Weld Seam Recognition and Tracking With Omnidirectional Wall-Climbing Robot for Spherical Tank Inspection

Spherical tanks are commonly used to store chemical liquids and liquefied petroleum natural gas in the petroleum, chemical, and power industries. Multiple metal arc plates are used for welding spherical tanks in large-scale pressure vessel equipment. The weld seams between these metal plates may develop some internal defects, such as cracks, pores, and slag inclusions after a long period of service. Regular inspection and maintenance of weld seams on spherical tank surfaces can ensure safe production and operation. Compared with manual work, wall-climbing robots can adsorb and climb on wall surfaces, improving inspection efficiency and reducing maintenance costs. However, it is difficult for inspection robots to automatically identify and track weld seams because of the large space size of spherical tanks and the harsh operating environment.

Wall-climbing robots using various adsorption methods, such as negative pressure adsorption [1–4], magnetic adsorption [5–9], thrust adsorption [10–12], and bionic adsorption [13–15], have been developed and deployed. Metal walls are typically used for wall-climbing robots with permanent magnet adsorption. Permanent magnet adsorption can provide a greater adsorption force and no energy consumption than other methods. Magnetic adsorption wall-climbing robots have significant advantages in the application of spherical tank inspection. However, unlike climbing on planar walls, spherical tanks of different diameters have different surface curvatures. Wall-climbing robots for spherical tank inspection must adapt to curved working surfaces to stably adsorb and climb, which requires more reliable and flexible adaptation mechanisms.

Weld seams are the detection target, and the accurate weld seam recognition by the wall-climbing robot is the premise of automatic spherical tank inspection. Several weld identification methods have been tested and applied, such as weld line localization method using a cross-structured light device [16,17], an improved Canny algorithm to detect the edges of seam and extract the characteristic parameters of welding images [18], a seam tracking system based on dual cameras with auxiliary LED lights to present real-time image processing algorithms [19], and a novel intelligent system for the automatic visual inspection of vessels based on an ensemble of classifiers [20]. Because of the painted surface of spherical tanks and various disturbances, such as rust, corrosion, smudges, and scratches, it is difficult to accurately identify weld seams only by image processing algorithms. Deep learning methods as emerging identification technologies based on convolutional neural networks (CNNs) can achieve faster and more accurate target detection and classification. Some welding identification methods based on deep learning have also been developed and studied [21,22]. During the wall-climbing robot inspection on the curved surface of the spherical tanks, weld seam recognition and weld path tracking are indispensable capabilities of the robot, which contribute to more automated navigation operations and improve the efficiency of inspection and maintenance.

In our earlier studies, a series of wall-climbing robots have been developed and applied [23,24], which can effectively climb on tank surfaces. Because the surface of the spherical tanks is spray-painted and may be covered with corrosion and stains, it is difficult to accurately judge the weld path using conventional image processing methods. This study presents a weld seam recognition and tracking method for the designed omnidirectional wall-climbing robot. The wall-climbing robot with an adjustable frame and suspension mechanisms can adequately absorb curved surfaces with different diameters. Mecanum wheels enable the robot to climb omnidirectionally on spherical surfaces. Based on deep learning networks, the robot can quickly identify weld seams and fit weld paths for tracking using a front-mounted camera. A complete weld seam tracking control model of the robot was proposed and simulated based on the robot’s kinematic model and sliding mode control. Multiple experiments on a cylindrical tank were performed to test the efficiency and accuracy of weld seam recognition and tracking by the robot. Field tests on a 3000 m³ spherical tank were conducted to verify the feasibility and practicality of the seam recognition and tracking robotic system.

Wall-climbing robots can be used to carry inspection equipment or maintenance tools to replace manual work on spherical tanks. These spherical tanks are welded by multiple arc-shaped plates, so the contact surfaces of wall-climbing robots are curved or spherical, and the weld seams are also distributed in the form of spatial curves, which puts higher requirements on climbing adaptability and flexibility of the robot. In this study, to achieve stable adsorption and flexible movement on the arc-shaped wall of the spherical tanks, an omnidirectional wall-climbing robot was designed.

