An ultrasound-guided puncture surgical robot precision positioning system

Abstract

This invention relates to the field of surgical navigation technology, specifically disclosing an ultrasound-guided puncture surgery robot precision positioning system, including a collaborative control center. The collaborative control center is communicatively connected to the following modules: a deformation displacement sensing module, used for hybrid optical flow field registration and the fruit fly algorithm, to sense in real time the nonlinear deformation displacement field of soft tissue caused by patient respiration and probe pressure, and to dynamically lock the optimal elastic matching point through intelligent iterative search, correcting the target point drift coordinates online. This invention, by fusing optical flow field registration and the fruit fly algorithm, can effectively sense and quantify the nonlinear deformation displacement field of soft tissue caused by the patient's respiratory rhythm and probe contact pressure, overcoming the limitation of traditional fixed templates that cannot track dynamic targets. It can jointly compensate for low-frequency drift caused by respiration and high-frequency abrupt changes caused by pressure, ensuring high accuracy and robustness of target positioning under dynamic physiological conditions.

Description

Technical Field

[0001] This invention relates to the field of surgical navigation technology, and in particular to a precise positioning system for an ultrasound-guided puncture surgery robot. Background Technology

[0002] Ultrasound-guided puncture surgery positioning technology stems from the clinical need for high-precision, real-time image navigation in minimally invasive interventional diagnosis and treatment. Traditional manual puncture relies on the operator's experience and spatial judgment ability, which carries the risk of target deviation and path deviation. Ultrasound imaging, due to its advantages such as no radiation, real-time dynamics, and portability, has become one of the most commonly used guidance methods for puncture surgery robots.

[0003] For example, Chinese Patent Publication No. CN116370089B describes a detection method and system for detecting the positioning accuracy of a puncture surgery robot. Through specific detection methods, the positioning accuracy of the puncture surgery robot during the positioning process using a 3D vision camera can be detected, and the measurement accuracy of the positioning marks can also be detected, ensuring that the robotic arm in the puncture surgery robot can accurately reach a certain fixed point.

[0004] In existing technologies, the patient's breathing movements and probe pressure during the puncture positioning process cause nonlinear deformation of soft tissue, resulting in real-time drift of the target point in the ultrasound image. The fixed template cannot track the real-time target point coordinates, and the puncture needle tip is affected by the anisotropic resistance of the tissue, making the actual advancement path prone to deviate from the planned straight line, causing the accumulated error to exceed the allowable range. In addition, the elastic target point drift and the needle tip deflection path are further coupled, which leads to the existence of end-point comprehensive error even after correcting the individual variables, forming a closed-loop residual of target-needle tip joint pose deviation. To address these issues, an ultrasound-guided puncture surgery robot precision positioning system is proposed. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention provides an ultrasound-guided puncture surgery robot precision positioning system, which can effectively solve the problems involved in the prior art.

[0006] The objective of this invention can be achieved through the following technical solution: This invention provides a precise positioning system for an ultrasound-guided puncture surgery robot, including a collaborative control center, which is communicatively connected to the following modules: The deformation displacement sensing module is used for hybrid optical flow field registration and fruit fly algorithm to sense the nonlinear deformation displacement field of soft tissue caused by patient breathing and probe pressure in real time. It also dynamically locks the optimal elastic matching point through intelligent iterative search and corrects the target point drift coordinates online, effectively compensating for the real-time target point drift caused by soft tissue deformation. The needle path adaptive control module integrates depth Q network and Kalman filter to establish a nonlinear mapping relationship between tissue resistance and needle tip deflection. It also integrates real-time force feedback and predicted yaw angle to dynamically adjust the puncture direction, suppress the deviation of the actual advancement path from the planned straight line, block the growth of path accumulation error, and realize real-time dynamic correction and error blocking of the puncture path. The clustering coupling mode identification module is used to perform cluster analysis on real-time and historical target-tip joint pose deviation data, identify the coupling mode types of errors in space and time, and classify and label the modes to achieve automatic identification and classification labeling of error coupling modes. The particle coupling parameter search module is used to search for the optimal decoupling parameter combination based on the identified coupling mode type using the particle swarm algorithm. It decomposes the coupling error into target-related components and needle-tip-related components that can be handled independently, thereby achieving effective separation and independent processing of the coupling error. The end-effector residual collaborative compensation module, based on the output coupling mode label, quickly matches the joint pose deviation type to which the current residual belongs, and performs collaborative iterative search on the matched end-effector composite residual to generate joint pose fine-tuning instructions for the end effector, thereby eliminating the target-tip composite deviation in a closed loop and achieving closed-loop elimination of end-effector composite deviation and high-precision positioning throughout the entire process.

[0007] Preferably, the deformation displacement sensing module includes an optical flow deformation extraction unit and a fruit fly target tracking unit; The optical flow deformation extraction unit, based on optical flow field registration technology, calculates the nonlinear displacement vector of soft tissue at the pixel level between consecutive ultrasound image frames, constructs a deformation displacement field that reflects the influence of breathing and pressure, quantifies the target drift trajectory, and achieves sub-pixel level accurate construction of the soft tissue deformation displacement field. The fruit fly target tracking unit is used to perform odor-guided iterative search in the constructed deformation displacement field using the fruit fly algorithm, quickly converge to the optimal elastic matching point at the current moment, correct the real-time coordinates of the target point, and output the dynamic target point position, thereby realizing rapid dynamic tracking and online correction of the target point position.

