An acoustic visual positioning laser cutting method for underwater unmanned underwater vehicles

Abstract

The present application relates to underwater laser cutting technical field, especially in a kind of acoustic vision positioning laser cutting method for underwater unmanned underwater vehicle, the multi-source data of typical target under water is synchronously collected, and the multi-source data is preprocessed;Construct the acoustic vision fusion positioning model based on joint error minimization, utilize the acoustic vision fusion positioning model, output the three-dimensional space accurate pose of target to be cut;Flexible connection mechanical arm configuration planning and drive control module are constructed, and the space configuration suitable for underwater different depth operation is planned, and the collaborative control instruction of drive motor is mapped generation;Through constructing the acoustic vision fusion positioning model based on joint error minimization, the long-distance detection capability of sonar and the high-resolution identification capability of vision are fused, the limitation of single sensor in underwater environment is overcome, the high-precision three-dimensional space positioning of target to be cut is realized, and accurate space reference is provided for subsequent laser cutting operation.

Description

Technical Field

[0001] This invention relates to the field of underwater laser cutting technology, and in particular to a laser cutting method for acoustic and visual positioning of underwater unmanned submersibles. Background Technology

[0002] Underwater special operations technology is not only an important means of maintaining marine infrastructure, but also an effective way to salvage sunken ships and destroy underwater targets. Due to limitations such as the extreme underwater environment and unstructured physical space, complex waters suffer from extreme turbidity, drastic light attenuation, and dense and unpredictable obstacles, resulting in traditional underwater operation equipment being less effective in approaching targets, achieving high-precision positioning, and ensuring safe operation. Therefore, automated precision operation technology mounted on unmanned underwater vehicles (UUVs) is becoming an important research direction. This technology is of great significance for promoting the efficient development of marine resources and ensuring the safety of deep-sea engineering. With the rapid development of marine resource development and underwater engineering construction, the application of unmanned underwater vehicles (UUVs) in areas such as subsea pipeline maintenance, underwater structure dismantling, and marine scientific research is becoming increasingly widespread. In these operational scenarios, precise cutting is one of the key operational steps.

[0003] In recent years, research on underwater special operations technology has mainly focused on multi-sensor fusion perception and robotic arm control. Exploration and research on underwater robots equipped with robotic arms to perform precision cutting operations have developed rapidly, with in-depth research into the application of laser cutting technology. Although underwater precision cutting tasks present unique challenges, such as the intense absorption of laser energy by the medium, localized blindness caused by cutting debris, and interference from water flow disturbances, the characteristics of novel sensing and actuator mechanisms make them potential effective tools for solving these problems. However, existing underwater operation systems mostly use traditional rigid multi-joint robotic arms, which are prone to collisions and interference in confined, narrow spaces. Simultaneously, single-vision positioning is easily ineffective in turbid water, and single-sonar positioning resolution cannot meet the millimeter-level requirements of laser cutting. Furthermore, static cutting strategies cannot compensate for end-effector defocusing errors caused by water flow in real time. Summary of the Invention

[0004] The technical problems solved by this invention are: the failure of single sensing and positioning in complex underwater environments, the easy collision and interference of robotic arms in confined and narrow spaces, and the easy loss of focus at the end of the dynamic cutting process in related technologies.

[0005] To solve the above-mentioned technical problems, the present invention provides a laser cutting method for acoustic-visual positioning of underwater unmanned submersibles, comprising: Step S1: Simultaneously collect multi-source data of typical underwater targets and preprocess the multi-source data; Step S2: Construct an acoustic-visual fusion localization model based on minimizing joint error, and use the acoustic-visual fusion localization model to output the precise three-dimensional spatial pose of the target to be cut. Step S3: Construct a flexible connection robotic arm configuration planning and drive control module, plan a spatial configuration suitable for underwater operations at different depths, and map and generate collaborative control commands for the drive motors; Step S4: Construct a compliant obstacle avoidance and cutting posture compensation mechanism based on state feedback to perform real-time posture compensation for the small end deformation caused by water flow disturbance. Step S5: Determine the coordinates of the center of the molten pool generated by the laser action through acoustic and visual fusion positioning, start the end-effector integrated laser, and complete the laser cutting operation.

[0006] As a preferred embodiment of the acoustic-visual positioning laser cutting method for underwater unmanned submersibles described in this invention, the method involves: simultaneously sampling typical underwater targets using forward-looking sonar and binocular vision camera to acquire the multi-source data; The multi-source data includes sonar metadata and binocular vision images.

[0007] As a preferred embodiment of the acoustic-visual positioning laser cutting method for underwater unmanned submersibles described in this invention, the preprocessing of the multi-source data includes: Step S11, set the forward-looking sonar sampling frequency to The sampling frequency of the binocular vision camera is And satisfy , 'n' represents the ratio between the sampling frequency of the binocular vision camera and the sampling frequency of the forward-looking sonar, where the deviation between the sonar point cloud data and the visual image frame on the time axis is less than a preset threshold. ; Step S12: Extract polar coordinate information from sonar metadata. Transform it into the carrier coordinate system using the transformation formula. Cartesian coordinates below Its mathematical expression is: , The sonar point pose is projected onto the global map coordinate system according to the spatial pose projection formula. The mathematical expression for the spatial pose projection formula is: , in, To detect distance, It is the azimuth angle. The pitch angle, A point vector in the carrier coordinate system. A point vector in the world coordinate system. It is an extrinsic parameter matrix. Here is the pose transformation matrix; Sonar images are acquired frame by frame, and all sonar images are uniformly scaled to a specific size using a bilinear interpolation algorithm. Pixel resolution, combined with synchronously acquired binocular vision images, target bounding boxes are labeled on the scaled sonar images to construct a sonar image detection training set containing location information and category labels; Step S13: Extract the precise pixel-level mask of the target using a region growing model. ,in To represent the target region, a one-to-one correspondence between sonar images and mask annotations is established, and a mask annotation set is constructed. During the training phase of the region growing model, the mask annotation set is input into the hierarchical attention block of the feature fusion network of the region growing model, and the cross-entropy loss between the predicted mask and the ground truth mask is calculated. Its mathematical expression is: , in, Here, N represents the mask loss value, and N is the total number of pixels in the current image. For the first The real label of each pixel The first prediction for the region growth model The probability value of a pixel belonging to the target object.

