract key features from the captured molten pool images, this study proposes a method combining a K-means clustering algorithm and threshold segmentation to extract molten pool information. The expression for the K-means clustering algorithm is as follows:

$$J={{\displaystyle {\sum }_{i=1}^m{\displaystyle {\sum }_{j=1}^K‖x_i-{\mu }_j‖}}}^2•I\left(c_i=j\right)$$

(3)

where μ_j is the cluster center of the j cluster, and I(c_i = j) is the indicator function that represents whether data point x_i belongs to the j cluster.

The K-means algorithm can be used for fast image segmentation. As shown in Figure 5a, the molten pool image is initially segmented into three parts: background, plume, and molten pool, using the K-means clustering algorithm. Figure 5b further refines the segmentation using an iterative threshold segmentation method, where the molten pool region appears white and the rest of the image is black as the background. It can be observed that the combined method of K-means clustering segmentation and iterative threshold segmentation accurately extracts the molten pool contours, reducing the impact of small light spots caused by the heat-affected zone, plume, and powder spatter reflections on the molten pool image. The edges of the molten pool are also relatively smoother. This precise segmentation of the molten pool region avoids interference and noise from other areas.

2.2.2. Molten Pool Edge Monitoring and Feature Extraction

During the DED-LB process, the arc and heat-affected zone can impact the edges of the molten pool, leading to erroneous calculations of the actual molten pool features. Therefore, in this process, the edge detection algorithm is needed to extract the real edge of the molten pool. We chose several edge detection methods that are widely recognized in academia, such as Canny, Sobel, Laplacian, and Scharr operators. Figure 6 presents the edge extraction diagrams of the four operators along with their detailed magnified views. From the extraction results, it is observed that the Canny operator achieves the best extraction effect. In the edge detection process, the Canny operator employs the principle of gradient to locate pixels with the maximum gradient values. It suppresses the thickness and noise of edges through multi-level threshold processing and non-maximum suppression, accurately delineating edges. Both the Sobel and Scharr operators perform smoothing on the pixel points on both sides of the edges during the edge detection process, which can cause a shift in edge position and thickening of lines, thereby reducing detection accuracy. Moreover, compared to the Sobel operator, the Scharr operator can effectively extract weak edges, but it is highly sensitive to noise in the image, potentially leading to incorrect edge detection results. The Laplacian operator is an isotropic operator that disregards pixel intensity differences during edge detection; as can be seen from the image, the edge of the molten pool is somewhat blurred; hence, it is not suitable for the detection of molten pools in DED-LB processes. Therefore, we use the Canny edge detection algorithm to extract the true molten pool edges [30].

In the process of DED-LB, the direction of the molten pool changes with the scanning direction. To quickly and accurately extract the molten pool width from the image, this study employs the minimum bounding rectangle image algorithm. The principle of this method is to obtain the minimum outer rectangle by continuously rotating and translating an initial outer rectangle and comparing their areas. The specific method is as follows:

(a)

Obtain the initial external rectangle length l and width w from the image boundaries. Assume the coordinates of the four vertices of the rectangle before rotation are (x₁, y₁), (x₂, y₂), and (x₃, y₃).

(b)

Set the center point of the rectangle as O(x₀, y₀), then rotate the principal axis around the center point O by an angle θ. The four edges of the rotated rectangle are tangent to the edges of the molten pool, resulting in new vertex positions: (x_1′, y_1′), (x_2′, y_2′), (x_3′, y_3′), (x_4′, y_4′).

(c)

By continuously rotating the rectangle by θ degrees and calculating the rectangle’s area, the smallest external rectangle area can be obtained by comparing the areas. The width of the rectangle w’ is the width of the molten pool.

As shown in Figure 7, after detecting the edges of the molten pool, the width of the molten pool is extracted, where the length of the black double-arrow line indicates the width of the molten pool.

2.3. Control Principle

Experimental studies have shown that during the DED-LB process, as the substrate temperature increases, the molten pool temperature also rises. This heat accumulation causes instability in the molten pool width, leading to uneven wall thickness in the formed parts and directly affecting the dimensional accuracy of the parts. In our previous research, we designed single-factor experiments to investigate the sensitivity of four characteristic parameters on the molten pool formation. The results showed significant variations in the molten pool width and area, while the direction angle and centroid position were not significantly affected. Considering that the molten pool width is primarily sensitive to thermal parameters, we selected laser power as the control parameter and molten pool width as the output parameter in the real-time process control system. Due to the nonlinear relationship between the molten pool width and laser power, a PID controller was used to adjust the laser power to stabilize the molten pool width. The system information collection and transmission scheme is shown in Figure 8. First, the upper computer obtains images captured by the camera through a serial port, processes the images in real-time to calculate the molten pool characteristic information, and then transmits this information to the PLC via the Profibus network interface. The PID control program written in the PLC further calculates the analog voltage value U_x required to adjust the laser power output. The Siemens S7-1200 analog control module outputs the calculated analog voltage value U_x to the AD bus in the laser INTERFACE interface. U_x is then used to control the laser output power. This achieves the real-time control process from molten pool image acquisition to feedback control, with laser power selected as the control parameter and molten pool width and area as output parameters.