[0008] Preferably, the optical flow deformation extraction unit specifically includes: By using a puncture surgical robot to acquire a continuous sequence of ultrasound images of the patient's puncture site, and based on the pyramid hierarchical optical flow field registration algorithm, the sub-pixel displacement vector of each pixel in the soft tissue between adjacent frames is calculated to construct an initial nonlinear deformation displacement field. The position of each pixel in this deformation displacement field is the candidate matching point, which realizes high-resolution dynamic capture of soft tissue deformation and lays the foundation for target tracking. The displacement field is decomposed in Gaussian scale along the time axis to extract the low-frequency periodic displacement component caused by respiratory rhythm and the high-frequency local abrupt change component caused by probe pressure. The target drift trajectory is quantified, and the spatial coordinate mapping relationship of each candidate matching point before and after decomposition is preserved to achieve physical separation of respiratory and pressure interference and ensure complete traceability of deformation characteristics. The decomposed multi-component displacement field is projected onto the base coordinate system of the puncture surgical robot to generate the deformation displacement mapping matrix at the current moment. The positions of each grid node in the matrix are used as the output set of candidate matching points to achieve spatial unification and structured output of candidate matching points, which facilitates rapid tracking in the future.

[0009] Preferably, the fruit fly target tracking unit specifically includes: Each candidate matching point in the deformation displacement field is encoded as the odor concentration sampling position of the fruit fly individual. The odor concentration fitness function is constructed using the inverse square of the predicted residual of the target drift trajectory, thereby amplifying the sensitivity of the target drift residual and improving the tracking accuracy. By using the odor-guided random search mechanism of the fruit fly population, dense iterative tracking is performed in the local extreme region of the deformation displacement field, which quickly converges to the optimal elastic matching point with the highest fitness at the current moment, avoiding global traversal search and significantly shortening the optimal matching point location time. Based on the spatial coordinate offset of the optimal matching point, the real-time coordinates of the original planned target point are corrected online, and the dynamic target point position after respiratory and pressure compensation is output to the collaborative control center, realizing real-time dynamic tracking and compensation of the target point under soft tissue deformation.

[0010] Preferably, the needle path adaptive control module includes a drag deflection mapping unit and a yaw dynamic filtering unit; The resistance deflection mapping unit is used to learn the mapping relationship between tissue anisotropic resistance and needle tip deflection angle using a deep Q-network, predict the deflection trend caused by resistance, and achieve accurate prediction of the needle tip deflection trend caused by tissue resistance. The yaw dynamic filtering unit is used to fuse real-time force sensor feedback and predicted yaw angle using Kalman filtering, and combined with the predicted yaw trend to dynamically correct the puncture direction, suppress path accumulation error, and achieve high-precision estimation of yaw angle and real-time direction correction.

[0011] Preferably, the drag deflection mapping unit specifically includes: The state space of the deep Q-network is constructed as the real-time needle tip position, needle insertion speed and tissue resistance vector fed back by multi-axis force sensors, and the action space is the predicted compensation amount of the needle tip deflection angle, which comprehensively characterizes the dynamic physical characteristics of the puncture process. By using an online interactive sampling method, the deep Q network learns the nonlinear mapping relationship between anisotropic drag and needle tip deflection angle, and outputs the deflection trend prediction value under different drag conditions, thus achieving accurate modeling and prediction of complex drag characteristics. The predicted deflection trend is injected into the yaw dynamic filter unit as a feedforward control quantity to achieve pre-compensation for the needle tip path deviation caused by drag, and actively suppress the path deviation trend before deflection occurs.

[0012] Preferably, the yaw dynamic filtering unit specifically includes: A discrete state-space model of the puncture needle tip is established. The observation equation of Kalman filter is constructed by the yaw angle observation value fed back by the real-time force sensor and the yaw trend prediction value output by the drag deflection mapping unit. This realizes the fusion observation of prediction and measurement, and improves the robustness of state estimation. By employing a two-step recursive algorithm of time update and measurement update using Kalman filtering, the observed and predicted values ​​are fused to dynamically estimate the optimal yaw angle and its covariance uncertainty at the current moment, thereby reducing single noise interference and obtaining a high-confidence optimal yaw angle estimate. Based on the optimal yaw angle estimation results, the direction fine-tuning command of the puncture robot end effector is generated in real time, the puncture propulsion path is dynamically corrected, the growth of path cumulative error is suppressed, and the real-time closed-loop correction of the needle tip path is realized, thus blocking the stepwise accumulation of error.

[0013] Preferably, the clustering coupler identification module specifically includes: The target drift residual sequence and the needle tip deflection residual sequence are collected in real time and historical time windows. The two are spliced ​​together to form a joint pose deviation feature vector. Normalization and dimensionality reduction preprocessing are then performed to effectively eliminate dimensional differences and improve the computational efficiency and accuracy of cluster analysis. A density-based clustering algorithm is used to perform unsupervised classification of joint pose deviation feature vectors. Based on the deviation amplitude, directional angle, and time delay correlation, different types of coupling pattern clusters are identified, including four types: breathing-dominant, pressure drift, drag deflection, and hybrid coupling. This enables automatic identification of error coupling types without the need to preset the number of classifications. A unique mode label is generated for each coupled mode cluster, and the joint deviation at the current moment is matched to the nearest cluster. The coupled mode label is then output to the collaborative control center to achieve rapid matching of deviation types, providing an accurate basis for subsequent decoupling compensation.

[0014] Preferably, the particle coupling parameter search module specifically includes: Based on the coupling mode label, the corresponding decoupling parameter search space is retrieved, and the decoupling parameter vector of the particle swarm algorithm is defined, including the target-needle tip weight allocation coefficient and the decoupling time window length, so as to achieve accurate matching between the mode and the decoupling strategy and improve the optimization targeting. With the norm of the final integrated residual as the fitness objective, the optimal combination of parameters that minimizes the coupling error is searched in the decoupling parameter space through the velocity-position iterative update formula of the particle swarm, ensuring efficient convergence of the search process and obtaining the globally optimal decoupling parameters. Based on the optimal decoupling parameter combination, the real-time joint pose deviation is decomposed into target-related components and needle-tip-related components that can be processed independently, and output to the end residual collaborative compensation module respectively, so as to achieve complete separation of coupling error and provide accurate input for independent compensation.

