Automatic deviation correction method and system for operation of underwater robot, medium and product

Abstract

The invention discloses an automatic deviation correction method and system for operation of an underwater robot, a medium and a product, and relates to the field of general control or adjustment systems.The method comprises the steps that a weld seam image is obtained, filtering parameters are adjusted according to the noise intensity of a local image in the weld seam image, and a noise reduction image is obtained; performing double-layer threshold segmentation on the noise-reduced image to obtain a binary image; setting a plurality of detection areas in the binarized image, determining an effective detection area, converting edge feature points in the effective detection area into coordinate data, and performing partition weighted fitting to obtain an initial weld direction; constructing a dynamic error band, taking the coordinate points outside the dynamic error band as candidate abnormal points, determining a final abnormal point in the candidate abnormal points in combination with continuity criteria of adjacent feature points, and re-fitting the remaining coordinate points to obtain a target weld direction; and the target welding seam direction is compared with the reference direction, and a graded deviation correction control signal is generated according to the deviation degree. According to the invention, the correction accuracy can be improved.

Description

Technical Field

[0001] The present application relates to the general field of control or regulation systems, and in particular to a method, system, medium and product for automatic deviation correction of underwater robot operations. Background Art

[0002] In the nuclear industry, spent fuel pools are critical facilities for the storage and disposal of spent fuel, and their structural integrity is directly related to nuclear safety. Spent fuel pools consist of a bottom and four side surfaces, and the welds between these surfaces are crucial for structural strength. To ensure the safe operation of the pools, regular weld inspection and maintenance are required. Because the pools are underwater and the welds are fixed in position, weld inspection requires underwater robots equipped with inspection equipment.

[0003] In related technologies, underwater robots use cameras to capture weld images, employ fixed-parameter image filtering to remove underwater noise, and then extract weld features using a global threshold segmentation method. After feature extraction, a simple edge detection algorithm is used to obtain weld contour points. Linear fitting is then used to calculate the weld direction, and the robot's motion is adjusted based on the calculated results.

[0004] However, in actual underwater environments, the robot's movement will cause water flow disturbances and the influence of suspended particles, and the imaging quality of the weld image shows obvious differences in different areas. The feature processing method in related technologies is easily affected by fluctuations in local imaging quality, resulting in large deviations in the calculation of the weld direction. Summary of the Invention

[0005] The present application provides a method, system, medium and product for automatic deviation correction of underwater robot operations, which are used to improve the accuracy of deviation correction.

[0006] In the first aspect, the present application provides an automatic deviation correction method for underwater robot operations, which is applied to a control system. The method comprises: obtaining a weld image captured by a camera of the underwater robot, adjusting a filter parameter according to the noise intensity of a local image in the weld image, and obtaining a denoised image; performing double-layer threshold segmentation on the denoised image to obtain a binary image; the first layer threshold of the double-layer threshold is used to extract the main contour of the weld, and the second layer threshold of the double-layer threshold is used to retain the edge transition area features of the weld; setting multiple detection areas in the binary image, independently extracting edge feature points in each detection area, and extracting edge feature points according to the number of feature points in each detection area. The effective detection area is determined by the quantity distribution, and the edge feature points in the effective detection area are converted into coordinate data; the weight coefficient is determined according to the distribution density of the edge feature points in the detection area, and the coordinate data is partitioned and weighted fitted to obtain the initial weld direction; a dynamic error band is constructed based on the initial weld direction, and the coordinate points outside the dynamic error band are used as candidate abnormal points. The final abnormal point among the candidate abnormal points is determined in combination with the continuity criterion of the adjacent feature points, and the remaining coordinate points outside the final abnormal point are re-fitted to obtain the target weld direction; the target weld direction is compared with the reference direction, and a graded correction control signal is generated according to the degree of deviation.

[0007] In the above embodiment, the control system adjusts the filtering parameters according to the local image noise intensity to improve the adaptability of image preprocessing, the double-layer threshold segmentation retains the weld body and edge transition features, the independent extraction of feature points in multiple detection areas avoids local interference, the weighted fitting based on distribution density improves the contribution of key areas, and the dynamic error band filters out abnormal points to enhance the stability of the fitting results, thereby achieving improved accuracy of underwater weld tracking.

[0008] In combination with some embodiments of the first aspect, in some embodiments, the step of performing double-layer threshold segmentation on the denoised image to obtain a binary image specifically includes: performing grayscale normalization processing on the denoised image to obtain a standard grayscale image; calculating the grayscale histogram and cumulative distribution function of the standard grayscale image to obtain grayscale distribution characteristics; determining the first layer threshold based on the grayscale distribution characteristics, performing initial segmentation to obtain the weld candidate area; calculating the second layer threshold according to the grayscale distribution of the weld candidate area, performing secondary segmentation to retain the edge transition area; marking the area above the first layer threshold as foreground pixel value, and marking the area below the second layer threshold as background pixel value to obtain a binary image.

[0009] In the above embodiment, the grayscale normalization processing of the control system eliminates the brightness differences between different areas. The grayscale histogram and cumulative distribution function reflect the overall feature distribution of the image. The first-layer threshold ensures the complete extraction of the weld body, and the second-layer threshold retains the detailed features of the edge transition area. The binarization processing of the foreground and background lays a reliable foundation for subsequent feature extraction and improves the segmentation accuracy of weld features in underwater environments.

[0010] In combination with some embodiments of the first aspect, in some embodiments, the step of determining the weight coefficient based on the distribution density of the edge feature points in the detection area, performing partitioned weighted fitting on the coordinate data, and obtaining the initial weld direction specifically includes: calculating the horizontal and vertical projection distribution of the edge feature points in each detection area to obtain a two-dimensional density distribution matrix; calculating the clustering degree of the feature points in each area based on the two-dimensional density distribution matrix, and determining the weight coefficient based on the clustering degree of the feature points; dividing the detection area into several sub-areas of equal length along the direction of the weld, and applying the corresponding weight coefficient to weight the feature point coordinates in each sub-area; performing least squares fitting on all weighted feature point coordinates to obtain the spatial equation of the weld, and determining the direction vector of the spatial equation as the initial weld direction.

[0011] In the above embodiment, the control system obtains a two-dimensional density matrix through projection distribution calculation to reflect the spatial distribution characteristics of the feature points, determines the weight coefficient based on the concentration of the feature points to highlight the contribution of the high-quality area, sub-area division and weighted processing balance the influence of different areas, and the least squares fitting ensures the accuracy of the spatial equation.

[0012] In combination with some embodiments of the first aspect, in some embodiments, after obtaining the weld image captured by the underwater robot camera, adjusting the filtering parameters according to the noise intensity of the local image in the weld image, and obtaining the noise reduction image, the method also includes: obtaining the posture data collected by the underwater robot posture sensor, and establishing a spatial transformation matrix between the image plane and the weld plane; performing perspective correction on the weld image according to the spatial transformation matrix to eliminate the perspective deviation; extracting the moving bubble trajectory in the weld image, establishing the bubble motion vector field, and determining the direction of water flow disturbance; adjusting the image sampling timing according to the direction of water flow disturbance to avoid the peak moment of water flow disturbance.

[0013] In the above embodiment, the control system's posture data and spatial transformation matrix achieve image perspective correction, bubble trajectory analysis establishes a water flow disturbance model, and the sampling timing is adjusted to avoid interference peaks, thereby improving image acquisition quality.

[0014] In combination with some embodiments of the first aspect, in some embodiments, the steps of extracting the trajectory of moving bubbles in the weld image, establishing a bubble motion vector field, and determining the direction of water flow disturbance specifically include: performing inter-frame difference operations on continuously collected multiple-frame images, extracting the position changes of the moving targets, and screening according to the target area and grayscale distribution characteristics to obtain bubble targets; establishing a time tracking sequence for the bubble targets, calculating the movement speed and acceleration of each bubble target, and constructing a local bubble motion vector; spatially interpolating each local bubble motion vector on the image plane to obtain a continuous bubble motion vector field distribution; performing cluster analysis on the bubble motion vector field, extracting the direction of water flow disturbance, and identifying the periodic parameters of the water flow disturbance based on the gradient change of the vector field.

[0015] In the above embodiment, the control system's inter-frame difference operation extracts moving targets, area and grayscale feature screening ensures accurate identification of bubble targets, the time tracking sequence reflects the law of bubble movement, vector field spatial interpolation constructs a continuous disturbance distribution, cluster analysis extracts the disturbance direction and identifies periodic parameters, and realizes accurate modeling of water flow disturbances.

[0016] In combination with some embodiments of the first aspect, in some embodiments, after the step of comparing the target weld direction with the reference direction and generating a graded correction control signal according to the degree of deviation, the method also includes: obtaining contact pressure data of the robot adsorption device and generating a wall contact state distribution map; identifying abnormal weld contour areas according to the contact state distribution map and determining the weld structure type of the abnormal weld contour areas; when the weld structure type changes, updating the detection area division scheme; and selecting a corresponding correction execution strategy according to the weld structure type.

[0017] In the above embodiment, the contact pressure data of the control system reflects the wall state distribution, the identification of abnormal weld contour areas provides structural type information, the dynamic update of the detection area division scheme adapts to structural changes, and the selection of the correction execution strategy ensures the control accuracy in different situations, thereby improving the adaptability of underwater robot operations.