[0008] As a preferred embodiment of the acoustic-visual positioning laser cutting method for underwater unmanned submersibles described in this invention, step S2 includes: Step S21: Construct a pose transformation model based on Lie algebras. Define the relative pose between the end effector of the flexible connecting robotic arm and the target to be cut as the transformation matrix, the mathematical expression of which is: , in, Let be the pose transformation matrix to be solved. It is a six-dimensional vector, where the first three dimensions represent translation offsets and the last three dimensions represent rotation vectors. This means converting a Lie algebra into a Lie group through an exponential mapping. for Rotational torque, for Translation vector; Step S22: Acquire the two-dimensional pixel coordinate observations of the target under binocular vision. Compared with polar coordinate observations in the sonar coordinate system Construct a joint least squares optimization function Its mathematical expression is: , in, , These represent the three-dimensional spatial positions of the feature points on the surface of the target to be cut in the world coordinate system. For the first The pixel coordinates of each visual feature point on the imaging plane of the stereo camera. This is a visual projection function that projects 3D points into 2D pixel coordinates based on camera intrinsic parameters. For the first Observations of a sonar feature point within the sonar detection range. This is a sonar observation mapping function that maps 3D points to sonar polar coordinates, serving as a local 3D visual point cloud of the target. and These are acoustic and visual observation information matrices, used to characterize the confidence levels of different sensors; Step S23: Perform iterative optimization based on the Levenberg-Marquardt algorithm for the objective function. Performing a first-order Taylor expansion yields information about the pose increment. The linearized equations are derived, and a damping factor is introduced. The mathematical expression for iterative solution is: , in, To comprehensively observe residuals, Let be the Jacobian matrix, representing the partial derivatives of the error term with respect to the pose transformation increment. It is the identity matrix. The damping factor, The pose update increment is based on the initial pose and the pose update increment of the cutting target. Output the precise pose of the target to be cut .

[0009] As a preferred embodiment of the acoustic-visual positioning laser cutting method for underwater unmanned submersibles described in this invention, step S3 includes: Step S31, set the flexible connection robotic arm to be... A kinematic model based on standard DH parameters is established by cascading segments of a continuum, and the kinematics of the first segment are extracted. Link length, link torsion angle, joint offset, and real-time joint rotation angle of each joint. The first is calculated by homogeneous transformation matrix. Local spatial pose transformation of each joint relative to the previous joint The global positive kinematics mapping matrix from the base of the flexible connecting robotic arm to the end effector The mathematical expression is: , Step S32: Obtain the precise pose of the target to be cut. Combining underwater obstacle avoidance and joint physical constraints, an inverse kinematics objective optimization function is constructed. The optimal joint space configuration for reaching the work surface is calculated, and its mathematical expression is: , in, This is a set of joint angle parameters for the entire arm. and These are the current 3D position vector and rotation matrix of the end effector, respectively. and These are the desired position and rotation matrix, respectively. , , These are the weighting coefficients; The singularity and joint deviation penalty term is used to constrain the smooth movement of the robotic arm in a confined underwater space to avoid interference between the joints; Step S33: Iteratively solve the objective optimization function using the damped least squares method. Output the optimal set of joint angle parameters The optimal joint angle parameter set The underlying servo motors of each rotary joint are mapped in real time to generate collaborative control commands for the drive motors. These collaborative control commands are then used to guide the end laser to align with the surface to be cut.

[0010] In a preferred embodiment of the acoustic-visual positioning laser cutting method for underwater unmanned submersibles described in this invention, step S4 includes: An external collision torque observation model based on flexible joints is constructed. Flexible connection components are installed at each rotary joint of the robotic arm. The external collision torque is then extracted. Real-time angle of the encoder on the motor side at each joint Real-time angle with the connecting rod side encoder , and They are respectively , The first derivative; based on the equivalent joint stiffness of the flexible connection component. With damping Calculate the real-time elastic internal torque of each joint. Its mathematical expression is: , Construct an external environment disturbance torque observer based on generalized momentum to calculate the actual external collision torque experienced by each joint. When there is When the collision threshold exceeds the preset safety threshold, it is determined that the corresponding link has a rigid interference collision with an underwater unstructured obstacle.

[0011] As a preferred embodiment of the acoustic-visual positioning laser cutting method for underwater unmanned submersibles described in this invention, wherein: after determining that a rigid interference collision has occurred, the current Jacobian matrix of the robotic arm is used... The external collision moment matrix Mapped to equivalent external contact force at the end This triggers a compliant obstacle avoidance law based on admittance control, establishing a force-displacement response equation, the mathematical expression of which is: , in, , , These are the defined desired mass matrix, damping matrix, and stiffness matrix, respectively. The amount of obstacle avoidance posture retraction generated to adapt to external collision forces. , They are respectively The first and second derivatives; The robotic arm performs the obstacle avoidance pose retraction. When the joint undergoes flexible deflection due to water flow disturbance, the pseudo-inverse of the Jacobian matrix is ​​utilized. The mathematical expression for the inverse kinematics compensation solution is as follows: , in, This is the angle compensation vector that needs to be added for the remaining joints that are not subject to collision constraints. To maintain the target's precise pose The required desired end-effector change will be sent to the underlying actuators of each joint in real time after compensation of the angle parameters.