[0015] Preferably, the end residual collaborative compensation module specifically includes: Receive the coupling mode label and decoupling component of the output, quickly match the joint pose deviation type to which the current end-point integrated residual belongs, and ensure that the compensation measures correspond accurately to the current error characteristics; Under the constraint of the deviation type after matching, a collaborative iterative search strategy is adopted to simultaneously optimize the position fine-tuning and attitude fine-tuning of the end effector, generating a joint pose fine-tuning candidate instruction set to avoid new secondary coupling deviations caused by individual corrections. The fine-tuning instruction that minimizes the target-needle tip combined deviation norm is selected from the candidate instruction set and sent to the puncture robot controller. The closed loop eliminates residual coupling deviation, achieving high-precision positioning throughout the puncture process and stable, low-error tracking of the target point throughout the puncture process.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This ultrasound-guided puncture surgery robot precision positioning system, by integrating optical flow field registration and fruit fly algorithm, can effectively sense and quantify the nonlinear deformation displacement field of soft tissue caused by the patient's respiratory rhythm and probe contact pressure. It breaks through the limitation of traditional fixed templates that cannot track dynamic target points. It can jointly compensate for low-frequency drift caused by breathing and high-frequency mutation caused by pressure, ensuring high accuracy and strong robustness of target positioning under dynamic physiological environment.

[0017] 2. This ultrasound-guided puncture surgery robot precision positioning system establishes a nonlinear mapping model between tissue anisotropic resistance and needle tip deflection, and combines adaptive filtering technology with real-time force feedback and deflection prediction information. It can provide feedforward compensation before the actual deflection of the needle tip, dynamically correct the puncture direction, significantly reduce path deviation caused by tissue resistance differences, and effectively block the growth chain of error accumulation with increasing puncture depth.