[0018] In combination with some embodiments of the first aspect, in some embodiments, the step of selecting the corresponding correction execution strategy according to the weld structure type specifically includes: classifying the weld structure type into straight welds, fillet welds and lap welds according to the geometric characteristics and pressure distribution characteristics of the abnormal weld contour area; when the weld structure type is a straight weld, adopting a unidirectional progressive correction strategy, and gradually adjusting the robot posture according to the deviation direction until it returns to the reference direction; when the weld structure type is a fillet weld, adopting a bidirectional coordinated correction strategy, and simultaneously adjusting the lateral position and steering angle of the robot to keep the relative position of the detection area and the weld contour stable; when the weld structure type is a lap weld, adopting a step transition correction strategy, pausing the correction action at the weld step, and re-establishing the correction reference after the robot crosses the step.

[0019] In the above embodiment, the control system classifies the weld structure according to the geometric characteristics and pressure distribution characteristics, adopts corresponding correction strategies for different types of welds, and flexibly applies strategies such as one-way progressive, two-way coordination and step transition to achieve precise correction control in complex environments.

[0020] In a second aspect, an embodiment of the present application provides a control system comprising: one or more processors and a memory; the memory is coupled to the one or more processors, the memory being used to store computer program code, the computer program code comprising computer instructions, the one or more processors calling the computer instructions to enable the control system to execute the method described in the first aspect and any possible implementation of the first aspect.

[0021] In a third aspect, an embodiment of the present application provides a computer program product comprising instructions, which, when the computer program product is run on a control system, enables the control system to execute the method described in the first aspect and any possible implementation of the first aspect.

[0022] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium comprising instructions, which, when executed on a control system, causes the control system to execute the method described in the first aspect and any possible implementation of the first aspect.

[0023] It is understandable that the control system provided in the second aspect, the computer program product provided in the third aspect, and the computer storage medium provided in the fourth aspect are all used to execute the methods provided in the embodiments of the present application. Therefore, the beneficial effects that can be achieved can be referenced to the beneficial effects of the corresponding methods and will not be repeated here.

[0024] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. Due to the adoption of technical means such as dynamic filtering parameter adjustment based on local noise intensity, double-layer threshold segmentation, multi-region independent feature extraction, and weighted fitting based on distribution density, the filtering effect can be adaptively optimized according to the imaging quality of different regions in the image preprocessing stage. In the feature extraction stage, the main contour of the weld is retained while the detailed information of the edge transition area is protected. In the feature analysis stage, the propagation of local interference is avoided through independent processing of multiple detection areas. At the same time, the weight coefficient is determined by using the distribution density of feature points to highlight the contribution of high-quality areas. This effectively solves the problems in the existing technology such as fixed parameter filtering easily causing over-smoothing, single threshold segmentation losing edge details, and global feature extraction being easily affected by local interference. Therefore, high precision and high robustness of underwater weld tracking are achieved, and the reliability of underwater robot operations is improved.

[0025] 2. Due to the use of technical means such as grayscale normalization processing, double-layer threshold determination based on grayscale distribution characteristics, and binarization processing for foreground and background separation, the brightness differences between different areas are eliminated in the image preprocessing stage. In the threshold determination stage, the image feature distribution law is accurately grasped by analyzing the grayscale histogram and cumulative distribution function. While the first layer of threshold is used to ensure the complete extraction of the weld body, the second layer of threshold is used to retain the detailed information of the edge transition area. This effectively solves the problems of uneven lighting in underwater environments, the difficulty of a single threshold to take into account both the integrity of the body and the edge details, and the unstable segmentation results in the existing technology. It thus achieves high-quality segmentation of weld features in underwater environments, laying a reliable data foundation for subsequent feature extraction and direction calculation.

[0026] 3. Due to the use of technical means such as wall state analysis based on contact pressure data, identification of abnormal weld contour areas, dynamic division of detection areas, and graded correction strategies, it is possible to monitor the contact state between the robot and the wall in real time, promptly detect changes in weld structure and identify its type, adjust the division scheme of the detection area according to the structural changes, and select the most suitable correction strategy for different types of weld structures. This effectively solves the problems in the existing technology of untimely response to weld structure changes, fixed detection area division leading to feature extraction failure, and single correction strategy that is difficult to adapt to complex environments. It thus realizes precise correction control of underwater robots in complex environments and improves operational adaptability and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a flow chart of an automatic deviation correction method for an underwater robot in an embodiment of the present application; Figure 2 This is another flowchart of the automatic deviation correction method for an underwater robot in an embodiment of the present application; Figure 3It is a schematic diagram of the structure of a physical device of the control system in an embodiment of the present application. DETAILED DESCRIPTION

[0028] The terms used in the following examples of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification of the present application, the singular expressions "a", "an", "above", "the", and "this" are intended to include plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in the present application refers to any or all possible combinations of one or more of the listed items.

[0029] In the following, the terms "first" and "second" are used for descriptive purposes only and should not be understood to imply or suggest relative importance or implicitly indicate the number of the technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of this application, unless otherwise specified, "plurality" means two or more.

[0030] For ease of understanding, the application scenarios of the embodiments of the present application are introduced below.

[0031] A spent fuel pool maintenance project at a nuclear power plant requires automated welding repairs on underwater storage racks and pipes. Due to the unique environment of spent fuel pools, with water depths often exceeding 10 meters and the presence of radiation, operations must be performed using remotely controlled underwater robots. The pool's circulating water system generates complex flow disturbances, while thermal convection caused by spent fuel decay heat also causes water disturbances. This, combined with radiation interference with camera equipment, results in blurred underwater vision and severe image noise. Especially in areas near spent fuel assemblies, strong radiation causes speckle noise in image sensors, while Cherenkov radiation reflected from the pool walls creates uneven blue light interference. These factors significantly impact weld tracking accuracy, increasing the difficulty and risk of maintenance operations.

[0032] In related technologies, underwater weld feature extraction and tracking can be achieved by employing fixed-parameter image filtering algorithms and single-threshold segmentation methods. Specifically, the image is first subjected to Gaussian filtering for noise reduction. The OTSU algorithm is then used to determine a global threshold for binary segmentation. Finally, a Hough transform is used to detect straight line features and determine weld orientation. The following describes a scenario using the related-art automatic deviation correction method for underwater robots.

[0033] A nuclear industry research institute used traditional image processing methods for spent fuel pool welding operations. They used a fixed-parameter median filter to remove impulse noise caused by radiation and a single-threshold Canny operator for edge detection. However, in practice, when the underwater robot approached areas of varying radiation intensity, image quality varied, and the fixed processing parameters were unable to adapt. Furthermore, blue light interference from Cherenkov radiation affected edge detection accuracy, resulting in unstable weld feature extraction. Furthermore, because global feature extraction was employed, strong radiation interference in a specific area affected overall tracking, causing the robot to deviate from its intended trajectory.

[0034] The automatic deviation-correction method for underwater robots in the embodiments of this application, however, achieves precise tracking of underwater welds by adjusting filtering parameters and performing double-layer threshold segmentation to preserve edge details. It also utilizes multiple detection regions to independently extract feature points to avoid local interference propagation. This not only improves the accuracy of feature extraction but also enhances the system's adaptability to changes in the underwater environment. The following describes scenarios in which the automatic deviation-correction method for underwater robots in this application is used.

[0035] By applying the basic scheme of this application, a nuclear power plant has implemented a radiation-proof adaptive weld tracking system in the maintenance of spent fuel pools. By analyzing the degree of radiation interference in the image in real time, the system can adjust the filter parameters to suppress radiation noise. The double-layer threshold segmentation method can not only effectively separate the interference of Cherenkov radiation, but also ensure the integrity of the weld contour. The system divides the detection area into multiple independent units. Even if some areas are affected by strong radiation, other areas can still provide reliable feature information. The weighted fitting method based on radiation intensity distribution gives priority to the feature data of low-radiation areas, and the dynamic error band can effectively screen out abnormal points affected by radiation interference.

[0036] It can be seen that the automatic deviation correction method for the underwater robot in the embodiment of the present application can not only realize the accurate extraction of weld features, but also effectively solve problems such as underwater environmental noise interference, unstable image quality and local feature abnormalities, thereby realizing high-precision and high-reliability tracking control of the underwater welding robot.

[0037] For ease of understanding, the following describes the process of the method provided by this implementation in combination with the above scenario. Figure 1 , which is a flow chart of the automatic deviation correction method for the underwater robot in the embodiment of the present application.

[0038] S101, obtaining a weld image captured by an underwater robot camera, adjusting a filtering parameter according to noise intensity of a local image in the weld image, and obtaining a denoised image.

[0039] Among them, the underwater robot camera refers to the image sensor device installed on the underwater robot for image acquisition; the weld image refers to the digital image containing the weld area captured by the camera; the local image noise intensity represents the degree of noise interference in a specific area of ​​the image; the filtering parameters are used to represent the various control parameters in the image filtering algorithm, including window size, weight coefficient, etc.; the denoised image refers to the image after noise removal processing.