The schematic of the developed wall-climbing robot for the weld seam inspection on spherical tanks is shown in Fig. 1(a). The omnidirectional wall-climbing robot includes the following components (Fig. 1(b)): four Mecanum wheels, four drive motors, an adjustable robot frame, four permanent magnets, four shock-absorbing suspension mechanisms, as well as robot control computer.

The driving form of four Mecanum wheels enables the wall-climbing robot to move in all directions on tank surfaces, and the robot can adjust its position and speed more flexibly. When climbing on spherical tanks, which is more difficult than flat metal walls, the four wheels of the robot must adjust their inclination angle to adapt to spherical surfaces; otherwise, stable adsorption cannot be guaranteed. The adjustable robot frame can change the inclination angle of four Mecanum wheels to ensure full contact with spherical tank surfaces of different diameters. Four permanent magnets are located near the Mecanum wheels to provide adsorption force, and their positions change with the wheel's inclination angles. When surmounting obstacles or weld seams, four independent shock-absorbing suspension mechanisms keep the robot stable. The performance parameters of the omnidirectional wall-climbing robot are as follows: the outer dimension is 466 × 520 × 498 mm, the mass is 13.75 kg, the maximum payload is 10 kg, and the maximum climbing velocity is 0.23 m/s. Robot force analysis and climbing tests indicate that the robot can stably adsorb and climb on spherical surfaces with different diameters [25].

A camera is installed at the front of the robot to capture weld seam images on tank surfaces. The position parameters of the Mecanum wheels and camera are shown in Fig. 2. The Mecanum wheel radius r is 63.5 mm, the distance between the front and rear wheels $Lw$ is 400 mm, and the distance between the left and right wheels $Ww$ is 300 mm. $hc$ and $lc$ represent the installation height and distance of the camera, and their values are 200 and 100 mm, respectively. These robot design parameters determine the relative geometric relationship between the captured weld seam image and the center of the robot and are used to calculate the deviations during robot tracking weld paths.

Figure 3 shows the weld recognition and robot control system. The entire robotic system includes an omnidirectional wall-climbing robot, a wireless router, and a remote computer. The electronic hardware of the wall-climbing robot consists of a robot control computer, a camera, a central controller, and encoders. Real-time images of weld seams are captured using the camera. The robot control computer is used to process these images and identify weld seams. A NVIDIA RTX2060 graphics processing unit (GPU) is equipped to accelerate deep learning network learning and weld seam recognition. The main functions of the robot control computer include image processing, weld seam recognition, robot position estimation, robot motion control, and data communication. The central controller is used to control the rotation speed of four motors and calculate the robot’s running speed through four encoders.

Workers can use the remote control unit and remote computer to send commands and control the motion of the omnidirectional wall-climbing robot. The remote computer communicates with the robot through the wireless router. The remote computer can display real-time weld seam paths, motion states, and the spatial position of the robot.

To realize automated weld seam inspection by the wall-climbing robot, weld seam recognition and path extraction are one of the most important tasks. After capturing the weld seam images, the robot needs to complete weld seam recognition, weld path fitting, and path deviation calculation before it performs autonomous weld tracking and inspection operations.

After a long period of outdoor work, the surface of spherical tanks is covered in dirt and rust, affecting the recognition accuracy of weld seams. Weld seam images on the tank surface captured by the wall-climbing robot are shown in Fig. 4. Weld seams on tank surfaces are classified as single transverse, single longitudinal, crossed, and T-shaped weld seams based on the characteristics of distribution locations. Different categories make fitting weld paths for tracking more difficult.

Accurate identification of weld paths is beneficial to the automatic inspection and maintenance of wall-climbing robots. However, weld seam identification and path line extraction with high precision have always been arduous tasks. It is difficult to obtain intact seam path trajectories only by image processing because of the small distinction between weld seams and surroundings. Furthermore, the recognition method of image processing is very sensitive to illumination changes and image disturbances. Even if some features are obtained, the path information is mostly discontinuous, unclear, and distorted.