[0018] 3. This ultrasound-guided puncture surgery robot precision positioning system addresses the end-effector deviation caused by the coupling between elastic target drift and needle tip deflection path. It employs cluster analysis technology to automatically identify the coupling mode of error in the spatiotemporal domain and combines it with adaptive search for optimal decoupling parameters. This allows the highly coupled joint pose deviation to be decomposed into mutually independent target-related components and needle tip-related components, laying the foundation for independent error compensation. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the workflow of a precision positioning system for an ultrasound-guided puncture surgery robot according to the present invention. Figure 2 This is a schematic diagram of the module structure of a precision positioning system for an ultrasound-guided puncture surgery robot according to the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0021] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: a precision positioning system for an ultrasound-guided puncture surgery robot, comprising a collaborative control center, which is connected to the following modules for communication: The deformation displacement sensing module is used for hybrid optical flow field registration and fruit fly algorithm to sense the nonlinear deformation displacement field of soft tissue caused by patient breathing and probe pressure in real time. It dynamically locks the optimal elastic matching point through intelligent iterative search and corrects the target point drift coordinates online, effectively compensating for the real-time target point drift caused by soft tissue deformation. The deformation displacement sensing module includes an optical flow deformation extraction unit and a fruit fly target tracking unit. The optical flow deformation extraction unit, based on optical flow field registration technology, calculates pixel-level nonlinear displacement vectors of soft tissue from continuous ultrasound image frames, constructing a deformation displacement field reflecting the effects of respiration and pressure. It quantifies the target point drift trajectory, achieving sub-pixel-level accurate construction of the soft tissue deformation displacement field. Using a puncture surgical robot, it acquires a sequence of continuous ultrasound image frames of the patient's puncture site. Based on a pyramid-layered optical flow field registration algorithm, it calculates the sub-pixel-level displacement vectors of each pixel in the soft tissue between adjacent frames, constructing an initial nonlinear deformation displacement field. The positions of each pixel in this deformation displacement field are the candidate matching points, achieving high-resolution dynamic capture of soft tissue deformation, providing a target... Point tracking lays the foundation by performing Gaussian-scale spatial decomposition of the displacement field along the time axis, extracting the low-frequency periodic displacement components caused by respiratory rhythm and the high-frequency local abrupt change components caused by probe pressure, quantifying the target drift trajectory, and preserving the spatial coordinate mapping relationship of each candidate matching point before and after decomposition, realizing the physical separation of respiratory and pressure interference, ensuring the complete traceability of deformation features, and projecting the decomposed multi-component displacement field onto the base coordinate system of the puncture surgical robot to generate the deformation displacement mapping matrix at the current moment. The positions of each grid node in the matrix are used as the output set of candidate matching points, realizing the spatial unification and structured output of candidate matching points, which facilitates rapid tracking in the future. It should be noted that before the puncture procedure, the end effector of the puncture robot holds the ultrasound probe and continuously acquires a sequence of ultrasound images of the patient's puncture site at a fixed frame rate of 25 frames / second. The acquisition time span covers at least two complete respiratory cycles, totaling no less than 50 consecutive images. Adjacent frames are input into a three-layer pyramid optical flow registration algorithm. The bottom layer of the pyramid retains the original image resolution of 512×512 pixels, while the top layer is downsampled to 64×64 pixels. Sub-pixel displacement vectors of each pixel in the soft tissue between adjacent frames are calculated layer by layer from the top to the bottom, with a pixel displacement calculation accuracy of [insert accuracy here]. Using 0.1 pixels as the basis, an initial nonlinear deformation displacement field covering the entire ultrasound image field of view is constructed based on the displacement vectors of all pixels. The coordinates of each pixel in this displacement field serve as candidate matching points for soft tissue deformation. Gaussian scale space decomposition is performed on the constructed initial nonlinear deformation displacement field along the time axis, with the Gaussian kernel scale parameter set to four octaves, each containing three scale intervals. Through scale space decomposition, low-frequency periodic displacement components with frequencies from 0.2 Hz to 0.8 Hz are extracted from the displacement field. These components correspond to the overall soft tissue deformation caused by the patient's respiratory rhythm. Simultaneously, [further details are needed]. High-frequency local abrupt change components with frequencies above 2 Hz are selected, corresponding to local tissue compression deformation caused by changes in probe contact pressure. The two types of displacement components are differentially analyzed with the original displacement field to quantize the real-time drift trajectory curve of the target point in the image coordinate system. Simultaneously, the spatial coordinate mapping relationship of each candidate matching point before and after decomposition is recorded, establishing a mapping lookup table from the original image coordinates to the coordinates of each component displacement field. The low-frequency respiratory component displacement field and the high-frequency pressure component displacement field obtained after Gaussian scale spatial decomposition are then used for coordinate analysis using the hand-eye calibration matrix between the ultrasound probe and the robot end effector. The transformation is uniformly projected onto the base coordinate system of the puncture surgical robot. The hand-eye calibration matrix adopts a 4×4 homogeneous transformation matrix form, in which the calibration accuracy of the rotation matrix is ​​0.5 degrees and the calibration accuracy of the translation vector is 0.2 millimeters. After the projection calculation is completed, the deformation displacement mapping matrix at the current moment is generated. This matrix has a dimension of 64 rows × 64 columns. Each grid node position in the matrix corresponds to the three-dimensional spatial coordinates of a candidate matching point in the robot's base coordinate system. This mapping matrix is ​​used as the final output result of the optical flow deformation extraction unit and passed to the fruit fly target tracking unit for subsequent optimal matching point iterative search. The Drosophila target tracking unit utilizes the Drosophila algorithm to perform odor-guided iterative search within a constructed deformation displacement field, rapidly converging to the optimal elastic matching point at the current moment. It then corrects the real-time target coordinates and outputs the dynamic target position, enabling rapid dynamic tracking and online correction of the target position. Each candidate matching point in the deformation displacement field is encoded as the odor concentration sampling position of an individual Drosophila. An odor concentration fitness function is constructed using the reciprocal of the squared residual of the target drift trajectory prediction, amplifying the sensitivity of the target drift residual and improving tracking accuracy. Through the odor-guided random search mechanism of the Drosophila population, dense iterative tracking is performed in the local extreme regions of the deformation displacement field, rapidly converging to the optimal elastic matching point with the highest fitness at the current moment. This avoids global traversal search and significantly shortens the optimal matching point location time. Based on the spatial coordinate offset of the optimal matching point, the real-time coordinates of the originally planned target are corrected online, and the dynamic target position after respiration and pressure compensation is output to the collaborative control center, achieving real-time dynamic tracking and compensation of the target under soft tissue deformation. It should be noted that in the 64×64 grid nodes of the deformation displacement mapping matrix, each node position is defined as the odor concentration sampling position of a fruit fly individual. Each sampling position corresponds to a three-dimensional spatial coordinate point in the robot's base coordinate system. Using the target point drift trajectory curve output by the optical flow deformation extraction unit as a reference, an odor concentration fitness function is constructed. The fitness function value is the reciprocal of the square of the target point drift trajectory prediction residual. The prediction residual is calculated by the Euclidean distance between the current candidate matching point position and the time-series prediction point on the trajectory curve. The smaller the residual, the larger the fitness function value, indicating that the candidate matching point is closer to the true elastic deformation target point position. Each fruit fly individual independently calculates the fitness value of its position, forming the initial odor concentration distribution field. The fruit fly population contains 25 individuals, and the population center is initially set within 5 mm around the optimal elastic matching point at the previous moment. Under the guidance of odor, each individual generates a position perturbation with a step size ranging from 0.3 mm to 1.2 mm in a random direction. The perturbation step size decreases linearly with the number of iterations, and the decay coefficient is set to 0.85. In each iteration... In the process, all individuals recalculate the fitness value of their current position. The group updates the global optimal position by comparing the fitness values. When the rate of change of the global optimal fitness value is less than 2% in three consecutive iterations, it is determined that it has entered a local extremum region. At this time, the search step size is automatically reduced to 0.1 mm, and dense iterative tracking is performed for a total of 40 iterations. Finally, it converges to the grid node with the highest fitness, which is the optimal elastic matching point at the current moment. Based on the spatial coordinate offset of the optimal elastic matching point, which is the vector difference between the coordinates of the current matching point and the coordinates of the original planned target point, the component offset values ​​of the X, Y, and Z axes are calculated in the robot base coordinate system. The offset calculation accuracy is 0.05 mm. The