[0040] When the underwater robot begins operation, the control system first needs to obtain a clear image of the weld seam as the basis for subsequent processing. Specifically, the control system uses the underwater robot's camera to capture the weld seam image in real time and divides the image into multiple local areas. For each local area, the control system calculates the variance and gradient distribution of the image grayscale values ​​to assess the noise intensity in that area. Based on the noise intensity, the control system adjusts the filter algorithm parameters applied to that area, using a higher filter intensity for areas with higher noise intensity and a lower filter intensity for areas with lower noise intensity. This approach removes noise while preserving image detail to the greatest extent possible.

[0041] In practical applications, uneven illumination in underwater environments can lead to different noise characteristics in different areas of the same image. To address this, the control system can incorporate a local contrast enhancement mechanism: First, a local contrast distribution map of the image is calculated to identify areas with abnormal contrast. Adaptive histogram equalization is then performed on these areas to improve the local contrast of the image. Noise estimation and filtering are then performed to ensure more accurate and effective adjustment of filtering parameters.

[0042] S102 , performing double-layer threshold segmentation on the denoised image to obtain a binary image.

[0043] Among them, the denoised image refers to a digital image that has been processed by noise removal; double-layer threshold segmentation refers to a processing method that uses two different thresholds for image segmentation; the first-layer threshold is used to represent the grayscale threshold for extracting the main contour of the weld; the second-layer threshold is used to represent the grayscale threshold for retaining the transition area of ​​the weld edge; the binary image refers to the processing result of converting the image pixel value into two discrete values, usually using 0 to represent the background and 1 to represent the target.

[0044] After obtaining the denoised image, the control system needs to accurately segment the weld area from the background. Specifically, the control system first performs grayscale normalization on the denoised image to eliminate the brightness differences between different areas. Then, the grayscale histogram and cumulative distribution function of the normalized image are calculated to analyze the grayscale distribution characteristics of the image. Based on the grayscale distribution characteristics, the control system uses the OTSU algorithm to determine the first-layer threshold to extract the main weld area. Next, the control system analyzes the grayscale distribution of the weld candidate area and uses an adaptive threshold algorithm to calculate the second-layer threshold to retain the transition area of ​​the weld edge. Finally, the control system marks the area above the first-layer threshold as the foreground and the area below the second-layer threshold as the background to obtain the final binary image.

[0045] It should be noted that dual-threshold segmentation is an adaptive segmentation method tailored to underwater weld image features. The system first preprocesses the image with grayscale normalization. Then, it determines two key thresholds by analyzing the grayscale histogram and cumulative distribution function. The first threshold is determined primarily based on the overall grayscale distribution of the image, by maximizing the inter-class variance, to reliably extract the main weld contour. The second threshold is determined by analyzing the local grayscale distribution of the weld candidate region, preserving the transition features of the weld edge. The area between these two thresholds is considered the edge transition zone, and the system dynamically determines pixel assignment based on local region characteristics. As the turbidity of the underwater environment changes, the system automatically adjusts the interval between the two thresholds, preserving as much edge detail as possible while maintaining the integrity of the weld main body. For example, when the water is turbid, the system will appropriately narrow the interval between the two thresholds to prevent feature loss due to over-segmentation. When the water is clear, the threshold interval will be widened to capture richer edge features.

[0046] In practical applications, the turbidity of underwater environments can vary over time, resulting in unstable image contrast. To address this, the control system can employ a dynamic threshold adjustment mechanism: by tracking changes in the grayscale distribution across multiple consecutive image frames, a dynamic model of the grayscale distribution is established. When a change in the grayscale distribution is detected, the control system automatically adjusts the threshold calculation strategy, perhaps by adjusting the inter-class variance weights of the OTSU algorithm or modifying the local window size of the adaptive threshold, to ensure the stability of the segmentation results. For example, as water turbidity increases, the interval between two thresholds can be appropriately reduced to better preserve weld features.

[0047] S103 , setting multiple detection areas in the binary image, independently extracting edge feature points in each detection area, and determining an effective detection area based on the number distribution of feature points in each detection area, and converting the edge feature points in the effective detection area into coordinate data.

[0048] Among them, the binary image represents the black and white image after double-layer threshold segmentation; the detection area refers to the local rectangular area divided in the image for feature extraction; the edge feature points are used to represent the key pixel points on the edge of the weld contour; the distribution of the number of feature points represents the density and distribution of edge feature points in each detection area; the effective detection area refers to the detection area containing effective weld features; the coordinate data is used to represent the position information of the edge feature points in the image coordinate system.

[0049] After obtaining the binary image, the control system needs to accurately extract the weld features. Specifically, the control system first sets up multiple rectangular detection areas at equal intervals along the general direction of the weld in the binary image. The size of each detection area is determined according to the image resolution and the weld width. For each detection area, the control system independently applies an edge detection algorithm to extract feature points, including steps such as calculating pixel gradients, non-maximum suppression, and dual threshold detection. Then, the control system analyzes the number and spatial distribution of feature points in each detection area, identifies areas with abnormal feature point distribution through cluster analysis, and excludes them to obtain effective detection areas. Finally, the control system maps the edge feature points in the effective detection area to the image coordinate system to obtain standardized coordinate data.

[0050] In actual applications, rust or attachments may exist on the weld surface, resulting in abnormal distribution of feature points in certain detection areas. To address this, the control system can introduce a feature point quality assessment mechanism: first, a feature template of the ideal weld edge is established, including the density distribution and spatial arrangement pattern of the feature points. For each feature point extracted from the detection area, its matching degree with the template is calculated. When the matching degree is lower than the threshold, the control system automatically expands the detection range of the area or adjusts the feature extraction parameters until a set of feature points that meets the quality requirements is obtained. For example, for areas with sparse feature points, the gradient threshold of edge detection can be lowered, while adding neighborhood constraints to ensure the reliability of the feature points.

[0051] S104 , determining a weight coefficient according to the distribution density of edge feature points in the detection area, performing partition weighted fitting on the coordinate data, and obtaining an initial weld direction.

[0052] Among them, the distribution density indicates the degree of aggregation of edge feature points in the detection area; the weight coefficient is used to indicate the importance of feature points in different regions in the fitting calculation; partition weighted fitting refers to the method of assigning different weights to feature points in different regions for curve fitting; the initial weld direction represents the preliminary estimate of the spatial direction of the weld obtained by fitting; the coordinate data refers to the position information of the edge feature points in the image coordinate system.

[0053] After obtaining the coordinates of the feature points in the effective detection area, the control system needs to accurately calculate the weld direction. Specifically, the control system first calculates the horizontal and vertical projection distribution of the feature points in each detection area and constructs a two-dimensional density distribution matrix. Based on this matrix, the control system calculates the aggregation index of the feature points in each area, including local density, spatial uniformity and other characteristics. According to the aggregation index, the control system assigns a weight coefficient to each detection area, and the area with the more concentrated and uniform distribution of feature points obtains a higher weight. Then, the control system divides the detection area into several equal-length sub-areas along the estimated weld direction, and applies the corresponding weight coefficient to weight the feature points in each sub-area. Finally, the control system uses the least squares method to perform spatial curve fitting on all weighted feature points to obtain the spatial equation of the weld, and determines the direction vector of the equation as the initial weld direction.

[0054] In practical applications, welds may contain curved or deformed areas, resulting in large errors in linear fitting. To address this, the control system can adopt a segmented weighted fitting strategy: First, based on the spatial distribution characteristics of the characteristic points, the turning points or curvature change points of the weld are identified. The weld area is then divided into multiple subsegments using these characteristic points as boundaries. Weight calculation and curve fitting are performed separately for each subsegment, and polynomial models of different orders can be selected. Finally, the control system smoothly connects the fitting results of each subsegment through methods such as spline interpolation to obtain a complete description of the weld direction. For example, for sharp turns, a higher-order polynomial model can be used, while increasing the weights of the characteristic points in this area to improve fitting accuracy.

[0055] S105. Construct a dynamic error band based on the initial weld direction, use the coordinate points outside the dynamic error band as candidate outlier points, determine the final outlier point among the candidate outlier points based on the continuity criterion of adjacent feature points, refit the remaining coordinate points outside the final outlier point, and obtain the target weld direction.

[0056] Among them, the dynamic error band represents a variable width tolerance area constructed according to the initial weld direction; the candidate outlier point refers to the suspicious feature point that falls outside the error band; the continuity criterion is used to represent the criterion for evaluating the spatial continuity of adjacent feature points; the final outlier point refers to the outlier feature point confirmed by the continuity judgment; the target weld direction refers to the final weld direction obtained by refitting after removing the outlier point.

[0057] After obtaining the initial weld direction, the control system needs to further optimize the fitting results. Specifically, the control system first constructs a dynamic error band along the initial weld direction. The width of the error band is adjusted according to the degree of discreteness of the local feature point distribution. The more concentrated the feature point distribution is, the narrower the error band is. Then, the control system identifies the feature points that fall outside the error band and marks them as candidate outlier points. For each candidate outlier point, the control system analyzes its spatial distance, angle, curvature and other features with adjacent feature points to establish a continuity evaluation model. Based on this model, the control system determines whether the candidate outlier point violates the continuity characteristics of the weld contour and determines the point that violates continuity as the final outlier point. Finally, the control system removes these outliers from the feature point set and re-weighted fits the remaining feature points to obtain a more accurate target weld direction.