With the development of artificial intelligence in recent years, deep learning methods have gradually become a feasible solution to achieve more accurate distinction of weld objects through training and learning. Deep learning is generally composed of multilayer CNNs, which can classify objects more accurately after learning and training. The instance segmentation algorithms can achieve pixel-level classification and segmentation, which provides new means for weld seam recognition and extraction.

Figure 5 shows the basic process of weld seam recognizing and weld path fitting. First, weld seam images were captured by the camera mounted on the wall-climbing robot. Mask R-CNN [26] was chosen to identify weld seams from captured images and perform instance segmentation. Compared with PA-Net [27], RetinaMask [28], FCIS [29], and Mask R-CNN have excellent processing speed (116.3 ms) and mask mean average precision (35.7%). Some deep learning networks (such as YOLACT [30,31]) have faster recognition speed, but the mask mean average precision still does not exceed Mask R-CNN. Subsequently, weld paths were extracted and fitted using image processing and the least squares method.

Mask R-CNN is a deep learning network based on CNNs. It can accomplish the instance segmentation of targets in images. Mask R-CNN's weld seam recognition process consists of feature extraction, candidate region processing, RoIAlign layer, classification, regression, and mask generation. First, a series of CNN layers (such as ResNet50 and ResNet101) combined with feature pyramid networks [32], as backbone networks, was used to extract different depth feature maps from weld seam images. The region proposal network can generate regions of interest (RoI). The corresponding features of each RoI on the feature maps were extracted by RoIAlign with bilinear interpolation to avoid deviations.

However, after the RoIAlign layer, RoIs are fed into two fully connected layers for image classification and bounding box regression to determine categories and boxes of the objects. Meanwhile, a mask prediction branch uses fully convolutional networks [33] to segment objects in the image in pixels, and predicts m × m masks from each RoI.

The wall-climbing robot has collected 1967 images of weld seams on the surface of a tank, with a resolution of 320 × 240 pixels. Smaller sizes can reduce training and recognition time during the training and testing process, but it has little impact on weld path fitting. The contours of the weld seams in these acquired images are labeled and masked in advance. Data augmentation methods, such as Gaussian blur, geometric transformation, brightness adjustment, and noise addition, are used to expand the training dataset. After 20 times of data augmentation, the training dataset includes 39,340 weld seam images. The number of categories in the dataset was 2, representing background and weld seam.

The network model was trained on the robot control computer with NVIDIA GeForce RTX2060 for GPU acceleration. Its parameters were set as follows: The initial learning rate of the network was 0.001; the weight attenuation coefficient was 0.0005; the momentum coefficient was 0.8. After 30,000 iterations of training, the total loss of the network model is less than 0.06, and the average accuracy is greater than 0.98. As shown in Fig. 6, the recognition results of weld seams include classification probabilities, regression boxes, and pixel-level masks. Based on the deep learning network model, the weld seams can be classified and recognized for further weld path extraction.

After weld seams were recognized, the binarized weld seam images were obtained through image processing methods such as color distinction, binarization processing, and Gaussian filter, and then, the least squares method was used to fit the centerlines of weld paths. Two-position parameters of weld paths were estimated, such as $αp$ and $dp$⁠. Figure 7 shows the fitted weld path and path parameters. $αp$ is the drift angle of the identified weld path in the image and $dp$ is the transverse distance (pixel) from the weld seam to the center of the image. When the robot reaches an intersected weld point, two weld seams will be recognized. The robot tracks one weld path according to manual selection or path planning.

The line equation of a weld path is given as $h(x)=kx+b$⁠. When the quadratic sum $G(x)$ of distances from all weld points to a line is the smallest, the line is considered the centerline of the weld path.