above offset is superimposed on the real-time coordinates of the original planned target point to complete the joint compensation for low-frequency drift caused by respiratory rhythm and high-frequency mutation caused by probe pressure. The corrected dynamic target point position is output to the collaborative control center in the form of a three-dimensional coordinate vector. The coordinate data update frequency is consistent with the ultrasound image acquisition frame rate, that is, it is updated once every 40 milliseconds to ensure the real-time and accuracy of target point positioning during puncture. The needle path adaptive control module is used to integrate the depth Q network and Kalman filter to establish a nonlinear mapping relationship between tissue resistance and needle tip deflection. It also integrates real-time force feedback and predicted yaw angle to dynamically adjust the puncture direction, suppress the deviation of the actual advancement path from the planned straight line, block the growth of path cumulative error, and realize real-time dynamic correction and error blocking of the puncture path. The needle path adaptive control module includes a resistance deflection mapping unit and a yaw dynamic filtering unit. The resistance deflection mapping unit utilizes a deep Q-network to learn the mapping relationship between tissue anisotropic resistance and needle tip deflection angle, predicts the deflection trend caused by resistance, and achieves accurate prediction of the needle tip deflection trend caused by tissue resistance. The state space of the deep Q-network is constructed as the real-time needle tip position, needle insertion speed, and tissue resistance vector fed back by multi-axis force sensors, and the action space is the predicted compensation amount of the needle tip deflection angle, comprehensively characterizing the dynamic physical characteristics of the puncture process. Through online interactive sampling, the deep Q-network learns the nonlinear mapping relationship between tissue anisotropic resistance and needle tip deflection angle, outputs the predicted value of deflection trend under different resistance conditions, and achieves accurate modeling and prediction of complex resistance characteristics. The predicted deflection trend is injected into the yaw dynamic filtering unit as a feedforward control quantity to achieve pre-compensation for the needle tip path deviation caused by resistance and actively suppress the path deviation trend before deflection occurs. It should be noted that before the puncture surgery robot performs the puncture operation, the state space configuration of the deep Q-network is first completed. The state space consists of three dimensions: real-time needle tip position, needle insertion speed, and tissue resistance vector fed back by the multi-axis force sensor. The real-time needle tip position is acquired by an electromagnetic positioning sensor with an accuracy of 0.10 mm in the robot's base coordinate system. The needle insertion speed is read in real-time by the robot's end effector encoder and set to a constant feed rate of 2.5 mm / s. The multi-axis force sensor is installed between the puncture needle clamping mechanism and the robot's end effector, with a sampling frequency of 100 Hz. The feedback resistance vector includes the axial force of needle insertion, lateral shear force, and torque. Each component in the state space has a range of 0 to 10 Newtons and a resolution of 0.01 Newtons. Each component is normalized before being input into the deep Q-network. The normalization range is set to -1 to +1 to ensure network training stability and convergence speed. The deep Q-network uses a three-layer fully connected neural network structure. The input layer has 12 nodes in the state space dimension, the hidden layers have 64 and 32 nodes respectively, and the output layer has 7 nodes in the action space dimension, corresponding to seven discrete actions within the range of -3 degrees to +3 degrees for the needle tip deflection angle prediction compensation. During online interactive sampling, the network selects actions with an initial exploration rate of 0.95, advancing each puncture. A state update and network learning are triggered when the distance reaches 0.50 mm. The experience replay pool capacity is set to 10,000 samples, and 128 samples are randomly selected each time for mini-batch gradient descent training. The target network is synchronized every 100 steps. After online learning of 50 consecutive puncture trajectories, the deep Q network can accurately establish the nonlinear mapping relationship between the tissue resistance vector and the needle tip deflection angle, and output the deflection trend prediction value under different resistance conditions. The prediction error converges to within 0.15 degrees. The deflection trend prediction value output by the deep Q network is injected as a feedforward control quantity into the yaw dynamic filter unit to achieve pre-compensation for the needle tip path deviation caused by resistance. The feedforward control quantity is sent to the yaw dynamic filter unit in the form of deflection angular velocity, with an update frequency of 40 Hz, which is synchronized with the frame rate of ultrasound image acquisition. When the multi-axis force sensor detects a lateral shear force exceeding 0.30 Newtons, the depth Q network immediately outputs a deflection angle prediction compensation of -2.10 degrees. The yaw dynamic filter unit pre-adds this feedforward quantity during the time update stage of the Kalman filter, so that the needle tip receives the reverse correction command before it actually deflects, reducing the needle tip path deviation caused by tissue anisotropic resistance from 1.20 mm without compensation to less than 0.25 mm, effectively suppressing the source of subsequent path accumulation error growth. The yaw dynamic filtering unit utilizes Kalman filtering to fuse real-time force sensor feedback and predicted yaw angles. Combined with the predicted yaw trend, it dynamically corrects the puncture direction, suppresses path accumulation error, and achieves high-precision yaw angle estimation and real-time direction correction. It establishes a discrete state-space model of the puncture needle tip, constructs the observation equation of Kalman filtering with the observed yaw angle value fed back by the real-time force sensor and the predicted yaw trend value output by the resistance deflection mapping unit, realizes the fusion of prediction and measurement observation, and improves the robustness of state estimation. Through a two-step recursive algorithm of time update and measurement update of Kalman filtering, it fuses the observed value and the predicted value, dynamically estimates the optimal yaw angle and its covariance uncertainty at the current moment, reduces single noise interference, and obtains a high-confidence optimal yaw angle estimate. Based on the optimal yaw angle estimation result, it generates direction fine-tuning commands for the puncture robot end effector in real time, dynamically corrects the puncture propulsion path, suppresses the growth of path accumulation error, realizes real-time closed-loop correction of the needle tip path, and blocks the stepwise accumulation of error. It should be noted that when establishing the discrete state-space model of the puncture needle tip, the yaw angle of the needle tip in the robot's base coordinate system is used as the state variable. The state transition matrix is ​​set as the identity matrix, and the process noise covariance matrix is ​​set to 0.05 square degrees. The yaw angle observation value fed back in real time by a six-dimensional force sensor installed at the end of the puncture needle clamping mechanism is used as the measurement input. The sampling frequency of this sensor is 100 Hz, and the yaw angle observation noise covariance is set to 0.10 square degrees. At the same time, the yaw trend prediction value output by the drag deflection mapping unit is used as the feedforward control quantity and introduced into the observation equation to construct an extended observation matrix. The weight coefficient of the prediction value is set to 0.85, and the weight coefficient of the observation value is set to 0.15, forming a joint observation equation that integrates prediction and measurement. The optimal yaw angle and its covariance uncertainty at the current moment are dynamically estimated through a two-step recursive algorithm. In the time update stage, the prior estimate at the current moment is calculated based on the optimal yaw angle estimate at the previous moment, and the process noise covariance is superimposed on the prior covariance matrix. In the measurement update stage, the Kalman gain moment is calculated. The value of the posterior covariance matrix depends on the relative magnitude of the prior covariance and the observation noise covariance. Substituting the joint observations into the update equation, the posterior yaw angle estimate and the posterior covariance matrix at the current moment are obtained. The posterior covariance value is controlled within 0.02 square degrees, indicating that the estimation result has a high confidence level and effectively suppresses the interference of single sensor noise or prediction model error on the yaw angle estimation. The posterior yaw angle estimate is converted into the rotation angle increment of the end effector. The angular displacement command of each joint is obtained by solving the robot inverse kinematics. The servo driver outputs the command to each joint motor at a control cycle of 1000 Hz. The direction fine-tuning command is updated every 25 milliseconds, keeping it synchronized with the frame rate of the ultrasound image acquisition. This allows the needle tip to complete a direction correction every 0.50 mm of advancement. After closed-loop control, the deviation of the needle tip path caused by tissue anisotropic resistance is reduced from 1.20 mm without compensation to within 0.25 mm. At the same time, the yaw angle estimation error converges to within 0.15 degrees, effectively blocking the growth chain of path accumulation error. The clustering coupling mode identification module is used to perform cluster analysis on real-time and historical target-tip joint pose deviation data, identify the coupling mode types of errors in space and time, and classify and label the modes to achieve automatic identification and classification labeling of error coupling modes. The particle coupling parameter search module is used to search for the optimal decoupling parameter combination based on the identified coupling mode type using the particle swarm algorithm. It decomposes the coupling error into target-related components and needle-tip-related components that can be handled independently, thereby achieving effective separation and independent processing of the coupling error. The end-effector residual collaborative compensation module, based on the output coupling mode label, quickly matches the joint pose deviation type to which the current residual belongs, and performs collaborative iterative search on the matched end-effector composite residual to generate joint pose fine-tuning instructions for the end effector, thereby eliminating the target-tip composite deviation in a closed loop and achieving closed-loop elimination of end-effector composite deviation and high-precision positioning throughout the entire process.