[0058] It should be noted that the dynamic error band enables adaptive screening based on the distribution of weld features. The system constructs a tolerance zone with variable width based on the weld orientation obtained from the initial fitting. The width of the error band is not a fixed value but is dynamically adjusted based on the distribution characteristics of local feature points: in areas with dense and regular feature points, the error band is narrowed to improve screening accuracy; in areas with sparse or irregular feature points, the error band is appropriately widened to accommodate reasonable fluctuations. For coordinate points that fall outside the error band, the system further analyzes continuity features such as spatial distance, angle, and curvature with adjacent feature points to comprehensively determine whether to mark them as outliers. In actual application, when a large number of continuous outliers are detected in a certain area, the system combines the pressure distribution characteristics and historical data in that area to assess whether they are caused by changes in the weld structure, thereby avoiding misidentification of normal structural feature points. This dynamic adjustment mechanism ensures accurate identification and elimination of true outliers in complex underwater environments, while retaining valid points that reflect actual weld characteristics.

[0059] In practical applications, local defects or bubble interference on the weld surface may cause a large number of continuous feature points to be misjudged as abnormal points. To this end, the control system can introduce an abnormal point verification mechanism based on regional consistency: first, the detected abnormal points are clustered according to spatial position to form abnormal regions. For each abnormal region, its shape characteristics and the distribution pattern of internal feature points are analyzed. If the abnormal region shows a pattern similar to the weld defect or bubble characteristics, the control system will re-evaluate the validity of the feature points in the region and adjust the error band parameters or continuity criterion threshold if necessary to avoid excessive deletion of valid feature information. For example, for areas suspected of weld pits, the constraints of the continuity criterion can be appropriately relaxed to retain feature points that reflect the actual weld morphology.

[0060] S106: Compare the target weld direction with the reference direction, and generate a graded deviation correction control signal according to the degree of deviation.

[0061] Among them, the target weld direction represents the final spatial direction of the weld; the reference direction is used to represent the ideal weld motion trajectory, which is set and input into the control system before the operation begins. It is a path pre-planned by analyzing the design drawings of the workpiece or through teaching, etc.; the degree of deviation refers to the angle and position deviation between the target direction and the reference direction; the graded correction control signal represents different levels of control instructions generated according to the degree of deviation, which is used to guide the underwater robot to adjust its posture.

[0062] After obtaining the target weld direction, the control system needs to generate corresponding correction control instructions. Specifically, the control system first converts the target weld direction into a direction vector in the standard coordinate system, and at the same time obtains the preset reference direction. Then, the control system calculates the angle and vertical distance between the two direction vectors, and comprehensively evaluates the degree of deviation in weld tracking. Based on the degree of deviation, the control system divides the deviation state into multiple levels, such as slight deviation, medium deviation, and severe deviation. For each deviation level, the control system designs a corresponding correction strategy, including adjusting the movement speed, steering angle, and posture parameters. Finally, the control system converts the correction strategy into a specific control signal sequence and sends it to the actuator of the underwater robot.

[0063] It should be noted that the reference direction, which serves as a reference for the ideal motion trajectory of the underwater robot during weld tracking, is determined in two stages: static presetting and dynamic updating. During the static presetting stage, the control system first extracts theoretical weld direction data from construction drawings and tank structural design documents, and then establishes an initial reference direction database based on historical inspection records. During actual operation, the control system dynamically updates the reference direction based on multi-source sensor data. First, it calculates the statistical average direction using weld features from multiple frames obtained through image processing. Spatial correction is then performed using attitude sensor data. Local reference parameters are then adjusted based on structural characteristics fed back by pressure sensors. Using a data fusion algorithm, the control system performs a weighted combination of the static presetting value and the dynamic inspection results to determine the final optimized reference direction. The weighting is adaptively adjusted based on weld quality requirements, environmental interference levels, and robot status.

[0064] For example, during a spent fuel pool inspection at a nuclear power plant, when an underwater robot passes through a straight weld, the control system primarily uses a preset theoretical direction as a reference. However, when approaching a corner, the system prioritizes real-time detection results, predicting the corner curvature through multi-frame image analysis and adjusting the reference direction in advance. In areas subject to significant flow disturbances, the system applies compensation corrections to the reference direction, combining analysis of bubble motion characteristics to ensure tracking control stability.

[0065] In practical applications, underwater robots' posture adjustments may experience response lag due to mechanical inertia or water flow disturbances. To address this, the control system can employ an adaptive control gain strategy: by monitoring the robot's motion state and environmental parameters in real time, a dynamic response model is established. When response lag is detected, the control system automatically adjusts the control gain parameters, increasing the control gain appropriately for directions with slower response and reducing it appropriately for directions prone to overshoot. Simultaneously, a velocity feedforward compensation mechanism is introduced to estimate the impact of water flow velocity and direction on the robot's motion and preemptively apply compensatory control. For example, in strong water flow, the system increases the control gain perpendicular to the flow direction and calculates the required compensation torque based on the water flow velocity to ensure the accuracy and stability of the corrective action.

[0066] This dynamic response model is used to adjust the control gain during the underwater robot's posture adjustment. During the training phase, the system collects the robot's motion state data (including velocity, acceleration, and attitude angle) and environmental parameters (such as water velocity and direction) under different water flow conditions as training samples. The deviation between the actual response time and the ideal response time is used as the loss function for training. The model itself is a state predictor. Its inputs include the current motion state parameter vector and the environmental parameter vector. It uses a multi-layer perceptron structure to perform feature extraction and nonlinear mapping, and outputs predicted response characteristics. During the operation phase, the control system inputs the real-time state data and environmental data into the model. Based on the predicted response characteristics, it calculates the control gain adjustment coefficients for each direction, increasing the gain in directions with slow response and reducing the gain in directions prone to overshoot. The system also calculates the required compensation torque based on the water flow velocity vector. For example, if a vertical response lag is detected, the model outputs a larger vertical gain coefficient and simultaneously calculates the required compensation torque based on the lateral component of the water flow, thereby achieving precise control of posture adjustment.

[0067] The above embodiment mainly describes basic feature extraction and direction estimation methods. In practical applications, functional modules such as posture compensation, water flow analysis, and structure recognition can be combined to further improve the system's environmental adaptability and control accuracy. The following supplements the scenarios of this embodiment.

[0068] During a grid maintenance project at a nuclear power plant's spent fuel pool, the engineering team further optimized system performance. By introducing a real-time radiation dose rate monitoring module, the system can predict high-radiation areas and adjust the work path in advance. The image sampling strategy automatically avoids periods of strong radiation and adjusts exposure parameters to reduce the impact of radiation on the sensor. The system can also adaptively update image processing parameters based on the radiation intensity distribution at different depths of the pool. For example, in high-radiation areas near the pool bottom, the system increases the sampling frequency and enables stronger filtering. At corner welds, the system selects the optimal detection area layout based on the local radiation environment to ensure the stability of feature extraction. These optimization measures improve the system's operational reliability in strong radiation environments, ensuring maintenance quality while also reducing radiation exposure for workers.

[0069] After combining the above scenarios, the following is a more detailed description of the process of the method provided by this implementation. Figure 2 , is another flow chart of the automatic deviation correction method for the underwater robot in the embodiment of the present application.

[0070] S201 , obtaining a weld image captured by a camera of an underwater robot, adjusting a filtering parameter according to noise intensity of a local image in the weld image, and obtaining a de-noised image.

[0071] Referring to step S101 , the control system dynamically optimizes the filtering effect according to the local noise intensity of the weld image.

[0072] In some embodiments, the control system will optimize image acquisition through posture data and bubble trajectory analysis, that is, the control system will obtain the posture data collected by the underwater robot's posture sensor, establish a spatial transformation matrix between the image plane and the weld plane; perform perspective correction on the weld image according to the spatial transformation matrix to eliminate perspective deviation; extract the moving bubble trajectory in the weld image, establish a bubble motion vector field, and determine the direction of water flow disturbance; adjust the image sampling timing according to the direction of water flow disturbance to avoid the peak moment of water flow disturbance.

[0073] Among them, the attitude sensor refers to the sensing equipment used to measure the spatial attitude of the underwater robot; the attitude data refers to the robot's spatial posture parameters such as pitch angle, roll angle and heading angle; the spatial transformation matrix is ​​used to represent the conversion relationship between different coordinate systems; perspective correction refers to the elimination of image deformation caused by changes in viewing angle; the motion bubble trajectory represents the spatial path of the bubble movement in the water; the bubble motion vector field refers to the vector distribution that describes the bubble movement trend; the water flow disturbance direction is used to represent the movement direction of the local water flow; the sampling timing represents the time series of image acquisition.