Figure 8 shows the positional relationship between the tracked weld path and the wall-climbing robot. Point $Oc$ is the center of the robot, point $Om$ is the center of the camera, and the robot coordinate system $xOcy$ is established. According to the robot design parameters, the distance between point $Oc$ and point $Om$ is $lc+Lw/2$⁠. $αc$ and $dc$ are the drift angle and offset distance of the robot center relative to the tracked weld path.

The deviation conversion in Eq. (8) indicates that $αc$ and $αp$ are linearly dependent, and changes in $αc$ or $αp$ will cause changes in $dc$⁠. The image range captured by the camera is limited, which places higher requirements on the robot weld seam tracking algorithm. If the robot cannot quickly correct its position and angle, the camera may lose weld seam images, and the robot will fail to track the weld path. At the beginning of weld seam tracking, the tracked weld seam should be within the maximum camera range.

The weld recognition method based on deep learning can identify and extract the weld paths, and the arc-shaped welds on the surface of the spherical tank bring more difficulty to the robot's autonomous operation. Fast and stable path-tracking control is conducive to improving the robot's inspection efficiency and reducing energy consumption. Based on the omnidirectional motion form of the robot and the calculation of the weld path deviations, the control method of tracking weld paths directly affects the accuracy of the robot's automatic inspection operation.

Based on the adjustable robot frame and shock-absorbing elastic suspension mechanisms, the robot can climb flexibly on the surface of spherical tanks. Because the robot is much smaller than spherical tanks, the weld seam tracking motion on tank surfaces can be simplified as a plane motion. During the weld seam tracking process, the wall-climbing robot continuously corrects the drift angle and offset distance to ensure that the robot’s climbing direction coincides with the weld path direction.

Figure 9 shows the position correction analysis during the robot weld seams tracking process. The weld seam coordinate system $XOwY$ is established, and the tracked weld path is distributed along the Y coordinate axis. The starting point of the robot is $Oc$⁠, and the desired point is $Od$⁠. The distance between point $Oc$ and point $Ow$ is $dc$⁠, the angle between the robot's forward direction and the Y coordinate axis is $ac$⁠. In the coordinate system $XOwY$⁠, the current state vector of the robot is $qc=[−dc⋅sinac$$dc⋅cosacac]T$⁠. When tracking the weld seam, the robot should reach point $Od$ on the weld path (Y coordinate axis), and the next desired state vector is $qd=[xd00]T$⁠.

Figure 10 shows the control diagram of the weld seam tracking by the omnidirectional wall-climbing robot using sliding mode control. In the initial stage, after the weld path was identified, the robot calculates and converts the deviations of the weld path. The sliding mode controller is used to adjust the robot's posture and position, and the three velocities $(vx,vy,andωc)$ of the robot will be corrected simultaneously. The velocities of robot wheels are modified based on the robot kinematic model. During the wheel speed regulation process, four independent PID controllers are used to complete the closed-loop control of wheel speeds using the data recorded by four encoders.

$vyd$ and $ωd$ are linearly related to $dc$ and $ac$⁠, respectively. When there is no offset distance and drift angle between the robot and the tracked weld path, $vyd$ and $ωd$ are 0.

If $Qi≥0$ and $Pi≥0$⁠, V is negative semidefinite. According to Barbalat's lemma, $S→0$ as $t→0$⁠, when $S→0$⁠, $ev=−K∫evdt$⁠, and $ev→0$ as $t→0$⁠.

Simulation experiments were performed to verify the correctness and feasibility of the weld path-tracking controller. During the weld path-tracking process, the maximum forward velocity and traverse velocity of the robot is 0.23 m/s, and the maximum rotational angular velocity is 0.94 rad/s. The simulation parameters are set as follows: the initial drift angle $αp$ in the weld image is $π/6$ rad, the initial offset distance $dp$ is 80 pixels, and the given tracking velocity $vxd$ is 0.1 m/s. The parameters of the sliding mode controller are as follows: $kω=0.1$⁠, $ky=0.8$⁠, $Q1=Q3=0.5$⁠, $Q2=0.75$⁠, $P1=P2=P3=0.01$⁠, and $k1=k2=k3=1$⁠.