3. Results and Discussion

3.1. The Influence of Process Parameters on the Characteristics of Molten Pool

Laser power is the most critical process parameter in DED-LB, determining the energy density of the laser beam. Higher power means more energy is delivered to the workpiece surface by the laser beam. To study the relationship between laser power and the geometric characteristics of the molten pool, six sets of experiments were designed with a scanning speed of 240 mm/s, a powder feed rate of 12.79 g/min, and laser power levels of 1000 W, 1200 W, 1400 W, 1600 W, 1800 W, and 2000 W.

Figure 9 shows the images and characteristic parameters of the molten pool under different laser power levels. From the images, it can be seen that as the power increases, more energy is input to the deposited material, generating higher temperatures. This enhances the radiation emitted by the molten pool, leading to a gradual increase in both the width and area of the molten pool. The most significant increase occurs when the power is raised from 1200 W to 1600 W, with the width increasing by 0.376 mm and the area by 1.242 mm². Between 1800 W and 2000 W, the molten pool also shows a noticeable growth due to the high power exceeding the material’s tolerance limit, causing irregular diffusion and distortion of the molten pool shape. Additionally, the heat-affected zone around the molten pool expands, affecting the material’s structure and properties.

Scanning speed refers to the speed at which the laser beam or energy source moves over the workpiece surface. This speed determines the irradiation time of the laser on the workpiece surface, thereby affecting the formation, size, and temperature distribution of the molten pool. To study the relationship between scanning speed and the geometric characteristics of the molten pool, six sets of experiments were designed with a laser power of 1600 W, a powder feed rate of 12.79 g/min, and scanning speeds of 210 mm/s, 240 mm/s, 270 mm/s, 300 mm/s, 330 mm/s, and 370 mm/s. Figure 10 shows the images and characteristic parameters of the molten pool at different scanning speeds. It can be seen that as the scanning speed increases, the molten pool width decreases from wide to narrow, and the area decreases from large to small, with the decreasing trend gradually slowing down. The most significant decrease occurs when the scanning speed is increased from 210 mm/min to 360 mm/min, with the width decreasing by 0.284 mm and the area by 1.326 mm². The reason is that with constant laser power, increasing the scanning speed shortens the laser interaction time, reducing the amount of powder and substrate that can be melted, thereby decreasing the size of the molten pool.

3.2. Process Control of Molten Pool Morphology

Given the significant sensitivity of molten pool width to thermal parameters and the observed decisive influence of laser power on molten pool width in previous single-factor experiments, this study defines laser power as the control parameter and molten pool width as the output parameter in the real-time process control system. To systematically evaluate the real-time adjustment capabilities of this control system, we meticulously designed and implemented a single-track multi-layer experiment with a 10-layer structure following a U-shaped path. The linear part of the U-shaped part is 15 mm long, the radius of the arc is 15 mm, the height is 9 mm, and the thickness is 3.1 mm. This experimental design cleverly integrates both straight and curved paths to comprehensively validate the system’s real-time responsiveness in controlling molten pool width. During the experiment, two modes were compared: constant power mode and width control mode. The initial conditions for the experiment were set with a laser power of 1600 W, a powder feed rate of 12.79 g/min, and a scanning speed of 240 mm/s. Figure 11 presents the morphology of the deposited U-shaped workpiece under both modes, with 21 key coordinate points marked to provide a precise reference framework for subsequent height and width measurements.

As shown in Figure 11, it can be observed that in constant power mode, the edges of the deposition layers exhibit irregular shapes, and even spheroidization phenomena occur. In contrast, in width control mode, the width distribution of each layer of the workpiece is more uniform, and no interlayer width inconsistencies are observed, as seen in constant power mode. However, both modes present certain issues: the height at the beginning and end of the deposition layers is uneven, which could be due to excessive heat accumulation at the ends during the reciprocating scanning process and instability in powder feed at the start of the DED-LB process, leading to a rapid increase and deformation of the molten pool width. To further analyze the characteristics of the deposition layers under both modes, we cut the U-shaped workpiece at coordinate points 3, 7, 11, 15, and 19 and compared the height and width of the deposition layers in both modes. Figure 12 shows a comparative diagram of the cross-sectional heights of the deposition layers under the two modes. Under the constant power mode, the workpiece height is 7.565 mm, which is 1.435 mm less than the standard dimension. Under the width control mode, the workpiece height is 8.487 mm, 0.513 mm less than the standard dimension. Both are below the designed height of the workpiece, but the height under the width control mode shows a smaller deviation from the design value. Additionally, the fluctuation in height per layer is more significant under the constant power mode, while under the width control mode, the height fluctuation exhibits a consistent increasing trend.