[0074] Before performing image processing, the control system needs to eliminate perspective deviation and optimize the sampling strategy. Specifically, the control system first reads the attitude angle data collected in real time by the attitude sensor and, combined with the camera calibration parameters, calculates the spatial transformation relationship between the image plane and the weld plane. Based on this transformation relationship, the control system constructs a perspective transformation matrix and performs geometric correction on the original image to eliminate image distortion caused by changes in the robot's attitude. The control system then identifies moving bubbles in the water through image sequence analysis and tracks and records the bubble's movement trajectory. Through statistical analysis of multiple bubble trajectories, the control system establishes a vector field model that reflects the local water flow motion characteristics. Finally, based on the time-varying characteristics of the water flow vector field, the control system predicts the peak moment of water flow disturbance and adjusts the image sampling interval to avoid image acquisition during periods of strong disturbance.

[0075] In some embodiments, posture compensation and sampling optimization can be achieved in a variety of ways: Optionally, the control system can use a posture estimation method based on Kalman filtering, first fusing the measurement data of the accelerometer and gyroscope to predict the robot's posture change trend, and then adaptively adjust the perspective transformation parameters based on the prediction results, while optimizing the image sampling timing to match the posture change cycle; Optionally, the control system can use a motion compensation method based on optical flow analysis, estimating relative motion by calculating the optical flow field between consecutive image frames, building a compensation model based on the posture data, and adjusting the image correction parameters and sampling strategy. It is understandable that other methods can also be used to achieve motion compensation and sampling control, which are not limited here.

[0076] In some embodiments, the control system tracks bubble motion characteristics through multi-frame image analysis, that is, the control system performs inter-frame difference operations on continuously collected multi-frame images, extracts the position changes of the moving targets, and screens them according to the target area and grayscale distribution characteristics to obtain bubble targets; establishes a time-series tracking sequence for the bubble targets, calculates the motion speed and acceleration of each bubble target, and constructs a local bubble motion vector; spatially interpolates each local bubble motion vector on the image plane to obtain a continuous bubble motion vector field distribution; performs cluster analysis on the bubble motion vector field, extracts the direction of water flow disturbance, and identifies the periodic parameters of the water flow disturbance based on the gradient change of the vector field.

[0077] Among them, inter-frame difference operation means calculating the pixel changes between adjacent image frames; moving target refers to the object whose position changes in the image; target area represents the number of pixels occupied by the bubble in the image; grayscale distribution feature is used to represent the brightness distribution pattern of the bubble; time tracking sequence refers to the time series that records the motion state of the bubble; local bubble motion vector represents the motion direction and speed of a single bubble; spatial interpolation refers to constructing a continuous distribution between discrete observation points; cluster analysis is used to represent the grouping of similar motion features; and period parameter refers to the temporal characteristics of water flow disturbance.

[0078] After acquiring the image sequence, the control system needs to analyze the characteristics of the water flow disturbance. Specifically, the control system first performs inter-frame difference calculations on the continuously acquired image sequence to extract dynamic targets in the image. By setting area thresholds and grayscale feature constraints, the control system filters out targets that meet the bubble characteristics and removes other interfering targets. For each identified bubble target, the control system establishes a motion state sequence containing position, velocity, and acceleration, and uses Kalman filtering to smooth the motion state. Then, the control system projects the motion vectors of each bubble onto the image plane, uses radial basis functions for spatial interpolation, and constructs a continuous motion vector field. Finally, the control system performs density clustering analysis on the vector field, extracts the main water flow motion direction, and identifies the periodic characteristics of the water flow disturbance through Fourier analysis.

[0079] In practical applications, the motion trajectories of multiple bubbles may interfere with or intersect each other, affecting the accuracy of water flow feature extraction. To address this, the control system can adopt a hierarchical trajectory management strategy: first, a multi-level trajectory tracking model is established, with the bottom layer responsible for estimating the state of individual bubbles, the middle layer processing trajectory interactions and associations, and the top layer analyzing the overall motion pattern. When trajectory intersection or occlusion is detected, the system initiates a trajectory prediction and reassociation mechanism to restore trajectory continuity through motion consistency and appearance similarity. For example, for a bubble that temporarily disappears, the system maintains its trajectory hypothesis for a period of time, waiting for it to reappear before reconnecting the trajectory to ensure the integrity of water flow feature extraction.

[0080] S202 : performing grayscale normalization processing on the denoised image to obtain a standard grayscale image.

[0081] Among them, the denoised image refers to the original image after noise suppression processing; grayscale normalization refers to the process of mapping image pixel values ​​to a standard range; standard grayscale image refers to an image whose pixel values ​​are normalized to the interval [0, 1] or [0, 255]; pixel intensity represents the brightness value of each point in the image; grayscale distribution characteristics are used to represent the overall brightness distribution of the image; dynamic range refers to the difference between the maximum and minimum grayscale values ​​of the image; contrast represents the difference in the brightness of the image; linear mapping function is used to represent the mathematical relationship of grayscale value conversion; histogram equalization refers to the processing method of optimizing grayscale distribution; normalization coefficient represents the scale factor of grayscale mapping.

[0082] After obtaining the denoised image, the control system needs to perform grayscale normalization to eliminate the effects of uneven illumination. Specifically, the control system first analyzes the grayscale histogram distribution of the denoised image and calculates the image's maximum, minimum, and average grayscale values. The control system then selects an appropriate normalization strategy based on the image's dynamic range characteristics: when the image contrast is low, the control system uses a linear stretching method to expand the grayscale dynamic range through remapping; when the image contains locally dark or bright areas, the control system uses piecewise linear mapping, applying different mapping coefficients to different grayscale intervals; and when the image's grayscale distribution is uneven, the control system combines histogram equalization technology to optimize the grayscale distribution. Finally, the control system maps the processed grayscale values ​​uniformly to a standard range, generating a standardized grayscale image. To ensure processing stability, the control system limits abnormal grayscale values ​​to avoid overflow.

[0083] S203 , calculating the grayscale histogram and cumulative distribution function of the standard grayscale image to obtain grayscale distribution characteristics.

[0084] Among them, the standard grayscale image represents an image that has undergone grayscale normalization processing; the grayscale histogram is used to represent the statistical distribution of the number of pixels at each grayscale level in the image; the cumulative distribution function refers to the cumulative sum curve of the grayscale histogram; and the grayscale distribution feature represents the statistical feature that reflects the overall distribution law of the grayscale value of the image.

[0085] After obtaining a standard grayscale image, the control system needs to analyze the image's grayscale distribution characteristics. Specifically, the control system first counts the number of pixels corresponding to each grayscale level in the image to generate a grayscale histogram. Then, starting from the minimum grayscale value, the control system accumulates the number of pixels at each grayscale level and calculates the cumulative distribution function. By analyzing the peak and valley positions of the grayscale histogram and the changing trend of the cumulative distribution function, the control system obtains key grayscale features of the image, including the grayscale distribution range of the background and target areas, and the grayscale span of edge transition areas.

[0086] In practical applications, images may contain localized areas of strong reflection or shadow, resulting in abnormal peaks in the grayscale distribution. To address this, the control system can employ an adaptive histogram smoothing strategy: First, it detects abnormal fluctuations in the grayscale histogram and preprocesses the histogram using median filtering or Gaussian smoothing. It then analyzes the local variation trends of the smoothed histogram to identify the true grayscale distribution characteristics. For example, in areas with strong reflection, the system corrects for abnormal peaks by analyzing the grayscale distribution of the surrounding area, ensuring a more reliable distribution characteristic.

[0087] S204: Determine the first layer threshold based on the grayscale distribution characteristics, and perform initial segmentation to obtain weld candidate areas.

[0088] Among them, the grayscale distribution feature represents the statistical distribution law of the grayscale value of the image; the first-layer threshold is used to represent the grayscale critical value for distinguishing the weld body from the background; the initial segmentation refers to the image segmentation operation using the first-layer threshold; the weld candidate area refers to the image area that may contain the weld obtained after the initial segmentation.

[0089] After obtaining the grayscale distribution characteristics, the control system needs to determine an appropriate segmentation threshold. Specifically, the control system first analyzes the bimodal characteristics of the grayscale histogram, where higher grayscale values ​​correspond to the weld area and lower grayscale values ​​correspond to the background area. The control system then uses the OTSU algorithm to calculate the inter-class variance and determines the optimal segmentation threshold by maximizing the inter-class variance. To improve segmentation reliability, the control system also considers the rate of change of the cumulative distribution function, searching for the location with the steepest grayscale change near the point of maximum inter-class variance. The grayscale value at this location is used as the first-level threshold. This threshold is used to segment the image and obtain preliminary candidate weld areas.

[0090] In some embodiments, the first-layer threshold can be determined in a variety of ways: Optionally, the control system can use an improved iterative threshold method, first estimating an initial threshold, repeatedly calculating the average grayscale value of the target area and the background area, obtaining a stable segmentation threshold through iterative optimization, and modifying it in combination with image entropy information; Optionally, the control system can use a threshold selection method based on local adaptation to divide the image into multiple sub-areas, calculate the local optimal threshold for each sub-area, and obtain the global threshold through weighted fusion. It is understood that other methods can also be used to determine the first-layer threshold, which are not limited here.