[0012] As a preferred embodiment of the acoustic-visual positioning laser cutting method for underwater unmanned submersibles described in this invention, wherein: Step S5 includes: Step S51: Extract the surface unit normal vector at each interpolation point s on the trajectory to be cut in the local three-dimensional visual point cloud of the target. Establish the desired spatial pose matrix of the end-effector laser cutting head. , Describe the three-dimensional special Euclidean group, its positional components The rotational component of the three-dimensional cutting curve that fits the plan Subject to pose alignment constraints, its mathematical expression is: , in, This is the optical axis direction vector of the laser beam emitted from the end effector coordinate system; Step S52: Control the flexible connecting robotic arm to move the end effector to the cutting start point, prioritize the activation of the high-pressure jet device arranged around the laser emitter to displace the turbid water, activate the fiber laser, and dynamically compensate and control the output power based on the underwater laser transmission attenuation model. Its lasing power output function... The mathematical expression is: , , in, The constant cutting threshold energy required to melt through the target material. The real-time working focal length from the terminal exit port to the target surface. For wavelength The total attenuation coefficient of the laser in water; The inherent absorption coefficient of the water body. The dynamic scattering coefficient caused by the suspended slag and bubbles generated during cutting is determined by the function. Real-time adjustment of laser output to compensate for energy loss caused by focal length fluctuations and sudden changes in local water turbidity; Step S53: Use the end-effector binocular vision camera to capture in real time the normalized pixel coordinates of the center of the molten pool generated by the laser on the image plane. The projection point of the desired cutting trajectory on the image Visual feature error The mathematical expression for the image-based visual servoing feedforward-feedback control law is as follows: , in, For space velocity spinor, For proportional control gain, The expected feedforward cutting speed for moving forward along the planned trajectory; The spatial velocity spinor is the Moore-Ponros generalized inverse of the image Jacobian matrix. Mapped to the control angular velocity of each joint Its mathematical expression is: , The control angular velocity closed-loop drive is used to drive each joint of the flexible connected robotic arm to eliminate cutting trajectory offset in real time until the laser cutting operation is completed.

[0013] The beneficial effects of the present invention are as follows: (1) The present invention constructs an acoustic-visual fusion positioning model based on the minimization of joint error, which integrates the long-range detection capability of sonar and the high-resolution recognition capability of vision, overcomes the limitations of a single sensor in the underwater environment, and realizes high-precision three-dimensional spatial positioning of the target to be cut, providing a precise spatial reference for subsequent laser cutting operations. (2) The present invention constructs a flexible connection robotic arm configuration planning and drive control module, and combines the DH parameter kinematic model and inverse kinematic optimization solution to realize the smooth motion planning and precise control of the robotic arm in the narrow underwater space, avoiding interference between joints and improving the flexibility and adaptability of the operation. (3) The present invention constructs a compliant obstacle avoidance and cutting posture compensation mechanism based on state feedback, uses flexible connection components and torque observers to detect external collisions and water flow disturbances in real time, and realizes compliant response and posture correction through admittance control and inverse kinematic compensation, ensuring the safety and cutting accuracy of the operation and avoiding equipment damage. (4) This invention uses a high-pressure jet device to discharge turbid water, combines an underwater laser transmission attenuation model to dynamically compensate and control the output power, and uses visual servo control to correct the cutting trajectory in real time, which significantly improves the accuracy, efficiency and quality of underwater laser cutting and can adapt to complex and ever-changing underwater operating environments. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0015] Figure 2 This is a schematic diagram of the overall hardware structure of the present invention.

[0016] Figure 3 This is a partially enlarged view of the end effector structure of the present invention.

[0017] Figure 4 This is a block diagram of the admittance obstacle avoidance and dynamic pose compensation control of the present invention.

[0018] Figure 5 The diagram shows the dynamic power compensation curves of this invention, where (a) is the laser cutting trajectory tracking error curve and (b) is the laser dynamic power compensation curve.

[0019] Among them, 1. Underwater unmanned submersible, 2. Robotic arm, 3. End effector, and 4. Laser cutting head. Detailed Implementation

[0020] Example 1, referring to Figure 1 As an embodiment of the present invention, a laser cutting method for acoustic-visual positioning of an underwater unmanned submersible is provided, comprising: Step S1: Simultaneously collect multi-source data of typical underwater targets and preprocess the multi-source data; use forward-looking sonar and binocular vision camera to simultaneously sample typical underwater targets and obtain multi-source data; the multi-source data includes sonar metadata and binocular vision images.

[0021] In this embodiment, the multi-source data specifically includes sonar metadata and binocular vision images; wherein, the sonar metadata refers to the acoustic reflection information of underwater targets acquired through forward-looking sonar, including polar coordinate information (range r, azimuth θ, pitch angle). ( ) and sonar images; binocular vision images refer to pairs of optical images of underwater targets acquired through binocular vision cameras, including left and right views.