The simulation results of the robot tracking weld path are shown in Fig. 11. The drift angle and offset distance of the robot can be quickly corrected, the robot state errors $ye$ and $ae$ are also corrected accordingly, and the robot tracking error $xe$ is kept at 0.1 m. The simulation results indicate that the designed controller can accurately adjust robot velocities based on the weld path deviations and the robot can successfully track the weld path. In subsequent experiments and field tests, the relevant parameters of the seam path-tracking controller will be optimized and adjusted to ensure that the actual wall-climbing robot stably tracks the weld seams on tank surfaces.

To verify the accuracy of weld seam recognition and tracking, we conducted experiments with the climbing robot on an experimental platform. The experiments mainly include weld recognition experiments and weld seam tracking experiments. In addition, field tests on a real spherical tank were also carried out to verify the feasibility and stability of the weld seam recognition and tracking control method as well as the developed robotic system in application.

The experimental platform, which is a cylindrical tank, was welded from multiple arc-shaped plates (Fig. 12). The cylindrical tank has a diameter of 4000 mm, a height of 2.8 m, and a wall thickness of 10 mm. Multiple longitudinal and transverse weld seams are distributed on the tank surface for robot recognition and tracking. The omnidirectional wall-climbing robot can stably adsorb and climb on the vertical curved surface of the cylindrical tank.

In the weld seam recognition experiments, the robot used the camera to capture weld seam images of the tank surface, recognized weld seams, and extracted weld paths based on deep learning networks. The robot's real-time recognition results of weld seams are shown in Fig. 13. The results indicate that the weld seam recognition method based on deep learning can accurately identify weld seams and is not easily disturbed by the surrounding environment, such as rust, stains, and scratches.

Figure 14 shows the real-time weld path extraction results, including original weld images and binary path images. The robot could extract the angle and offset distance of weld paths in real time, which were then used to feedback on the position of the robot relative to the weld seam. The total processing time of each image was kept between 0.22 and 0.25 s, including image loading loss time (0.006 s), deep learning identification time (0.21–0.23 s), and image processing time (0.018 s). The experimental results indicate that the robot can successfully identify and extract weld paths. The robotic system can obtain at least 4 sets of weld path data per second, which can meet the inspection requirements of the wall-climbing robot.

Weld seam tracking experiments were performed to verify the tracking accuracy of the wall-climbing robot. In the experiments, the robot adsorbed on the vertical arc-shaped wall of the experimental platform and automatically recognized and tracked the weld seams. The experiments included longitudinal weld seam tracking experiments and transverse weld seam tracking experiments. The given desired tracking velocity of the robot included ±0.05 m/s, ±0.1 m/s, ±0.15 m/s, and ±0.2 m/s. The real-time drift angle and offset distance of the robot relative to the tracked weld seam were recorded during weld seam tracking.

Figure 15 shows the process of longitudinal weld seam tracking experiments. In the beginning, it is necessary to ensure that the robot camera can capture images of weld seams. The initial drift angle and offset distance were random. The robot tracked a longitudinal weld seam upward at different velocities and then tracked downwards (without turning around) at the same velocity.

Figure 16 shows the experimental results of the robot longitudinal seam tracking at different velocities. The robot angle (drift angle) and offset distance fluctuate significantly during the initial stage of tracking weld seams, but they are quickly adjusted back. In subsequent stages, the robot can stably track weld seams, and the robot angle and offset distance are continuously corrected. During the upward tracking weld seams, the robot angle deviations at different velocities are kept within ±3 deg (Fig. 16(a)). As shown in Fig. 16(b), at low tracking velocities (0.05 and 0.1 m/s), the maximum offset distances of the robot are less than ±20 mm. The offset distances slightly increase as the tracking velocity increases (0.15 and 0.2 m/s), but it can still be maintained within ±30 mm. Just like upward tracking seams, the downward tracking seam performance of the robot is also excellent (Figs. 16(c) and 16(d)). In longitudinal seam tracking experiments, the robot could flexibly adjust its velocities and position to smoothly track the weld seams with high precision. The experimental results indicate that the wall-climbing robot with a designed weld seam tracking controller can accurately track longitudinal seams on vertical tank surfaces.