[0091] In practical applications, reflections or oxidation on the weld surface can lead to uneven grayscale distribution, affecting the accuracy of threshold selection. To address this, the control system can introduce a threshold optimization mechanism based on morphological features: First, a tentative segmentation is performed using multiple candidate thresholds. The resulting candidate regions are analyzed for morphological features, including connectivity, boundary smoothness, and regional compactness. The threshold that best matches the weld morphological characteristics is selected as the final first-layer threshold. For example, when excessive fragmentation is detected in the segmentation results, the system will appropriately adjust the threshold to maintain the integrity of the weld region.

[0092] S205 , calculating the second layer threshold according to the grayscale distribution of the weld candidate area, and performing secondary segmentation to retain the edge transition area.

[0093] Among them, the weld candidate area represents the potential weld area obtained by the initial segmentation; the second-layer threshold is used to represent the grayscale critical value for retaining edge transition features; secondary segmentation refers to the fine segmentation operation performed using the second-layer threshold; the edge transition area represents the gradient area between the weld body and the background.

[0094] After obtaining the weld candidate region, the control system needs to further preserve edge detail information. Specifically, the control system first extracts the boundary of the weld candidate region and sets a local analysis window centered on the boundary. Within each window, the control system calculates the gradient distribution and direction of the grayscale values ​​to construct a grayscale variation model for the edge region. Based on this model, the control system analyzes the continuity characteristics of the grayscale transition and determines the optimal threshold that preserves the edge transition characteristics. To adapt to the edge characteristics of different regions, the control system adjusts the second-layer threshold based on the local gradient strength, ensuring edge integrity while avoiding the introduction of excessive noise.

[0095] In some embodiments, the calculation of the second-layer threshold can be achieved through various methods: Optionally, the control system can use an adaptive threshold method based on edge strength, first calculating the gradient amplitude map of the candidate region boundary, determining the gradient threshold through statistical analysis, and adjusting the second-layer threshold based on the local grayscale distribution characteristics; Optionally, the control system can use a threshold determination method based on region growing, using the main weld region as the seed region and gradually expanding to the edge transition region by controlling the grayscale similarity threshold of the growth condition. It is understood that other methods can also be used to calculate the second-layer threshold, which are not limited here.

[0096] In practical applications, weld edges may exhibit irregular variations or blurred areas, leading to unstable edge feature extraction. To address this, the control system can employ a multi-scale edge analysis strategy: This involves decomposing the image at multiple scales using a Gaussian pyramid and analyzing the consistency of edge features at different scales. When edge feature instability is detected at a particular scale, the system references analysis results at other scales for correction. For example, for blurred edge regions, the system determines the approximate boundary location at a coarse scale, then precisely locates the edge transition region at a fine scale, ultimately determining an appropriate second-layer threshold.

[0097] S206 , marking the area with a value higher than the first layer threshold as a foreground pixel value, and marking the area with a value lower than the second layer threshold as a background pixel value, to obtain a binary image.

[0098] Among them, the foreground pixel value represents the binary value used to mark the weld area; the background pixel value is used to represent the binary value of the non-weld area; the binary image refers to the processing result image containing only two pixel values; the first layer threshold and the second layer threshold represent the grayscale critical values ​​for distinguishing the weld body and the edge transition area, respectively.

[0099] After determining the two thresholds, the control system needs to generate the final binarization result. Specifically, the control system first determines the grayscale value of each pixel: Pixels with grayscale values ​​above the first-level threshold are identified as the main weld area and marked as foreground pixels (usually 255 or 1); pixels with grayscale values ​​below the second-level threshold are identified as background areas and marked as background pixels (usually 0). For pixels with grayscale values ​​between the two thresholds, the control system analyzes their spatial relationship with the marked areas and grayscale continuity, dynamically determining their location based on local region characteristics. Finally, the control system performs a connectivity analysis on the marked results, eliminating isolated small areas to obtain a complete binary image.

[0100] In practical applications, the transition region between two thresholds may contain complex grayscale changes, resulting in fragmented or discontinuous labeling results. To address this, the control system can introduce a topology-based optimization mechanism: first, the topological features of the binarization result are analyzed, including the number of connected components, region shape, and boundary curvature. When a topological anomaly is detected, the system automatically adjusts the labeling strategy for the intermediate region. For example, for a broken weld region, the system analyzes the morphological features on both sides of the break and restores regional connectivity through an intelligent completion algorithm. At the same time, the system also considers the overall direction of the weld to ensure the consistency and integrity of the labeling results.

[0101] S207 , setting multiple detection areas in the binary image, extracting edge feature points independently in each detection area, determining an effective detection area based on the number distribution of feature points in each detection area, and converting the edge feature points in the effective detection area into coordinate data.

[0102] Referring to step S103 , the control system extracts valid feature points from multiple detection areas and converts them into coordinate data.

[0103] S208 , determining a weight coefficient according to the distribution density of the edge feature points in the detection area, performing partition weighted fitting on the coordinate data, and obtaining an initial weld direction.

[0104] Referring to step S104 , the control system performs weighted fitting based on the distribution density of the characteristic points to calculate the weld direction.

[0105] In some embodiments, the control system determines the regional weight coefficient based on the density distribution of feature points, that is, the control system calculates the horizontal and vertical projection distribution of the edge feature points in each detection area to obtain a two-dimensional density distribution matrix; calculates the feature point aggregation degree of each area based on the two-dimensional density distribution matrix, and determines the weight coefficient based on the feature point aggregation degree; divides the detection area into several sub-areas of equal length along the direction of the weld, and applies the corresponding weight coefficient in each sub-area to weight the feature point coordinates; performs least squares fitting on all weighted feature point coordinates to obtain the spatial equation of the weld, and determines the direction vector of the spatial equation as the initial weld direction.

[0106] Among them, the detection area represents the local image area used for feature extraction; the edge feature point refers to the key pixel point on the weld contour; the projection distribution is used to represent the density distribution of the feature points in the horizontal and vertical directions; the two-dimensional density distribution matrix represents the numerical expression of the spatial distribution of the feature points; the feature point aggregation degree refers to the spatial concentration degree of the feature points; the weight coefficient is used to represent the importance of different areas in the fitting calculation; the sub-area refers to the local area of ​​equal length divided along the weld direction; the spatial equation represents the mathematical expression describing the weld trajectory; the direction vector refers to the unit vector characterizing the direction of the weld.

[0107] After completing the extraction of edge feature points, the control system needs to determine the spatial direction of the weld. Specifically, the control system first counts the number of projections of the feature points in each detection area in the horizontal and vertical directions, and generates a two-dimensional matrix reflecting the distribution density of the feature points. Based on this matrix, the control system calculates the feature point aggregation index of each area, including statistical features such as the local density mean, variance, and peak value. According to the aggregation index, the control system assigns a weight coefficient to each area, and the area with more concentrated and uniform distribution of feature points obtains a higher weight. Then, the control system divides the detection area into multiple sub-areas along the estimated weld direction, and applies the corresponding weight coefficient to perform a weighted transformation on the coordinates of the feature points in each sub-area. Finally, the control system uses the least squares method to perform curve fitting on all weighted coordinate points to obtain a mathematical equation describing the spatial trajectory of the weld, and normalizes the tangent vector of the equation as the initial weld direction.

[0108] In some embodiments, spatial distribution analysis and weighted fitting of feature points can be achieved in a variety of ways: Optionally, the control system can adopt a distribution analysis method based on kernel density estimation, first select an appropriate kernel function to calculate the local density distribution of the feature points, then obtain the edge distribution through integral projection, determine the weight coefficient based on the statistical characteristics of the edge distribution, and finally use the weighted RANSAC algorithm for robust fitting; Optionally, the control system can adopt a direction estimation method based on principal component analysis, first perform principal component analysis on the feature point set to obtain the main direction, then calculate the contribution of each feature point to the main direction as the weight basis, adjust the weight distribution through iterative optimization, and finally use the weighted orthogonal distance regression method for accurate fitting. It is understandable that other methods can also be used to achieve distribution analysis and curve fitting, which are not limited here.

[0109] S209. Construct a dynamic error band based on the initial weld direction, use the coordinate points outside the dynamic error band as candidate outlier points, determine the final outlier point among the candidate outlier points based on the continuity criterion of adjacent feature points, refit the remaining coordinate points outside the final outlier point, and obtain the target weld direction.

[0110] Referring to step S105 , the control system will filter out abnormal points through the dynamic error band and re-fit to determine the target direction.

[0111] S210: Compare the target weld direction with the reference direction, and generate a graded deviation correction control signal according to the degree of deviation.

[0112] Referring to step S106 , the control system generates a graded control signal according to the deviation from the reference direction.

[0113] S211. Acquire contact pressure data of the robot adsorption device and generate a wall contact state distribution map.

[0114] Among them, the adsorption device refers to the mechanism used by the underwater robot to attach to the wall; the contact pressure data is used to represent the pressure value at the contact point between the adsorption device and the wall; the wall contact state distribution diagram refers to a two-dimensional distribution diagram reflecting the contact situation between the robot and the wall; the pressure data collection point represents the layout position of the pressure sensor.

[0115] During weld tracking, the control system needs to monitor the contact status between the robot and the wall in real time. Specifically, the control system first acquires the pressure data collected by each pressure sensor of the adsorption device and constructs a continuous pressure distribution field by processing these discrete pressure values ​​through time averaging and spatial interpolation. The control system then normalizes the pressure distribution to eliminate systematic errors between sensors. Based on the normalized pressure data, the control system uses the contour method to generate a pressure distribution cloud map, and combines it with color mapping to intuitively display pressure changes. Finally, threshold segmentation is used to convert the pressure distribution into discrete contact state levels, generating a wall contact state distribution map.