[0022] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: the clustering coupling mode identification module specifically includes: collecting the target drift residual sequence and the needle tip deflection residual sequence within the real-time and historical time windows, splicing the two to form a joint pose deviation feature vector, and performing normalization and dimensionality reduction preprocessing to effectively eliminate dimensional differences and improve the computational efficiency and accuracy of clustering analysis. A density-based clustering algorithm is used to perform unsupervised classification of the joint pose deviation feature vector. Based on the deviation amplitude, direction angle, and time delay correlation, different types of coupling mode clusters are identified, including four types: breathing-dominant, pressure drift, resistance deflection, and hybrid coupling. This achieves automatic identification of error coupling types without the need to preset the number of classifications. A unique mode label is generated for each coupling mode cluster, and the joint deviation at the current moment is matched to the nearest neighbor cluster. The coupling mode label is output to the collaborative control center to achieve rapid matching of deviation types and provide accurate basis for subsequent decoupling compensation. It should be noted that before the actual puncture operation is performed by the surgical robot, the target drift residual sequence and needle tip deflection residual sequence are synchronously collected within a real-time and historical time window at a period of 40 milliseconds. The target drift residual comes from the Euclidean distance difference between the dynamic target position output by the deformation displacement sensing module and the original planned target position. The needle tip deflection residual comes from the angle difference between the posterior yaw angle output by the yaw dynamic filtering unit and the initial planned path direction. The time window length is set to 2 seconds, containing a total of 50 sets of continuous sampling points. The two sets of residual sequences are spliced ​​together end to end in chronological order to form a 100-dimensional joint pose deviation feature. The eigenvectors were then normalized to eliminate the influence of dimensional differences on clustering analysis. Principal component analysis was used for dimensionality reduction preprocessing, compressing the original 100-dimensional eigenvectors to 15-dimensional principal components, retaining more than 95% of the variance contribution rate in the original information. The density-based clustering algorithm DBSCAN was used for unsupervised classification of the dimensionality-reduced joint pose deviation eigenvectors. The neighborhood radius parameter epsilon was set to 0.35, and the minimum number of neighborhood samples minPts was set to 4. During the clustering process, different types of coupling modes were identified based on three discriminant dimensions: deviation amplitude, orientation angle, and time delay correlation. The pattern clusters are defined as follows: deviation amplitude refers to the Euclidean norm of the target drift residual and the needle tip deflection residual; direction angle refers to the cosine similarity of the two types of residual vectors in high-dimensional space; and time delay correlation refers to the lag time of the needle tip deflection response after the target drift occurs, with a lag time calculation accuracy of 40 milliseconds. Through joint discrimination of these three dimensions, historical deviation data are divided into four typical pattern clusters: breathing-dominated, pressure drift, resistance deflection, and hybrid coupling. The center point coordinates and boundary thresholds of each cluster are recorded in the pattern knowledge base. For the joint pose deviation feature vector acquired at the current moment, the same normalization and dimensionality reduction preprocessing are performed, followed by calculation. The Mahalanobis distance between the sample and the cluster centers of each pattern in the pattern knowledge base is calculated. During the Mahalanobis distance calculation, the covariance matrix of each pattern cluster is used for scale correction to eliminate the influence of the correlation between feature dimensions on the distance metric. At the current time, the deviation vector is matched to the pattern cluster with the smallest Mahalanobis distance and the corresponding coupled pattern label is output to the collaborative control center. The confidence threshold for pattern matching is set to 0.75. When the maximum confidence is lower than this threshold, the sample is marked as an unrecognized pattern, triggering the online incremental update of the clustering model. The new deviation sample is included in the training set and the cluster center is recalculated to achieve adaptive evolution of coupled pattern recognition capability. The particle coupling parameter search module specifically includes: retrieving the corresponding decoupling parameter search space based on the coupling mode label, defining the decoupling parameter vector of the particle swarm algorithm, including the target-tip weight allocation coefficient and the decoupling time window length, to achieve accurate matching between the mode and the decoupling strategy and improve the optimization targeting. Taking the minimization of the norm of the end-comprehensive residual as the fitness objective, the optimal parameter combination that minimizes the coupling error is searched in the decoupling parameter space through the velocity-position iterative update formula of the particle swarm, ensuring efficient convergence of the search process and obtaining the globally optimal decoupling parameters. Based on the optimal decoupling parameter combination, the real-time joint pose deviation is decomposed into independently processable target-related components and tip-related components, which are output to the end-residual collaborative compensation module to achieve complete separation of coupling error and provide accurate input for independent compensation. It should be noted that, based on the coupling mode label type output by the clustering coupling mode identification module, the decoupling parameter search space corresponding to that mode is automatically retrieved. For the breathing-dominant mode, the decoupling parameter search space is set to a target-tip weight allocation coefficient between 0.70 and 0.90, and the decoupling time window length is set to 1.60 seconds; for the pressure drift mode, the weight allocation coefficient is set to a range of 0.40 to 0.60, and the time window length is set to 0.80 seconds; for the resistance deflection mode, the weight allocation coefficient is set to a range of 0.20 to 0.40, and the time window length is set to 0.40 seconds; for the mixed coupling mode, the weight allocation coefficient is set to a range of 0.50 to 0.70, and the time window length is set to 1.20 seconds. The decoupling parameter vector of the particle swarm optimization algorithm is defined in two dimensions, including two components: the target-tip weight allocation coefficient and the decoupling time window length. The particle swarm size is set to 30 particles, the maximum number of iterations is set to 50, the inertia weight is set to 0.70, and the learning factors c1 and c2 are both set to 1.50. The velocity upper limit constraint is 20% of the parameter range, and the position boundary constraint is limited to the upper and lower limits of the search space corresponding to each mode. The fitness objective of particle swarm optimization is to minimize the norm of the end-effector composite residual. The end-effector composite residual is defined as the square root of the weighted sum of squares of the target drift residual and the tip deflection residual, where the weighting coefficient is the weight allocation coefficient to be optimized. In each iteration, each particle updates its position according to its current velocity and calculates its current position. The corresponding fitness values ​​are set, and the individual historical best position and the group global best position are updated respectively. The velocity update formula adopts the linear combination form of the standard particle swarm optimization algorithm, which includes three parts: inertial component, individual cognitive component, and social cognitive component. When the rate of change of the global best fitness value is less than 0.5% in five consecutive iterations, the optimization process is considered to have converged, and the iteration is terminated early. After optimization, the optimal decoupling parameter combination that minimizes the final integrated residual is output, including the optimal weight allocation coefficient and the optimal decoupling time window length, so as to effectively separate the coupled joint pose deviation into two independent components. Based on the optimal decoupling parameter combination output by the particle coupling parameter search module, the real-time joint pose deviation is decoupled and decomposed. Specifically, the optimal decoupling parameter combination is used to decouple and decompose the real-time joint pose deviation. The weighting coefficient is a scaling factor. The joint pose deviation vector is projected onto the target-related subspace and the tip-related subspace, and the target-related component and the tip-related component are extracted respectively. The decoupling time window length is used to control the participation of historical deviation data within the sliding window. Historical data outside the time window is assigned zero weight in the decoupling calculation. After decomposition, the target-related component and the tip-related component are output to the end residual collaborative compensation module in the form of independent data streams. After receiving the two components, the end residual collaborative compensation module generates position compensation command and attitude compensation command respectively for the target drift error and the tip deflection error, so as to realize the separate independent compensation of coupling error and avoid the secondary deviation of the other variable caused by the separate correction of one variable. The end-effector residual collaborative compensation module specifically includes: receiving the output coupling mode label and decoupling component, quickly matching the joint pose deviation type to which the current end-effector comprehensive residual belongs, ensuring that the compensation measures correspond precisely to the current error characteristics, and under the constraint of the matched deviation type, adopting a collaborative iterative search strategy to simultaneously optimize the position fine-tuning amount and attitude fine-tuning amount of the end effector, generating a joint pose fine-tuning candidate instruction set, avoiding new secondary coupling deviations caused by individual corrections, selecting the fine-tuning instruction that minimizes the target-needle tip comprehensive deviation norm from the candidate instruction set, and sending it to the puncture robot controller to eliminate residual coupling deviations in a closed loop, achieving high-precision positioning throughout the puncture process, and achieving stable, low-error tracking of the target point throughout the puncture process; It should be noted that before the actual puncture operation is performed by the puncture surgical robot, it receives a coupling pattern label output from the clustering coupling pattern recognition module. This label is one of four types: breathing-dominated, pressure-drift, resistance-deflection, or hybrid coupling. Simultaneously, it receives the target-related component and needle-tip-related component output from the particle coupling parameter search module. The target-drift residual and needle-tip deflection residual at the current moment are weighted and combined to calculate the Euclidean norm of the end-effector composite residual. This Euclidean norm is then compared with the norm thresholds corresponding to various deviation types stored in the pattern knowledge base to match the specific joint pose deviation type to which the current composite residual belongs. A matching confidence threshold is then set. The maximum confidence level is set to 0.75. When the maximum confidence level falls below this threshold, an online incremental update mechanism is triggered. After matching is completed, deviation type information is transmitted for collaborative iterative search. The collaborative iterative search strategy is initiated, and the position and attitude fine-tuning of the end effector are optimized simultaneously. The search range for position fine-tuning is set to ±2 mm for each of the X, Y, and Z axes, and the search range for attitude fine-tuning is set to ±3 degrees for each of the yaw, pitch, and roll angles. The search step size is set to 0.10 mm for position and 0.15 degrees for attitude, respectively. The collaborative iterative search uses an improved particle swarm optimization algorithm with a particle swarm size of 20 and a maximum number of iterations set to... The iteration count was set to 30, with an inertia weight of 0.65 and a learning factor of 1.50. In each iteration, each particle simultaneously updated candidate commands for both position and attitude, generating a candidate set of 20 joint pose fine-tuning candidate commands. Each candidate command included the target position coordinates and target attitude angle of the end effector in the robot's base coordinate system. From the generated joint pose fine-tuning candidate command set, the target-tip integrated deviation norm corresponding to each candidate command was calculated one by one. This norm was defined as the square root of the weighted sum of the squared target drift residual and the squared tip deflection residual, where the weighting coefficients were obtained using the particle-coupled parameter search module. The optimal weight allocation coefficient is set to 0.65. The fine-tuning command that minimizes the overall deviation norm is selected as the optimal command. The position and attitude fine-tuning values ​​in the optimal command are converted into the target pose of the robot end effector. The angular displacement commands of each joint are obtained through inverse kinematics calculation. The servo driver outputs the commands to each joint motor at a control cycle of 1000 Hz. The direction fine-tuning command is updated every 40 milliseconds, keeping it synchronized with the frame rate of the ultrasound image acquisition. This allows the needle tip to complete a direction correction every 0.50 mm of advance. The closed loop eliminates residual coupling deviation and achieves high-precision positioning throughout the puncture process.