[0022] Preprocessing of multi-source data includes: Step S11, set the forward-looking sonar sampling frequency to The sampling frequency of the binocular vision camera is And satisfy Time synchronization mechanisms, achieved through hardware-triggered or IMU-based pre-integration timestamp alignment, are techniques used to ensure that sonar and visual data are synchronized at the time of acquisition. This eliminates time discrepancies caused by differences in sensor sampling frequencies. The goal is to ensure that the time-axis deviation between sonar point cloud data and visual image frames is less than a preset threshold. Preset threshold The typical value of t ranges from 10ms to 50ms, depending on the sensor sampling frequency and the required operational accuracy.

[0023] In this embodiment, the time synchronization mechanism specifically includes hardware synchronization and software synchronization. Hardware synchronization refers to simultaneously activating sonar and camera data acquisition via an external trigger signal. Software synchronization refers to matching data frames based on the timestamps of each sensor, selecting frames with time deviations less than a threshold. t is a data pair.

[0024] Step S12: Extract polar coordinate information from sonar metadata. Transform it into the carrier coordinate system using the transformation formula. Cartesian coordinates below Its mathematical expression is: ; In this embodiment, coordinate transformation refers to the process of converting measured values ​​in the sensor coordinate system to a unified reference coordinate system, which is used to achieve spatial registration of multi-source data. The coordinate transformation receives sonar polar coordinate data, and after calculation of the transformation matrix, outputs three-dimensional point cloud data in the carrier coordinate system and the global map coordinate system, which is used in conjunction with visual data for fusion positioning.

[0025] The sonar point pose is projected onto the global map coordinate system according to the spatial pose projection formula. The mathematical expression for the spatial pose projection formula is: ; in, For detection range, it represents the spatial distance corresponding to the time difference between the sonar transmitting sound waves and receiving the echo; The azimuth angle represents the angle of offset of the sonar beam on the horizontal plane. The elevation angle represents the angle of offset of the sonar beam in the vertical plane. It is a point vector in the carrier coordinate system. That is, the vector calculated by the above formula. Representation in the robot's body coordinate system; It is a point vector in the world coordinate system, that is, the final absolute coordinates of the target in the global map, which is the direct basis for laser cutting path planning; This is the extrinsic parameter matrix. It describes the specific position and rotation angle of the sonar sensor mounted on the unmanned underwater vehicle; This is the pose transformation matrix. It describes the real-time position and attitude of the unmanned underwater vehicle underwater.

[0026] Sonar images are acquired frame by frame, and all sonar images are uniformly scaled to a specific size using a bilinear interpolation algorithm. By combining pixel resolution with synchronously acquired binocular vision images, target bounding boxes are annotated on the scaled sonar images to construct a sonar image detection training set containing location information and category labels.

[0027] In this embodiment, raw sonar images and visual images are received, scaled, and labeled to output a standardized training dataset, which is used in conjunction with a region growing model for target detection training. Image preprocessing specifically includes image scaling and target labeling; wherein, image scaling uses a bilinear interpolation algorithm to unify sonar images of different resolutions to 1280×720 pixels; target labeling uses bounding box annotation, and the annotation content includes target location and category label.

[0028] Step S13: Extract the precise pixel-level mask of the target using a region growing model. ,in The target region is represented, a one-to-one correspondence between the sonar image and the mask annotation is established, and a mask annotation set is constructed.

[0029] During the training phase of the region growing model, the mask annotation set is input into the hierarchical attention block of the feature fusion network of the region growing model, and the cross-entropy loss between the predicted mask and the ground truth mask is calculated. Its mathematical expression is: ; in, is the mask loss value, used to evaluate the performance of region growing on pixel-level classification. The smaller the value, the closer the shape predicted by the model is to the real target. N is the total number of pixels in the current image. For the first The real label of each pixel The first prediction of the region growth model The probability value of a pixel belonging to the target object.

[0030] In this embodiment, mask extraction refers to an image processing technique that precisely segments the target region from an image to obtain pixel-level contour information of the target. Mask extraction receives a sonar image detection training set, processes it through a region growing model, and outputs a precise mask of the target, which works in conjunction with the fusion localization module for target pose estimation. The mask M is a two-dimensional matrix with dimensions H×W (image height × image width), and its elements are either 0 or 1; where M... u,v =1 indicates that the pixel at coordinates belongs to the target region, M u,v =0 indicates that the pixel at coordinates belongs to the background area. Data is acquired synchronously from sonar and visual sensors. Through steps such as time synchronization, coordinate transformation, image scaling, and mask extraction, the raw data is converted into standardized fusion positioning input data. This eliminates data incompatibility issues caused by factors such as differences in sensor sampling frequencies, inconsistent coordinate systems, and inconsistent image sizes, laying a data foundation for subsequent high-precision fusion positioning.

[0031] Step S2: Construct an acoustic-visual fusion localization model based on minimizing joint error. Using this model, output the precise 3D spatial pose of the target to be cut, including: Step S21: Construct a pose transformation model based on Lie algebras. Define the relative pose between the end effector of the flexible connecting robotic arm and the target to be cut as the transformation matrix, the mathematical expression of which is: ; in, Let be the pose transformation matrix to be solved. It is a six-dimensional vector, where the first three dimensions represent translation offsets and the last three dimensions represent rotation vectors. This means converting a Lie algebra into a Lie group through an exponential mapping. for Rotational torque, for Translation vector.

[0032] In this embodiment, the pose transformation model refers to a mathematical model describing the position and attitude changes of a rigid body in three-dimensional space, used to represent the transformation relationship between the sensor coordinate system, the carrier coordinate system, and the target coordinate system. The pose transformation model receives preprocessed multi-source data, and outputs a pose transformation matrix after Lie algebra parameterization and Lie group transformation, which is used in conjunction with the joint optimization function for pose estimation.