[0116] S212. Identify the abnormal weld contour area according to the contact state distribution diagram, and determine the weld structure type of the abnormal weld contour area.

[0117] Among them, the contact state distribution diagram represents the two-dimensional distribution of the contact conditions between the robot and the wall; the abnormal weld contour area is used to indicate the area where the contact pressure distribution changes significantly; the weld structure type refers to the weld shape with different geometric characteristics and pressure distribution characteristics, such as straight welds, corner welds, lap welds, etc.

[0118] After obtaining the contact state distribution map, the control system needs to analyze the weld structural characteristics. Specifically, the control system first performs morphological processing on the contact state distribution map to extract areas with abnormal pressure distribution changes. The control system then analyzes the geometric characteristics of the abnormal area, including parameters such as area, shape, and direction, while also considering the pressure gradient distribution characteristics within the area. Based on these characteristic parameters, the control system constructs feature vectors and classifies the abnormal areas into different weld structural types using a pattern recognition algorithm. For complex abnormal areas, the control system also analyzes the pressure distribution pattern in the surrounding area to improve the accuracy of structural type identification.

[0119] In some embodiments, the identification of weld structure types can be achieved through a variety of methods: Optionally, the control system can adopt a deep learning-based structural recognition method, using a pre-trained convolutional neural network to extract deep features of the pressure distribution map, identify the weld structure type through a multi-layer classifier, and continuously optimize the recognition model through online learning; Optionally, the control system can adopt a rule-based reasoning-based structural analysis method to establish a weld structure feature knowledge base, determine the structure type through feature matching and logical reasoning, and verify it with historical data. It is understandable that other methods can also be used to achieve structural type recognition, which are not limited here.

[0120] In practical applications, complex weld structures may exhibit multiple types of features, leading to ambiguous structural type identification. To address this, the control system can employ a multimodal feature fusion strategy: First, features are extracted from multiple dimensions, such as pressure distribution, geometry, and temporal variation. A feature importance assessment model is then established, adjusting the weights of different features based on their contribution to structural identification. When conflicting features arise, the system analyzes the reliability of each feature and prioritizes the most reliable. For example, in weld corners, the system considers both pressure mutation characteristics and geometric inflection point features to comprehensively determine the structural type.

[0121] A feature importance assessment model is used for feature fusion in weld structure type identification. During the training phase, the system collects a labeled dataset containing multi-dimensional features such as pressure distribution, geometry, and temporal variations, with each sample annotated with its corresponding weld structure type. A feature weighting network based on an attention mechanism is trained using an objective function that minimizes the structural type identification error. This model uses a multi-channel input structure, accepting feature vectors of different types. A self-attention layer calculates the correlation between each feature and then dynamically assigns weight coefficients based on the feature's recognition contribution. In practice, when new feature data is input, the model first evaluates the reliability of each feature, assigning a higher weight to features with high reliability. The final structural type determination is then made by combining the weighted results of these different features. For example, at weld corners, when both pressure spikes and geometric inflection points are prominent, the model assigns higher weights to these two features while reducing the weight of potentially interfering temporal features, thereby improving structural identification accuracy.

[0122] S213. When the weld structure type changes, update the detection area division scheme.

[0123] Among them, the weld structure type represents the currently identified weld morphological characteristics; the detection area division scheme is used to represent the layout strategy of the image feature extraction area; and the update refers to adjusting the parameters such as the position, size, and shape of the detection area according to the structural changes.

[0124] When a change in weld structure type is detected, the control system needs to promptly adjust its feature extraction strategy. Specifically, the control system first analyzes the impact of the new structure type on feature extraction, including factors such as changes in weld orientation, width, and local feature distribution. The control system then adjusts the parameters of the inspection zones based on these changes: for changes in weld orientation, the orientation and layout of the inspection zones are adjusted; for changes in width, the size and spacing of the inspection zones are modified; and for changes in local feature distribution, the shape and density of the inspection zones are optimized. Furthermore, the control system establishes relationships between inspection zones to ensure the continuity and integrity of the zone divisions.

[0125] In some embodiments, dynamic updating of the detection area can be achieved through a variety of methods: Optionally, the control system can use a prediction-based regional adjustment method to predict the optimal layout of subsequent areas based on the trend of weld structure changes, and achieve gradual adjustment of the detection area through smooth transition while maintaining detection continuity; Optionally, the control system can use a feedback-based adaptive adjustment method to monitor the feature extraction effect in real time, automatically optimize regional parameters when a decrease in feature quality is detected, and find the optimal partitioning solution through iterative optimization. It is understandable that other methods can also be used to achieve detection area updates, which are not limited here.

[0126] S214. Select a corresponding correction execution strategy according to the weld structure type.

[0127] Among them, the weld structure type represents the geometric characteristics of the weld; the correction execution strategy is used to represent the control scheme for the underwater robot's posture adjustment; the corresponding relationship refers to the matching rules between different structure types and corresponding correction strategies; and the control parameters represent the specific execution parameters required to implement the correction action.

[0128] After determining the weld structure type, the control system needs to select an appropriate correction strategy. Specifically, the control system first retrieves a basic strategy template that matches the current structure type from the strategy library. Then, the control system optimizes the parameters of the basic strategy based on real-time pressure distribution data and motion status, including adjusting dynamic parameters such as correction speed, acceleration, and transition time. For complex structures, the control system decomposes the correction process into multiple sub-stages, configures an independent control strategy for each stage, and designs a smooth transition mechanism for strategy switching. Finally, the control system converts the optimized strategy into a specific execution instruction sequence and sends it to the robot's motion control unit.

[0129] In some embodiments, the control system will select an appropriate correction strategy based on the structural characteristics of the weld, that is, the control system will classify the weld structure types into straight welds, fillet welds and lap welds based on the geometric characteristics and pressure distribution characteristics of the abnormal weld contour area; when the weld structure type is a straight weld, a unidirectional progressive correction strategy is adopted, and the robot posture is gradually adjusted according to the deviation direction until it returns to the reference direction; when the weld structure type is a fillet weld, a bidirectional coordinated correction strategy is adopted, and the lateral position and steering angle of the robot are adjusted at the same time to keep the relative position of the detection area and the weld contour stable; when the weld structure type is a lap weld, a step transition correction strategy is adopted, and the correction action is suspended at the weld step, and the correction reference is re-established after the robot crosses the step.

[0130] Among them, the abnormal area of ​​the weld contour indicates the location where the pressure distribution changes; the geometric characteristics are used to represent the shape and size parameters of the weld; the pressure distribution characteristics refer to the numerical characteristics that reflect the contact state; the weld structure types include three basic forms: straight welds, fillet welds and lap welds; the unidirectional progressive correction strategy refers to the control method of gradually adjusting along a single direction; the bidirectional coordinated correction strategy is used to represent the control method of adjusting two degrees of freedom at the same time; the step transition correction strategy refers to the control method of pausing at a specific position and then rebuilding the reference; the reference direction represents the ideal weld tracking trajectory.

[0131] When a change in the weld structure is detected, the control system needs to select an appropriate correction strategy. Specifically, the control system first comprehensively analyzes the geometric characteristics and pressure distribution characteristics of the contour abnormality area, including parameters such as regional shape, area, pressure gradient and distribution symmetry. Based on these characteristic parameters, the control system classifies the weld structure into three basic types. For straight welds, the control system adopts a unidirectional progressive strategy to gradually eliminate deviations through small-step posture adjustments while maintaining the smoothness of the movement. For fillet welds, the control system enables a bidirectional coordination strategy, establishes a coupled control model of lateral position and steering angle, and coordinates the adjustment of the two degrees of freedom to keep the relative position of the detection area stable. For lap welds, the control system executes a step transition strategy, reduces the movement speed when approaching the step position, suspends the correction action until the step is crossed, and then re-collects the baseline data to start a new correction process.

[0132] In some embodiments, structural recognition and correction control can be achieved in a variety of ways: Optionally, the control system can adopt a control strategy generation method based on deep reinforcement learning to establish a state space model that includes weld characteristics, robot status, and environmental factors, output the optimal control action through the policy network, and introduce a reward mechanism to optimize the control effect, and continuously update the policy network according to the execution results; Optionally, the control system can adopt an adaptive control method based on model prediction to construct a dynamic model of weld tracking, optimize the control sequence in the rolling time domain, and continuously update the model parameters through online parameter identification to achieve adaptive control of different structural types. It is understandable that other methods can also be used to achieve structural adaptive control, which is not limited here.

[0133] In practical applications, the transition region of weld structures may contain complex geometric features, leading to instability in structural type determination and strategy switching. To address this, the control system can adopt a hybrid control architecture: first, a state-space representation of multiple basic control strategies is established, and a smooth transition mechanism for strategy switching is designed. When fuzzy structural features or the coexistence of multiple features is detected, the system simultaneously activates multiple related control strategies and generates the final control output through a weighted combination. For example, in the region where a straight weld transitions to a fillet weld, the system gradually reduces the weight of the unidirectional strategy while increasing the weight of the bidirectional strategy to ensure the continuity and stability of the control action. At the same time, the system adjusts the strategy weights based on the execution results to optimize control performance.

