When the titanium alloy compressor blades of aircraft engines are working, they will deform, dent, wear, crack or even break due to long-term high-intensity service and foreign body damage. Laser cladding technology has become one of the important methods for blade repair due to its small heat-affected zone, good deposition performance, high forming accuracy and automation. The geometric characteristics of the molten pool are the key factors affecting the quality of cladding. Therefore, this paper proposes a recognition and measurement algorithm based on image processing for real-time monitoring of the molten pool. First, the ROI area is extracted by image mask, and then the ROI area is gamma transformed and threshold binarized to achieve the segmentation of the molten pool area; then the contour area features are calculated for denoising; finally, the AABB bounding box is used to extract the geometric features of the molten pool, and the length and width of the molten pool are monitored in real time during the cladding process. Finally, through multi-parameter orthogonal experiments, the average recognition error of the algorithm is verified to be 0.24 mm.

As the core parts of the engine, the compressor blades of aircraft engines undertake the important work of compressing air to provide high-pressure air for the combustion chamber. During the service of the aircraft, the blades work in an extreme working environment of high speed and high intensity for a long time. As the thrust-to-weight ratio of aircraft engines continues to increase, the pressure on the compressor increases, which directly increases the probability of blade deformation, cracking, or even fracture [1]. Therefore, in order to ensure the normal operation of the engine, it is necessary to regularly inspect the blades and replace or repair damaged blades. The maximum operating temperature of TC17 titanium alloy is 427 °C. It has the characteristics of high strength and good toughness and is widely used in the core components of aircraft engine compressors [2]. Statistics show that new blades need to be inspected after 3500 hours of work. The cost of blade replacement is 5 times that of blade repair, while the repaired blade can continue to work for 3000 hours. Therefore, repairing damaged blades is a more economical method [3].

At present, the main methods for surface repair of aviation parts include argon arc welding, laser cladding, electron beam welding, linear friction welding, etc. [4]. Argon arc welding has low forming accuracy and a large heat-affected zone, which makes it difficult to meet the surface accuracy and high strength requirements of aircraft engine blades. Laser cladding technology uses a small laser spot to accurately control heat input and is suitable for repairing various parts with complex structures. It has the characteristics of dense structure, small deformation, good flexibility, and easy integration. It is one of the trends for high-quality and high-efficiency repair of aircraft engine blades [5]. Real-time monitoring of the characteristics of the laser cladding process is a key way to intelligentize laser cladding [6–8]. In modern manufacturing technology, product quality control is often achieved through real-time monitoring of the production process, rather than testing and compensating the processed products one by one [9]. Therefore, in order to ensure the quality of laser cladding and give full play to its technical advantages, it is necessary to monitor the characteristics of the laser cladding process.

In recent years, with the growing demand for real-time monitoring of laser cladding and the development of information acquisition technology, research on it has continued to develop at home and abroad. Thompson et al. [10] proposed a laser deposition monitoring system based on beam coaxial imaging, which collected images through a camera and a narrowband filter to explore the effect of laser power on the laser deposition process. Wirth et al. [11] used a high-speed camera to monitor the motion trajectory of particles on the surface of the molten pool during laser cladding, and proposed that the particle flow trend is closely related to the process parameters. Gu Zhenjie et al. [12] developed a real-time monitoring system for melt pool spectra. The spectrometer fixed on the laser emitter collected spectral signals and studied the influence of plasma on laser energy transmission and cladding forming quality. Muvvala et al. [13] used a single-point monochromatic thermometer to monitor the thermal cycle process during cladding online and found that slow thermal cycles would lead to a decrease in the strength of the cladding body.

Laser cladding melt pool refers to the area of ​​molten liquid metal formed on the substrate by heating the metal powder fed synchronously to melt through the energy of the laser [14]. The quality of the melt pool is an important factor between the cladding quality and process variables. Real-time monitoring of the changes in the geometric characteristics of the melt pool is of great significance for cladding process analysis [15]. However, since the melt pool in the laser cladding process is moving and dynamically changing during melting, and there is high-brightness thermal radiation and a large amount of splashing powder, it is difficult to accurately obtain the shape and size of the melt pool in real time [16–17]. Therefore, the use of high-speed image acquisition technology and image processing algorithms to monitor the geometric shape and size of the melt pool in real time has become a hot topic in applied research. This paper takes TC17 titanium alloy as the research object and proposes a real-time monitoring method of molten pool geometry based on image processing. According to the characteristics of the highlight of the molten pool image, the contrast is enhanced by nonlinear transformation, and then the molten pool area is extracted by image segmentation, image denoising and other algorithms. Finally, the length and width of the molten pool are monitored in real time through the AABB bounding box to achieve the goal of real-time monitoring of the molten pool during the laser cladding process.