[0033] Step S22: Acquire the two-dimensional pixel coordinate observations of the target under binocular vision. Compared with polar coordinate observations in the sonar coordinate system Construct a joint least squares optimization function Its mathematical expression is: ; in, , This represents the three-dimensional spatial position of the feature points on the surface of the target to be cut in the world coordinate system. For the first The pixel coordinates of each visual feature point on the imaging plane of the stereo camera. This is a visual projection function that projects 3D points into 2D pixel coordinates based on camera intrinsic parameters. For the first Observations of a sonar feature point within the sonar detection range. This is a sonar observation mapping function that maps 3D points to sonar polar coordinates, serving as a local 3D visual point cloud of the target. and This is an acoustic-visual observation information matrix used to characterize the confidence levels of different sensors.

[0034] In this embodiment, the joint least squares optimization function refers to the objective function that minimizes the error of fusing sonar and visual observation data, and is used to solve for the optimal pose estimation. The joint least squares optimization function receives sonar observations zs and visual observations zv, calculates the error, and outputs the total error value after weighted summation. It is used in conjunction with the Levenberg-Marquardt algorithm for iterative optimization.

[0035] Step S23: Perform iterative optimization based on the Levenberg-Marquardt algorithm for the objective function. Performing a first-order Taylor expansion yields information about the pose increment. The linearized equations are derived, and a damping factor is introduced. The mathematical expression for iterative solution is: ; in, This is to synthesize the observation residuals, which are the deviations between the observed and predicted values; is the Jacobian matrix, representing the partial derivative of the error term with respect to the pose change increment, which determines the direction of searching for the optimal pose gradient; It is the identity matrix. This is the damping factor, used to balance the convergence rates of the Gauss-Newton method and the gradient descent method. The pose update increment is based on the initial pose and the pose update increment of the cutting target. Output the precise pose of the target to be cut .

[0036] In this embodiment, the Levonburg-Marquardt algorithm receives the joint least squares optimization function, and outputs the optimal pose estimate after iterative solution. This estimate is then used in conjunction with the subsequent robotic arm control module for cutting path planning. The acoustic-visual fusion positioning module effectively integrates the long-range detection capability of sonar and the high-resolution recognition capability of vision by constructing a joint error minimization model. Utilizing Lie algebra parameterization and iterative optimization with the Levonburg-Marquardt algorithm, it achieves high-precision three-dimensional pose estimation of underwater targets, solving the problem of insufficient positioning accuracy of a single sensor in underwater environments.

[0037] Step S3: Construct a flexible connected robotic arm configuration planning and drive control module, plan a spatial configuration suitable for underwater operations at different depths, and map and generate coordinated control commands for the drive motors, including: Step S31, set the flexible connection robotic arm to be... A kinematic model based on standard DH parameters is established by cascading segments of a continuum, and the kinematics of the first segment are extracted. Link length, link torsion angle, joint offset, and real-time joint rotation angle of each joint. The first is calculated by the homogeneous transformation matrix. Local spatial pose transformation of each joint relative to the previous joint The mathematical expression for the global positive kinematics mapping matrix from the base of the flexible connected robotic arm to the end effector is: .

[0038] Step S32: Obtain the precise pose of the target to be cut. Combining underwater obstacle avoidance and joint physical constraints, an inverse kinematics objective optimization function is constructed. The optimal joint space configuration for reaching the work surface is calculated, and its mathematical expression is: ; in, , represents the set of joint angle parameters for the entire arm. and Given the current 3D position vector and rotation matrix of the end point, and Given the desired position and rotation matrix, , , These are the weighting coefficients. The singularity and joint deviation penalty term is used to constrain the smooth movement of the robotic arm in a confined underwater space to avoid interference between the joints.

[0039] In this embodiment, the inverse kinematics objective optimization function refers to the optimization objective of solving the joint angles based on the desired end-effector pose, used to plan the optimal spatial configuration of the robotic arm. The inverse kinematics objective optimization function receives the target pose Tct* and constraints, and outputs the optimal joint angle parameters after optimization, which are used in conjunction with the drive control module for the motion control of the robotic arm.

[0040] Step S33: Iteratively solve the objective optimization function using the damped least squares method. Output the optimal set of joint angle parameters The optimal set of joint angle parameters The underlying servo motors of each rotary joint are mapped in real time to generate collaborative control commands for the drive motors. These commands are then used to guide the end laser to align with the surface to be cut.

[0041] In this embodiment, by establishing an accurate DH parameter kinematic model and combining inverse kinematics optimization and damped least squares method, smooth motion planning and precise control of the flexible connecting robotic arm in narrow underwater spaces are achieved, effectively avoiding interference and singularity problems between joints and improving the flexibility and safety of the operation.

[0042] Step S4: Construct a compliant obstacle avoidance and cutting pose compensation mechanism based on state feedback to perform real-time attitude compensation for minor end-effector deformations caused by water flow disturbances, including: An external collision torque observation model based on flexible joints is constructed. Flexible connection components are installed at each rotary joint of the robotic arm. The external collision torque is then extracted. Real-time angle of the encoder on the motor side at each joint Real-time angle with the connecting rod side encoder Based on the equivalent joint stiffness of the flexible connection component With damping Calculate the real-time elastic internal torque of each joint. Its mathematical expression is: ; Construct an external environment disturbance torque observer based on generalized momentum to calculate the actual external collision torque experienced by each joint. When there is When the collision threshold exceeds the preset safety threshold, it is determined that the corresponding link has a rigid interference collision with an underwater unstructured obstacle.