[0134] In the embodiments of the present application, due to the use of dynamic filtering parameter adjustment based on local noise intensity, double-layer threshold segmentation, independent feature extraction of multiple detection areas, weighted fitting based on distribution density, dynamic error band outlier screening, as well as posture compensation, water flow analysis and structural adaptation and other technical means, it is possible to achieve adaptive optimization of image processing parameters, complete extraction of weld features, effective isolation of local interference, outstanding contribution of key areas, accurate identification of outliers, active compensation of environmental disturbances and intelligent adaptation of complex structures, effectively solving the problems existing in the prior art such as excessive smoothing caused by fixed parameters, feature loss caused by a single threshold, susceptibility of global features to interference, poor adaptability to environmental changes, and insufficient ability to handle complex structures, thereby achieving high-precision, high-reliability and strong adaptability of weld tracking control in underwater environments, and improving the operation quality and efficiency of underwater welding robots.

[0135] The following describes the control system in the embodiment of the present invention from the perspective of hardware processing. Figure 3 , which is a schematic diagram of the structure of a physical device of the control system in an embodiment of the present application.

[0136] It should be noted that Figure 3 The structure of the control system shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0137] like Figure 3 As shown, the control system includes a CPU 301, which can perform various appropriate actions and processes according to the programs stored in the ROM 302 or the programs loaded from the storage unit 308 into the RAM 303, such as executing the methods described in the above embodiments. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are connected to each other via a bus 304. An I / O interface 305 is also connected to the bus 304.

[0138] The following components are connected to the I / O interface 305: an input section 306 including an audio input device, push button switches, and the like; an output section 307 including a liquid crystal display (LCD), an audio output device, indicator lights, and the like; a storage section 308 including a hard disk and the like; and a communication section 309 including a network interface card such as a LAN (Local Area Network) card or a modem. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to the I / O interface 305 as needed. Removable media 311, such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, is installed in the drive 310 as needed, so that computer programs read from the removable media can be installed in the storage section 308 as needed.

[0139] In particular, according to an embodiment of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, an embodiment of the present invention includes a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 309 and / or installed from the removable medium 311. When the computer program is executed by the CPU 301, the various functions defined in the present invention are performed.

[0140] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present invention. Each box in the flowchart or block diagram can represent a module, program segment, or part of the code, and the above-mentioned module, program segment, or part of the code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings.

[0141] Specifically, the control system of this embodiment includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, the automatic deviation correction method for the underwater robot provided in the above embodiment is implemented.

[0142] As another aspect, the present invention further provides a computer-readable storage medium, which may be included in the control system described in the above embodiments, or may exist independently and not be incorporated into the control system. The storage medium carries one or more computer programs, which, when executed by a processor of the control system, enable the control system to implement the automatic deviation correction method for underwater robot operations provided in the above embodiments.

[0143] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

[0144] As used in the above embodiments, the term “when” may be interpreted to mean “if” or “after” or “in response to determining that” or “in response to detecting that”, depending on the context. Similarly, the phrases “upon determining that” or “if (stated condition or event) is detected” may be interpreted to mean “if determining that” or “in response to determining that” or “upon detecting (stated condition or event)” or “in response to detecting (stated condition or event)”, depending on the context.

Claims

1. A method for automatically correcting the deviation of an underwater robot, characterized in that: Applied to a control system, the method includes: Acquire a weld image captured by an underwater robot camera, and adjust a filter parameter according to the noise intensity of a local image in the weld image to obtain a de-noised image; Performing double-layer threshold segmentation on the denoised image to obtain a binary image; the first layer threshold of the double-layer threshold is used to extract the main contour of the weld, and the second layer threshold of the double-layer threshold is used to retain the characteristics of the transition area of ​​the weld edge; Setting a plurality of detection areas in the binary image, independently extracting edge feature points in each detection area, determining an effective detection area based on the number distribution of feature points in each detection area, and converting the edge feature points in the effective detection area into coordinate data; Determine the weight coefficient according to the distribution density of the edge feature points in the detection area, perform partition weighted fitting on the coordinate data, and obtain the initial weld direction; A dynamic error band is constructed based on the initial weld direction, coordinate points outside the dynamic error band are used as candidate outlier points, a final outlier point among the candidate outlier points is determined in combination with a continuity criterion of adjacent feature points, and the remaining coordinate points outside the final outlier point are refitted to obtain a target weld direction; The target weld direction is compared with the reference direction, and a graded deviation correction control signal is generated according to the degree of deviation.

2. The method according to claim 1, characterized in that The step of performing double-layer threshold segmentation on the denoised image to obtain a binary image specifically includes: Performing grayscale normalization processing on the denoised image to obtain a standard grayscale image; Calculating a grayscale histogram and a cumulative distribution function of the standard grayscale image to obtain grayscale distribution characteristics; Determining a first layer threshold based on the grayscale distribution characteristics, performing a primary segmentation to obtain a weld candidate area; Calculating a second layer threshold according to the grayscale distribution of the weld candidate area, performing secondary segmentation and retaining the edge transition area; The area with a value higher than the first layer threshold is marked as a foreground pixel value, and the area with a value lower than the second layer threshold is marked as a background pixel value, thereby obtaining a binary image.

3. The method according to claim 1, characterized in that The step of determining a weight coefficient according to the distribution density of edge feature points in the detection area, performing partition weighted fitting on the coordinate data, and obtaining an initial weld direction specifically includes: Calculate the horizontal and vertical projection distribution of edge feature points in each detection area to obtain a two-dimensional density distribution matrix; Calculating the concentration of feature points in each region according to the two-dimensional density distribution matrix, and determining a weight coefficient according to the concentration of feature points; Divide the detection area into a number of sub-areas of equal length along the direction of the weld, and apply a corresponding weight coefficient to weight the coordinates of the feature points in each sub-area; Least squares fitting is performed on all weighted feature point coordinates to obtain a space equation of the weld, and a direction vector of the space equation is determined as an initial weld direction.

4. The method according to claim 1, wherein After the step of obtaining a weld image captured by a camera of an underwater robot and adjusting a filter parameter according to the noise intensity of a local image in the weld image to obtain a de-noised image, the method further includes: Obtain the posture data collected by the underwater robot's posture sensor and establish the spatial transformation matrix between the image plane and the weld plane; Performing perspective correction on the weld image according to the spatial transformation matrix to eliminate viewing angle deviation; Extracting the moving bubble trajectory in the weld image, establishing the bubble motion vector field, and determining the direction of water flow disturbance; The image sampling timing is adjusted according to the direction of the water flow disturbance to avoid the peak moment of the water flow disturbance.

5. The method according to claim 4, characterized in that The steps of extracting the moving bubble trajectory in the weld image, establishing the bubble motion vector field, and determining the direction of the water flow disturbance specifically include: Perform inter-frame difference calculation on the continuously acquired multi-frame images to extract the position change of the moving target, and filter the target according to its area and grayscale distribution characteristics to obtain the bubble target; Establishing a time-series tracking sequence for the bubble targets, calculating the motion speed and acceleration of each bubble target, and constructing a local bubble motion vector; Perform spatial interpolation on the image plane for each local bubble motion vector to obtain a continuous bubble motion vector field distribution; Cluster analysis is performed on the bubble motion vector field to extract the direction of water flow disturbance, and the periodic parameters of the water flow disturbance are identified based on the gradient change of the vector field.

6. The method according to claim 1, characterized in that After the step of comparing the target weld direction with the reference direction and generating a graded deviation correction control signal according to the degree of deviation, the method further includes: Obtain contact pressure data of the robot adsorption device and generate a wall contact state distribution map; Identify the weld contour abnormal area according to the contact state distribution map, and determine the weld structure type of the weld contour abnormal area; When the weld structure type changes, updating the detection area division scheme; Select the corresponding correction execution strategy according to the weld structure type.

7. The method according to claim 6, characterized in that The step of selecting a corresponding correction execution strategy according to the weld structure type specifically includes: According to the geometric characteristics and pressure distribution characteristics of the abnormal weld contour area, the weld structure type is classified into straight weld, fillet weld and lap weld; When the weld structure type is a straight weld, a one-way progressive deviation correction strategy is adopted to gradually adjust the robot posture according to the deviation direction until it returns to the reference direction; When the weld structure type is a fillet weld, a two-way coordinated correction strategy is adopted to simultaneously adjust the lateral position and steering angle of the robot to maintain a stable relative position between the detection area and the weld contour; When the weld structure type is a lap weld, a step transition correction strategy is adopted, and the correction action is suspended at the weld step, and the correction reference is re-established after the robot crosses the step.

8. A control system, characterized in that: The control system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to enable the control system to execute the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium comprising instructions, characterized in that: When the instructions are executed on a control system, the control system is caused to execute the method according to any one of claims 1 to 7.

10. A computer program product, characterized in that When the computer program product is run on a control system, the control system is caused to execute the method according to any one of claims 1 to 7.

Images