[0023] The following is a detailed description of the workflow of this ultrasound-guided puncture surgery robot precision positioning system.

[0024] After the system starts, the robot's end effector holds the ultrasound probe and continuously acquires a sequence of ultrasound images of the patient's puncture site. The deformation displacement sensing module first performs pyramid optical flow field registration on adjacent frames and calculates sub-pixel displacement vectors to construct an initial nonlinear deformation displacement field. Then, it extracts the low-frequency displacement component caused by respiratory rhythm and the high-frequency abrupt change component caused by probe pressure through Gaussian scale spatial decomposition. These components are projected onto the robot's base coordinate system via a hand-eye calibration matrix to generate a deformation displacement mapping matrix. The fruit fly target tracking unit encodes the grid nodes in the mapping matrix as candidate sampling positions and constructs a fitness function using the reciprocal of the square of the target drift trajectory prediction residual. The system then converges to the optimal elastic matching point through odor-guided iterative search. The original target coordinates are corrected online based on the spatial coordinate offset, and the dynamic target position is output to the collaborative control center. The needle path adaptive control module operates synchronously. The resistance deflection mapping unit uses the real-time needle tip position, needle insertion speed and tissue resistance vector fed back by the multi-axis force sensor to form the state space. The nonlinear mapping relationship between resistance and deflection angle is learned through the deep Q network, and the deflection trend prediction value is output as the feedforward control quantity. The yaw dynamic filtering unit establishes a discrete state space model of the needle tip yaw angle, and the yaw angle measured by the force sensor and the prediction value of the deep Q network are fused according to the weight to construct a joint observation equation. The optimal yaw angle is dynamically estimated through the two-step recursive algorithm of Kalman filtering, and the direction fine adjustment command is generated to dynamically correct the puncture path. The clustering coupling mode identification module synchronously collects target drift residual sequences and needle tip deflection residual sequences within a fixed-period time window, concatenating them to form a joint pose deviation feature vector. After normalization and principal component analysis for dimensionality reduction, a density-based clustering algorithm is used to identify the deviation amplitude, directional angle, and time delay correlation features. Historical data is divided into typical coupling mode clusters, and a mode knowledge base is established. The deviation vector at the current moment is matched to the nearest neighbor cluster using Mahalanobis distance to output the coupling mode label. The particle coupling parameter search module automatically retrieves the corresponding decoupling parameter search space based on the mode label, aiming to minimize the end-comprehensive residual norm, and iteratively searches for the optimal value using a particle swarm optimization algorithm. The weighting coefficient and decoupling time window length are used to decompose the joint pose deviation into independent target-related components and needle-tip-related components based on the optimal parameters. After receiving the above components, the end effector residual collaborative compensation module compares the current comprehensive residual with the norm threshold in the pattern knowledge base to match the deviation type. Under the constraint of the deviation type, a collaborative iterative search strategy is initiated. At the same time, the position fine-tuning amount and attitude fine-tuning amount of the end effector are optimized to generate a joint pose fine-tuning candidate instruction set. The optimal instruction that minimizes the comprehensive deviation norm is selected. After inverse kinematics calculation, the servo driver drives the joint motors to execute, forming a closed-loop control to achieve high-precision positioning throughout the puncture process.