In the continuous DED-LB process, the deposition layers of the first nine layers underwent laser remelting, leading to an increase in the width of the deposition layers and the difference in molten pool width. However, the tenth layer was not affected by this, as shown in Figure 13, which illustrates the width variation in the ten-layer deposition under both modes. The width variation range of the tenth deposition layer under the constant power mode is 0.382 mm, whereas under the width control mode, it is 0.247 mm. When the laser power was constant, the initial width variation in the molten pool remained within the range of 0.22 mm to 0.34 mm, with minor fluctuations. However, as the number of deposition layers increased, the fluctuation range of the molten pool width gradually expanded. Specifically, in the eighth and ninth layers, the difference between the widest and narrowest parts of the molten pool reached 0.45 mm and 0.52 mm, respectively. These large fluctuations can be attributed to the base material’s absorption and conduction of heat, causing instability at the molten pool boundary, which in turn affected the uniformity of the deposition layers. In contrast, under the width control mode, the overall fluctuation range of the molten pool width was significantly smaller compared to the constant power mode, and the width variation trend of each layer was more consistent. During the initial stage of deposition, the fluctuation range of the molten pool width was between 0.19 mm and 0.3 mm, which was not significantly different from the constant power mode. However, in the middle and later stages of deposition, the fluctuation range of the molten pool width increased significantly, especially in the eighth layer, where the maximum and minimum width difference reached 0.38 mm, while the fluctuation range in the ninth layer decreased to 0.19 mm. This could be due to the system automatically adjusting the laser power when it detected that the molten pool width exceeded the predicted value, thereby adapting to the high temperature and high absorption rate of the base material, gradually reducing the laser energy input, avoiding excessive melting, and stabilizing the molten pool width.

It is noteworthy that the width of the tenth layer was relatively narrower compared to the eighth and ninth layers because the tenth layer did not undergo the laser remelting step, resulting in a smaller width. However, its height was relatively larger compared to the remelted layers, which can be clearly seen from the height variation trend in Figure 13. Additionally, at the U-shaped bend, due to the larger curvature, the width fluctuation of the deposition layer under the width control mode was relatively small, maintaining an acceptable error range with the expected value. These comparisons demonstrate that the real-time control system can effectively maintain the stability of the processing process and improve processing accuracy.

4. Conclusions

This paper presents a study on molten pool image processing and feature extraction based on a laser-based direct energy deposition (DED-LB) coaxial vision detection system. We investigated the effects of different process parameters on the width and area of the molten pool and dynamically adjusted the laser power through a real-time feedback control system to achieve precise control of the molten pool width effectively.

(1) For the characteristics of coaxial molten pool images, a molten pool image monitoring algorithm was designed. Grayscale transformation and median filtering were used for noise reduction. To address the issues of the plume and heat-affected zones being mistakenly segmented into the molten pool area, a K-means algorithm combined with an iterative threshold segmentation method was proposed for accurate segmentation of the molten pool contour. Finally, the Canny algorithm was used to extract the molten pool edge information, and mathematical methods were used to extract the features of the molten pool.

(2) Variations in laser process parameters significantly impact the molten pool characteristics. An increase in laser power leads to an increase in the area and width of the molten pool. However, when the power is too high, irregular diffusion and distortion of the molten pool shape can occur. Increasing the scanning speed reduces the molten pool size, which is related to the shortened laser interaction time.

(3) A real-time feedback control system was implemented to dynamically adjust the mol-ten pool width during the laser-based direct energy deposition process. The experi-mental results show that the width of the deposited layer fluctuates from 0.22 mm to 0.52 mm under constant power mode, and from 0.19 mm to 0.38 mm under width control mode. The deposited layer with width control mode has smaller width fluctuation and higher machining accuracy. However, it has certain limitations in controlling the width difference between layers.

In summary, the proposed real-time monitoring method for the molten pool based on coaxial vision provides an effective technical means for stability control and quality assurance in the laser-based direct energy deposition process. This approach is significant for promoting the further development of DED-LB technology in industrial applications. Future work will focus on further optimizing algorithm performance and improving system response speed to meet the needs of more efficient and precise industrial production.