Figure 17 shows the experimental process of transverse weld seam tracking by the robot. The transverse weld seams are distributed in the middle of the experimental tank. Experimental results of robot transverse weld seam tracking at different velocities are shown in Fig. 18. The robot can correct deviations and track the weld seam with high precision for diverse initial angles and distances. Whether tracking the weld seam forward or backward, the drift angle of the robot can still be maintained within ±3 deg. When the robot tracking weld seam at 0.05 and 0.1 m/s, the offset distances are maintained within ±20 mm. The offset distance increases with the tracking velocity. For example, when the robot tracking velocity is 0.2 m/s, the maximum offset distance of the robot reaches −35 mm. The reason for this is that affected by gravity, the robot may slip slightly during transverse weld seams tracking, increasing the offset distance. The experimental results indicate that the maximum offset distance of transverse weld seams tracking at low velocity is ±20 mm, and the maximum offset distance at high velocity is ±35 mm. The weld seam tracking controller can enable the robot to track weld seams with high precision and correct the deviation caused by the sliding.

In the weld seam tracking experiments, based on the designed weld seam tracking controller, the wall-climbing robot could consistently track the longitudinal and transverse weld seams with high precision. After receiving the tracking command, the robot could quickly correct the random initial drift angle and offset distance. During the stable stage of weld seam tracking, the robot continuously adjusted its posture and corrected the deviations. The robot could still maintain stable tracking at a higher velocity, and the maximum drift angle and offset distance were maintained within ±3 deg and ±35 mm. The omnidirectional wall-climbing robot with high-precision weld seam tracking capability can improve the automatic inspection of spherical tanks.

Field tests were conducted to verify the adaptability and stability of automatic weld seam tracking for the omnidirectional wall-climbing robot in engineering applications. As shown in Fig. 19, the wall-climbing robot was tested on a real liquefied petroleum gas (LPG) spherical tank. It has a volume of 3000 m³, a diameter of 18 m, a maximum height of 20 m, and a thickness of 49 mm. Multiple weld seams are distributed on the surface of the spherical tank, the total length of weld seams is 455.46 m. Because of the large volume of industrial spherical tanks, the traditional manual inspection method will be expensive and time-consuming. Manual inspections often require a large amount of scaffolding to detect all weld seams, whereas the wall-climbing robot can flexibly reach any weld seam position for inspection.

The robot captured weld images of the spherical tank to expand the training dataset. Before the field tests, new 2172 weld images on the surface of the spherical tank were collected. After data augmentation, the training dataset was added with 43,440 weld images. On other different tanks, the training dataset of Mask R-CNN will continue to increase. Deep learning is a method of continuous learning. With more training and learning, the networks can recognize more types of weld seams. Presently, the robot has trained numerous weld images of different spherical tanks.

Figure 20 shows the real-time image results of the robot weld seam tracking on the spherical tank surface. Based on deep learning, the robot can accurately extract the weld paths on the spherical tank, the average recognition time of each weld image was approximately 0.25 s. Figure 21 shows the test process of robot tracking weld seams on the spherical tank. In the test, the robot tracked multiple longitudinal and transverse weld seams. The results indicate that the robot can quickly identify and extract weld paths with high precision, and automatically run along the weld seams. When the robot reaches one intersected weld point, the robot will automatically stop waiting for the control command, and workers can select the next tracked weld seam through the remote computer. If the robot fails to recognize or misidentifies weld seams, the robot will be switched to the remote control mode to resolve the issue.