[0043] In this embodiment, flexible connecting joints are set at each rotary joint of the robotic arm, and the motor output shaft and connecting rod are connected by elastic elements (such as springs, flexible couplings, etc.), and collision detection is achieved in conjunction with the torque observer.

[0044] After determining that a rigid interference collision has occurred, the current Jacobian matrix of the robotic arm is used. The external collision moment matrix Mapped to equivalent external contact force at the end This triggers a compliant obstacle avoidance law based on admittance control, establishing a force-displacement response equation, the mathematical expression of which is: ; in, These are the defined desired mass matrix, damping matrix, and stiffness matrix, respectively. The obstacle avoidance posture retreat amount generated to adapt to external collision force; thereby driving the joint that has collided to produce compliant reverse retreat to avoid damage to the mechanical structure.

[0045] The robotic arm retracts its obstacle avoidance posture. When the joints flexibly deflect due to water flow disturbance, to ensure that the laser cutting focus at the end of the robotic arm does not detach from the target working surface, the pseudo-inverse of the Jacobian matrix is ​​utilized. The mathematical expression for the inverse kinematics compensation solution is as follows: ; in, This is the angle compensation vector that needs to be added for the remaining joints that are not subject to collision constraints. To maintain the target's precise pose The required desired end effector change is sent to the underlying actuator of each joint in real time after compensation, ensuring that the end effector attitude is always locked to the cutting coordinates.

[0046] In this embodiment, external collisions and water flow disturbances are detected in real time by using flexible connecting joints and torque observers. Admittance control is used to achieve compliant obstacle avoidance response, and inverse kinematics compensation is used to eliminate posture deviations, effectively ensuring the operational safety and cutting accuracy of the robotic arm in complex underwater environments.

[0047] Step S5: Determine the coordinates of the center of the molten pool generated by the laser through acoustic-visual fusion positioning, activate the end-effector integrated laser, and complete the laser cutting operation, including: Step S51: Extract the surface unit normal vector at each interpolation point s on the trajectory to be cut in the local 3D visual point cloud of the target. Establish the desired spatial pose matrix of the end-effector laser cutting head. its positional components The rotational component of the three-dimensional cutting curve that fits the plan Subject to pose alignment constraints, its mathematical expression is: ; in, This is the direction vector of the laser beam's output direction in the end effector coordinate system. The pose alignment constraint forces the laser beam axis to remain perpendicular to the target surface or maintain a preset optimal cutting angle during the moving and cutting process to maximize laser energy absorption. The cutting trajectory planning receives the local 3D visual point cloud of the target, performs trajectory interpolation and normal calculation, and outputs the desired spatial pose sequence, which works in conjunction with the robotic arm control module to achieve cutting path tracking.

[0048] Step S52: Control the flexible connecting robotic arm to move the end effector to the cutting start point, prioritize the activation of the high-pressure jet device arranged around the laser emitter to displace the turbid water, activate the fiber laser, and dynamically compensate and control the output power based on the underwater laser transmission attenuation model. Its lasing power output function... The mathematical expression is: ; ; in, The constant cutting threshold energy required to melt through the target material. The real-time working focal length from the terminal exit port to the target surface. For wavelength The total attenuation coefficient of the laser in water; The inherent absorption coefficient of the water body. The dynamic scattering coefficient caused by the suspended slag and bubbles generated during cutting is expressed by a function. The laser output is adjusted in real time to compensate for energy loss caused by focal length fluctuations and sudden changes in local water turbidity; the high-pressure jet device is arranged around the laser output head and is activated before the cutting operation to achieve efficient cutting in conjunction with the laser.

[0049] Step S53: Finally, high-precision closed-loop cutting based on image Jacobian matrix is ​​performed using visual servoing. During continuous cutting, the normalized pixel coordinates of the center of the molten pool generated by the laser on the image plane are captured in real time using a binocular vision camera at the end. The projection point of the desired cutting trajectory on the image Visual feature error The mathematical expression for the image-based visual servoing feedforward-feedback control law is as follows: ; in, For space velocity spinor, For proportional control gain, The expected feedforward cutting speed for moving forward along the planned trajectory; The Moore-Ponous generalized inverse of the image Jacobian matrix, with the spatial velocity spinor Mapped to the control angular velocity of each joint Its mathematical expression is: ; By using a closed-loop control system to drive the flexible joints of the robotic arm, the cutting trajectory offset is eliminated in real time until the laser cutting operation is completed.

[0050] In this embodiment, the laser cutting execution module improves the local water environment through a high-pressure jet device, achieves dynamic power compensation by combining an underwater laser transmission attenuation model, and uses visual servo control to correct the cutting trajectory in real time, which significantly improves the accuracy, efficiency and quality of underwater laser cutting.

[0051] Example 2, refer to Figures 2 to 5 This is another embodiment of the present invention, based on the previous embodiment. The difference between this embodiment and the previous embodiment is that it specifically focuses on close-range laser cutting operations in an underwater unstructured environment. This method relies on... Figure 2 and Figure 3 The hardware physical architecture shown is used for underwater deployment. Figure 3 In the diagram, 1 represents an underwater unmanned submersible, 2 represents a robotic arm, 3 represents an end effector, and 4 represents a laser cutting head.