1 Experiment and method

1.1 Experimental materials and equipment

This paper carries out laser repair experiments on the lateral powder feeding laser melting deposition platform, as shown in Figure 1. The platform includes a DPSF powder feeder, an ABB robot and a HCFS-3000 laser. The laser is connected to the ZKSX-100TC equipment of Zhongke Sixiang through optical fiber, and the laser head is installed on the robot through the adapter tooling. The maximum power of the laser is 2000 W. The laser is integrated in the main chassis of the ABB robot, and the switch of the laser is controlled by the robot program. The experimental substrate uses TC17 titanium alloy. The experimental process is shown in Figure 2. Titanium alloy powder and protective gas are transported to the substrate along the lateral powder feeding tube, the laser scans along the top surface of the substrate, the CCD camera collects coaxial data of the molten pool through the reflection platform in the laser, and the data processing platform processes the image frame.

1.2 Image processing algorithm flow

Due to the high grayscale value of the image during laser cladding and the noise formed by powder splashing, it is necessary to design a suitable image processing algorithm to extract the characteristic information of the molten pool. This study takes TC17 titanium alloy material as the research object. For the video stream of the cladding area collected by the CCD camera, the geometric features of the molten pool are identified with the help of the OpenCV open source computer vision library. The algorithm flow is as follows.

(1) Obtain the ROI area through the mask.

During the laser cladding process, the industrial grayscale camera uses an optical path coaxial with the cladding laser to observe the cladding area. In order to avoid interference from irrelevant areas of the image and improve the recognition accuracy and speed, this paper selects a 480×640 pixel mask to perform a logical “AND” operation on the image to extract the melt pool and its surrounding images.

(2) Image enhancement based on gamma transform.

Since the grayscale value difference between the melt pool and the surrounding area is small, the robustness and effect of threshold segmentation will be poor. Therefore, it is necessary to enhance the melt pool image to improve the contrast of the melt pool area. Basic image enhancement methods are mainly based on linear transformation or nonlinear transformation. Linear enhancement can globally increase or decrease the overall grayscale of the image, but it cannot locally enhance the target object according to the spatial distribution of the image grayscale. Therefore, this paper uses a nonlinear image enhancement algorithm to enhance the highlight melt pool area. Gamma transform is a simple and effective nonlinear image enhancement algorithm that can enhance the contrast of scenes that are too dark or too bright [18–19]. For the input grayscale image, it is first normalized, the grayscale value of each pixel is divided by 255, and then the grayscale value of each pixel is gamma transformed. The mathematical expression of gamma transform is: O(r, c) = I(r, c)’γ, 0 ≤ r<H, 0 ≤ c ≤ W1 (1)
Wherein, H and W1 are the height and width of the image; r and c are the number of rows and columns of the image; O is the output image; I is the input image; γ is the parameter of gamma transform. When 0< γ <1, the gamma transform can enhance the contrast and make the darker ROI area more obvious; when γ = 0, the image does not change; when γ >1, the brighter ROI area can be extracted by reducing the contrast [20].

The ROI area in this paper is the bright melt pool area, as shown in Figure 3. In the grayscale histogram, the red framed area with a larger grayscale value is shown. Experiments were performed with γ = 1, 2, and 3. It can be found that as γ increases, the grayscale value of the red framed area becomes larger than that of other pixels, that is, the length of the blue arrow increases, and the melt pool becomes more obvious relative to the background and easier to segment. Through experiments, this paper finally selected γ = 3 as the parameter of the gamma transform.

(3) Melt pool extraction based on threshold segmentation.

Threshold segmentation is a common image processing algorithm. It is widely used in image segmentation scenarios due to its simple structure and stable performance [21]. It works on grayscale images and is suitable for situations where there is a significant difference in grayscale values ​​between the segmented target and the background. Its basic principle is: by setting a grayscale threshold, the pixels of the entire image are divided into two categories, and the pixels with grayscale greater than the threshold are set to white, and the areas with grayscale less than the threshold are set to black [22].