The field tests verified the efficiency of the weld seam recognition and control method in engineering applications. The developed robot control system enhances the practicability of the wall-climbing robot and improves the automatic inspection of spherical tanks. Equipped with inspection and maintenance mechanisms, the wall-climbing robot can complete more inspection and maintenance tasks.

There remains a lack of effective implementation methods for tracking weld seams and automatic inspection operation by the robots in real spherical tanks. In this study, we proposed a weld seam recognition and weld path-tracking control method suitable for the developed omnidirectional wall-climbing robot to implement weld seam inspection on arc-shaped working surfaces of cylindrical tanks and spherical tanks. The designed omnidirectional wall-climbing robot could provide flexible movement forms for automated weld seam tracking, and the recognition method-based deep learning networks enabled more accurate extraction of welding seam paths. By analyzing the weld path deviations, the control method based on sliding mode control was proposed and applied to adjust the movement velocities of the robot during the weld seam tracking process. In the cylindrical tank experiments and field tests of the real spherical tank, the robot could complete the recognition and tracking of the weld seams, with a maximum drift angle of ±3 deg and a maximum offset distance of ±35 mm. During the process of the robot tracking the weld seams, due to the continuous adjustment of the running velocities and angular velocity, the tracking route fluctuated slightly, but it could be corrected in time. This weld seam recognition and weld path-tracking control method is conducive to improving the autonomous operation capabilities of the spherical tank inspection robots.

There may be some possible limitations in this study, the first limitation is that the weld seam training dataset may be insufficient, and the deep learning networks can be continuously optimized to improve the accuracy and speed of weld seam recognition and extraction. The second limitation concerns that more spherical tanks should be tested and implemented to improve the recognition accuracy of weld seams. More in-depth follow-up research and applications will be conducted to optimize the performance of robot tracking weld seams, including establishing more comprehensive weld seam datasets, optimizing deep learning models, and conducting more field tests.

In this study, a weld seam recognition and tracking control method with the developed omnidirectional wall-climbing robot for spherical tank inspection is proposed. The designed wall-climbing robot adopts Mecanum wheels to achieve omnidirectional movement. With the adjustable robot frame and shock-absorbing suspension mechanisms, the robot can be applied to spherical tanks of different diameters. The robot captures real-time images of weld seams through the front-mounted camera. It can be controlled and monitored by a remote computer. The robot can quickly recognize weld seams and accurately extract weld paths based on deep learning networks. Real-time drift angle and offset distance are used to feedback on the position of the robot relative to the tracked weld seam. Based on the kinematic model and sliding mode control, the robot can quickly correct deviations and adjust its position to track weld seams. In the weld seam recognition experiments, the wall-climbing robot could identify weld seams and extract weld paths in real time, and the processing time of each weld image was approximately 0.25 s. In the weld seam tracking experiments, at different velocities, the robot successfully and continuously tracked longitudinal and transverse weld seams in two directions. At low velocity, the robot could achieve more accurate weld seam tracking, and the maximum drift angle and offset distance were ±3 deg and ±20 mm. The experimental results indicate that the robot can recognize and track weld seams with high accuracy.

Furthermore, field tests on a 3000-m³ spherical tank were conducted to verify the practicability and reliability of the developed robotic system. High precision and fast weld seam recognition and tracking facilitate automated and intelligent inspection of the wall-climbing robot. More tests on different spherical tanks will be conducted in the future to optimize the performance of the robotic system. More weld seam images will be collected to extend the training dataset, which will benefit the recognition accuracy and application range of the inspection robot.

This project is supported by the Natural Science Foundation of Jiangsu Province, China (BK202300367), the Science and Technology Project of the State Administration for Market Regulation, China (2023MK040), the Open Project of Anhui Province Key Laboratory of Special and Heavy Load Robot (TZJQR002-2024), and the Open Project of China International Science and Technology Cooperation Base on Intelligent Equipment Manufacturing in Special Service Environment (ISTC2023KF06).

There are no conflicts of interest.

The datasets generated and supporting the findings of this article are obtainable from the corresponding author upon reasonable request.