[0052] In the initial stage of the operation, the underwater unmanned vehicle used the M750d, employing its forward-looking active sonar combined with a self-developed sonar dataset to detect and locate distant targets. Upon entering the close-range cutting phase, facing unknown underwater rigid obstacles and sudden strong current disturbances, the system immediately triggered [measures such as...]. Figure 4 The admittance obstacle avoidance and dynamic pose compensation control are shown.

[0053] Taking a complex underwater dynamic disturbance environment as an example, the final closed-loop control and cut-off compensation results are as follows: Figure 5 As shown, the combined curve graphs sequentially record the time series nodes of the entire operation: combined with Figure 5 (a) It can be seen that a sudden physical collision occurred at the 2nd second of the operation, causing the trajectory error to increase instantaneously. However, thanks to the pose compensation algorithm of this invention, the error curve exhibits a rapid exponential decay and reconverges to zero within the following 2 seconds; combined with Figure 5(b) It can be seen that when the submersible enters heavily turbid water between the 4th and 8th second, the laser output power achieves adaptive smooth ascent and stable descent based on the underwater light attenuation model and coaxial clearing feedback. This demonstrates that the present invention exhibits extremely high robustness and high precision in dynamic cutting operations under conditions of sudden underwater physical collisions and extremely turbid small targets, thereby significantly improving the accuracy and system robustness of unmanned underwater vehicles performing laser servo cutting operations under dynamic disturbances.

Claims

1. A laser cutting method for acoustic-visual positioning of underwater unmanned submersibles, characterized in that, include: Step S1: Simultaneously collect multi-source data of typical underwater targets and preprocess the multi-source data; Step S2: Construct an acoustic-visual fusion localization model based on minimizing joint error, and use the acoustic-visual fusion localization model to output the precise three-dimensional spatial pose of the target to be cut. Step S3: Construct a flexible connection robotic arm configuration planning and drive control module, plan a spatial configuration suitable for underwater operations at different depths, and map and generate collaborative control commands for the drive motors; Step S4: Construct a compliant obstacle avoidance and cutting posture compensation mechanism based on state feedback to perform real-time posture compensation for the small end deformation caused by water flow disturbance. Step S5: Determine the coordinates of the center of the molten pool generated by the laser action through acoustic and visual fusion positioning, start the end-effector integrated laser, and complete the laser cutting operation.

2. The method as described in claim 1, characterized in that, The forward-looking sonar and binocular vision camera are used to simultaneously sample typical underwater targets to obtain the multi-source data; the multi-source data includes sonar metadata and binocular vision images.

3. The method as described in claim 2, characterized in that, Preprocessing the multi-source data includes: Step S11, set the forward-looking sonar sampling frequency to The sampling frequency of the binocular vision camera is And satisfy , 'n' represents the ratio between the sampling frequency of the binocular vision camera and the sampling frequency of the forward-looking sonar, where the deviation between the sonar point cloud data and the visual image frame on the time axis is less than a preset threshold. ; Step S12: Extract polar coordinate information from sonar metadata. Transform it into the carrier coordinate system using the transformation formula. Cartesian coordinates below Its mathematical expression is: ， The sonar point pose is projected onto the global map coordinate system according to the spatial pose projection formula. The mathematical expression for the spatial pose projection formula is: , in, To detect distance, It is the azimuth angle. The pitch angle, A point vector in the carrier coordinate system. A point vector in the world coordinate system. It is an extrinsic parameter matrix. Here is the pose transformation matrix; Sonar images are acquired frame by frame, and all sonar images are uniformly scaled to a specific size using a bilinear interpolation algorithm. Pixel resolution, combined with synchronously acquired binocular vision images, target bounding boxes are labeled on the scaled sonar images to construct a sonar image detection training set containing location information and category labels; Step S13: Extract the precise pixel-level mask of the target using a region growing model. ,in To represent the target region, a one-to-one correspondence between sonar images and mask annotations is established, and a mask annotation set is constructed. During the training phase of the region growing model, the mask annotation set is input into the hierarchical attention block of the feature fusion network of the region growing model, and the cross-entropy loss between the predicted mask and the ground truth mask is calculated. Its mathematical expression is: , in, Here, N represents the mask loss value, and N is the total number of pixels in the current image. For the first The real label of each pixel The first prediction of the region growth model The probability value of a pixel belonging to the target object.

4. The method as described in claim 3, characterized in that, Step S2 includes: Step S21: Construct a pose transformation model based on Lie algebras. Define the relative pose between the end effector of the flexible connecting robotic arm and the target to be cut as the transformation matrix, the mathematical expression of which is: ， in, Let be the pose transformation matrix to be solved. It is a six-dimensional vector, where the first three dimensions represent translation offsets and the last three dimensions represent rotation vectors. This means converting a Lie algebra into a Lie group through an exponential mapping. for Rotational torque, for Translation vector; Step S22: Acquire the two-dimensional pixel coordinate observations of the target under binocular vision. Compared with polar coordinate observations in the sonar coordinate system Construct a joint least squares optimization function Its mathematical expression is: ， in, , These represent the three-dimensional spatial positions of the feature points on the surface of the target to be cut in the world coordinate system. For the first The pixel coordinates of each visual feature point on the imaging plane of the stereo camera. This is a visual projection function that projects 3D points into 2D pixel coordinates based on camera intrinsic parameters. For the first Observations of a sonar feature point within the sonar detection range. This is a sonar observation mapping function that maps 3D points to sonar polar coordinates, serving as a local 3D visual point cloud of the target. and These are acoustic and visual observation information matrices, used to characterize the confidence levels of different sensors; Step S23: Perform iterative optimization based on the Levenberg-Marquardt algorithm for the objective function. Performing a first-order Taylor expansion yields information about the pose increment. The linearized equations are derived, and a damping factor is introduced. The mathematical expression for iterative solution is: ， in, To comprehensively observe residuals, Let be the Jacobian matrix, and Ω be the inverse of the observation error covariance matrix. It is the identity matrix. The damping factor, The pose update increment is based on the initial pose and the pose update increment of the cutting target. Output the precise pose of the target to be cut .