In this paper, the grayscale difference between the molten pool area and the background area is increased by gamma transformation, so that the threshold selection range is larger. According to Figure 3 (c), it can be observed that the grayscale values ​​of the molten pool area are concentrated between 200 and 225; the grayscale values ​​of the non-molten pool area are concentrated between 0 and 150. Therefore, this paper sets the thresholds to 150, 175, and 200 respectively. The threshold segmentation effects are shown in Figure 4 and Table 1. When the threshold is selected at 150 and 175, there is an under-segmentation phenomenon, the highlight powder or background cannot be completely segmented, and the molten pool recognition accuracy is less than 90%; when the threshold is selected at 200, the contour of the molten pool area can be relatively completely segmented, and the molten pool accuracy reaches 96.8%. Therefore, the threshold segmentation parameter of this study is selected as 200.

(4) Point cloud denoising based on the contour area of ​​the connected domain.

In the laser cladding process, in addition to the molten pool, there will also be high-heat and high-brightness non-molten pool areas left on the substrate. Therefore, there may be redundant discrete non-molten pool areas in the segmented binary image, which will interfere with the extraction of the geometric features of the molten pool. This paper screens the molten pool area and the non-molten pool area by finding the contour of each connected domain and calculating its area.

Contour extraction adopts the idea of ​​encoding to assign different values ​​to boundaries belonging to different levels. The specific idea is as follows: First, traverse each row of the image, f (i, j) represents the pixel value of the i-th row and j-th column of the image, and terminate when the pixel value meets one of the following conditions.
a. f (i, j–1) = 0, f (i, j) = 1, then define f (i, j) as the starting point of the outer boundary;
b. f (i, j) = 1, f (i, j + 1) = 0, then define f (i, j) as the starting point of the hole boundary.

Then, starting from the starting point, mark the elements on the boundary, with the initial identifier NBD = 1, and NBD plus 1 every time it touches a new boundary; if f(i, j) = 1, f(i, j+1) = 0, then f(i, j) is defined as –NBD, that is, the boundary end point. After determining the image hierarchy, the contour area is calculated, and this is used as a judgment condition to remove the contours with smaller areas.

The main factor affecting the size of the connected domain area is the linear energy density, that is, the amount of energy absorbed per unit length per unit time. This paper designs 27 groups of experiments based on the three variables of laser power, scanning speed and powder feeding speed that can be controlled in the experimental environment in Table 2 to count the size of the connected domain area of ​​the molten pool. The results of the melt pool pixel area are shown in Figure 5, where the minimum melt pool area is 410 pixel’2
, the maximum area is 494 pixel’2, and the average area is 454 pixel’2. The average area of ​​the irrelevant non-melt pool area is 100 pixel’2. Therefore, this paper takes 400 pixel’2 as the area critical value and only retains the connected domain contours with an area greater than 400 pixel’2 (Figure 6).

(5) Melt pool size extraction.

The geometric characteristic size parameters of the melt pool are shown in Figure 7. The melt pool is elliptical as a whole. Among them, x is the melt pool scanning direction; L is the length of the melt pool; W2 is the width of the melt pool.

This study extracts the melt pool geometric information by obtaining the AABB bounding moment of the melt pool contour. The bounding box is an algorithm for extracting the distribution range of the white area of ​​a binary image. Its basic principle is to use a simple geometric body to fit the range of the target object. As shown in Figure 8, the AABB bounding box is the minimum rectangle of the vertical boundary of the contour, and the side length is parallel to the upper and lower boundaries of the image.

It can be found that the width of the molten pool can be directly obtained from the height of the bounding box. As shown in Figure 9, due to the interference of strong light reflection and dust, one side of the tail of the molten pool sometimes has a false detection phenomenon similar to tailing, so the length of the molten pool cannot be simply obtained by the width of the bounding box. In this paper, the distance between the vertical midpoints A and B of the measured contour bounding box is taken as the length of the molten pool.

2 Experimental verification

The density and surface quality of laser cladding are closely related to the state of the molten pool, and the process parameters determine the geometric shape and fluctuation amplitude of the molten pool area. In order to more comprehensively verify the accuracy of the molten pool geometry monitoring algorithm based on TC17 titanium alloy, the experiment explored the recognition errors of the molten pool length and width under different process parameters.