5. The method as described in claim 4, characterized in that, Step S3 includes: Step S31, set the flexible connection robotic arm to be... A kinematic model based on standard DH parameters is established by cascading segments of a continuum, and the kinematics of the first segment are extracted. Link length, link torsion angle, joint offset, and real-time joint rotation angle of each joint. The first is calculated by the homogeneous transformation matrix. Local spatial pose transformation of each joint relative to the previous joint The global positive kinematics mapping matrix from the base of the flexible connecting robotic arm to the end effector The mathematical expression is: ， Step S32: Obtain the precise pose of the target to be cut. Combining underwater obstacle avoidance and joint physical constraints, an inverse kinematics objective optimization function is constructed. The optimal joint space configuration for reaching the work surface is calculated, and its mathematical expression is: ， in, This is a set of joint angle parameters for the entire arm. and These are the current 3D position vector and rotation matrix of the end effector, respectively. and These are the desired position and rotation matrix, respectively. , , These are the weighting coefficients; The singularity and joint deviation penalty term is used to constrain the smooth movement of the robotic arm in a confined underwater space to avoid interference between the joints; Step S33: Iteratively solve the objective optimization function using the damped least squares method. Output the optimal set of joint angle parameters The optimal joint angle parameter set The underlying servo motors of each rotary joint are mapped in real time to generate collaborative control commands for the drive motors. These collaborative control commands are then used to guide the end laser to align with the surface to be cut.

6. The method as described in claim 5, characterized in that, Step S4 includes: An external collision torque observation model based on flexible joints is constructed. Flexible connection components are installed at each rotary joint of the robotic arm. The external collision torque is then extracted. Real-time angle of the encoder on the motor side at each joint Real-time angle with the connecting rod side encoder , and They are respectively , The first derivative; based on the equivalent joint stiffness of the flexible connection component. With damping Calculate the real-time elastic internal torque of each joint. Its mathematical expression is: ， Construct an external environment disturbance torque observer based on generalized momentum to calculate the actual external collision torque experienced by each joint. When there is When the collision threshold exceeds the preset safety threshold, it is determined that the corresponding link has a rigid interference collision with an underwater unstructured obstacle.

7. The method as described in claim 6, characterized in that: After determining that a rigid interference collision has occurred, the current Jacobian matrix of the robotic arm is used. The external collision moment matrix Mapped to equivalent external contact force at the end This triggers a compliant obstacle avoidance law based on admittance control, establishing a force-displacement response equation, the mathematical expression of which is: ， in, , , These are the defined desired mass matrix, damping matrix, and stiffness matrix, respectively. The amount of obstacle avoidance posture retraction generated to adapt to external collision forces. , They are respectively The first and second derivatives; The robotic arm performs the obstacle avoidance pose retraction. When the joint undergoes flexible deflection due to water flow disturbance, the pseudo-inverse of the Jacobian matrix is ​​utilized. The mathematical expression for the inverse kinematics compensation solution is as follows: ， in, This is the angle compensation vector that needs to be added for the remaining joints that are not subject to collision constraints. To maintain the target's precise pose The required desired end-effector change will be sent to the underlying actuators of each joint in real time after compensation of the angle parameters.

8. The method as described in claim 7, characterized in that, Step S5 includes: Step S51: Extract the surface unit normal vector at each interpolation point s on the trajectory to be cut in the local three-dimensional visual point cloud of the target. Establish the desired spatial pose matrix of the end-effector laser cutting head. , Describe the three-dimensional special Euclidean group, its positional components The rotational component of the three-dimensional cutting curve that fits the plan Subject to pose alignment constraints, its mathematical expression is: ， in, This is the optical axis direction vector of the laser beam emitted from the end effector coordinate system; Step S52: Control the flexible connecting robotic arm to move the end effector to the cutting start point, prioritize the activation of the high-pressure jet device arranged around the laser emitter to displace the turbid water, activate the fiber laser, and dynamically compensate and control the output power based on the underwater laser transmission attenuation model. Its lasing power output function... The mathematical expression is: ， ， in, The constant cutting threshold energy required to melt through the target material. The real-time working focal length from the terminal exit port to the target surface. For wavelength The total attenuation coefficient of the laser in water; The inherent absorption coefficient of the water body. The dynamic scattering coefficient caused by the suspended slag and bubbles generated during cutting is determined by the function. Real-time adjustment of laser output to compensate for energy loss caused by focal length fluctuations and sudden changes in local water turbidity; Step S53: Use the end-effector binocular vision camera to capture in real time the normalized pixel coordinates of the center of the molten pool generated by the laser on the image plane. The projection point of the desired cutting trajectory on the image Visual feature error The mathematical expression for the image-based visual servoing feedforward-feedback control law is as follows: ， in, For space velocity spinor, For proportional control gain, The expected feedforward cutting speed for moving forward along the planned trajectory; The spatial velocity spinor is the Moore-Ponros generalized inverse of the image Jacobian matrix. Mapped to the control angular velocity of each joint Its mathematical expression is: ， The control angular velocity closed-loop drive is used to drive each joint of the flexible connected robotic arm to eliminate cutting trajectory offset in real time until the laser cutting operation is completed.

Images