The process parameters that dominate the changes in the molten pool morphology are scanning speed, powder feeding speed, laser power, etc. This paper uses the above-mentioned molten pool recognition and monitoring algorithm, and designs 3×3×3 groups of orthogonal experiments based on the experimental parameters in Table 3 to verify the accuracy of the molten pool recognition algorithm, and analyzes the influence of process parameters on the recognition accuracy of the molten pool length and width features. This paper verifies the accuracy of the algorithm by comparing the algorithm recognition value with the actual measurement value. As shown in Figure 10, the average width of the molten pool can be obtained by taking the average value of the molten channel width by using a vernier caliper for multiple measurements, while the length of the molten pool during laser cladding cannot be measured due to the continuous stacking of cladding materials. Therefore, this paper takes the length of the circular spot at the end of the melt path as the measured value of the melt pool length, and takes the melt pool length at the end of the melt path obtained by the algorithm as the recognition result.

Figure 11 shows the recognition screenshots of the real-time monitoring algorithm of the melt pool during the laser cladding process under different process parameters (the laser power, scanning speed and powder feeding speed are marked in the lower right corner of the image). Table 4 records the average width of the melt pool obtained by the recognition algorithm under different process parameters and the melt pool length at the end of the cladding process, and uses the actual measured data as a comparison.

Analysis of Table 4 shows that within the control range of the experimental variables, laser power is the main factor affecting the recognition accuracy of the melt pool width and length. Laser is the main energy source of the laser cladding process. The laser power is positively correlated with the energy density. A larger laser power will cause the melt pool to emit stronger and more unstable light, which will directly lead to a decrease in the recognition accuracy of the melt pool length and width. In the experiment, the average error of the molten pool when the laser power is 500 W, 1000 W, and 1500 W is 0.12 mm, 0.26 mm, and 0.36 mm respectively; the scanning speed is a secondary factor affecting the accuracy of molten pool recognition. With the increase of scanning speed, the movement speed of the molten pool on the substrate increases, and the stability of the molten pool decreases, resulting in a decrease in the accuracy of molten pool recognition. In the experiment, the average error of the molten pool when the scanning speed is 5 mm/s, 10 mm/s, and 15 mm/s is 0.22 mm, 0.26 mm, and 0.28 mm respectively; and the powder feeding speed has no direct effect on the accuracy of molten pool recognition.

The experimental results show that the overall average error of the molten pool geometry monitoring algorithm proposed in this paper is 0.24 mm, the maximum error is 0.48 mm, the minimum error is 0.06 mm, and the recognition speed is 0.04 s/frame, which can achieve real-time target monitoring.

3 Conclusions

This study aims at laser cladding TC17 titanium alloy materials, and proposes a molten pool geometric feature recognition algorithm based on image processing. The performance of the algorithm is verified and analyzed under different experimental parameter environments, and the following conclusions are drawn.

(1) The laser cladding coaxial monitoring algorithm extracts the image near the molten pool through the image mask, performs nonlinear transformation to improve the contrast, and binarizes the molten pool image. Noise removal is achieved based on the connected domain area characteristics of the splash powder and the molten pool, and the image and length and width values ​​of the molten pool itself are obtained.

(2) In view of the problem that the overall grayscale value of the molten pool image area is high and the visual scene distinction of the molten pool area is low, the gamma transform with γ = 3 can make the grayscale value of the molten pool area easier to separate. Combined with the binary segmentation with a threshold of 200, the molten pool area in the image can be extracted more completely.

(3) Within the test parameter range of TC17 titanium alloy, the area of ​​the connected domain of the molten pool is above 400 pixel’2, and the area of ​​the connected domain of the non-molten pool is about 100 pixel’2. The area of ​​the connected domain can be used as a judgment condition to realize the distinguishing and identifying features of the non-molten pool area and achieve the denoising function.

(4) The algorithm error range of the coaxial CCD monitoring of the length and width of the molten pool at the end of the melt channel is 0.06~0.48 mm; the average error is 0.24 mm; the recognition speed reaches 0.04 s/frame, which can meet the accuracy and real-time requirements of molten pool monitoring.