[0025] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A precision positioning system for an ultrasound-guided puncture surgery robot, comprising a collaborative control center, characterized in that, The collaborative control center has the following communication connection modules: The deformation displacement sensing module is used for hybrid optical flow field registration and fruit fly algorithm to sense the nonlinear deformation displacement field of soft tissue caused by patient breathing and probe pressure in real time, and dynamically locks the optimal elastic matching point through intelligent iterative search to correct the target point drift coordinates online. The needle path adaptive control module integrates depth Q network and Kalman filter to establish a nonlinear mapping relationship between tissue resistance and needle tip deflection. It also integrates real-time force feedback and predicted yaw angle to dynamically adjust the puncture direction, suppress the deviation of the actual advancement path from the planned straight line, and block the growth of path accumulation error. The clustering coupling mode identification module is used to perform cluster analysis on real-time and historical target-tip joint pose deviation data, identify the coupling mode types of errors in space and time, and classify and label the modes. The particle coupling parameter search module is used to search for the optimal combination of decoupling parameters based on the identified coupling mode types using the particle swarm algorithm, and decomposes the coupling error into target-related components and needle-tip-related components that can be processed independently. The end effector residual collaborative compensation module quickly matches the joint pose deviation type of the current residual based on the output coupling mode label, and performs collaborative iterative search on the matched end effector composite residual to generate joint pose fine-tuning instructions for the end effector, thus eliminating the target-needle tip composite deviation in a closed loop.

2. The ultrasound-guided puncture surgery robot precision positioning system according to claim 1, characterized in that: The deformation displacement sensing module includes an optical flow deformation extraction unit and a fruit fly target tracking unit. The optical flow deformation extraction unit, based on optical flow field registration technology, calculates the nonlinear displacement vector of soft tissue at the pixel level between consecutive ultrasound image frames, constructs a deformation displacement field that reflects the influence of breathing and pressure, and quantifies the target drift trajectory. The fruit fly target tracking unit is used to perform odor-guided iterative search in the constructed deformation displacement field using the fruit fly algorithm, quickly converge to the optimal elastic matching point at the current moment, correct the real-time coordinates of the target point, and output the dynamic target point position.

3. The ultrasound-guided puncture surgery robot precision positioning system according to claim 2, characterized in that: The optical flow deformation extraction unit specifically includes: A continuous sequence of ultrasound images of the patient's puncture site is acquired using a puncture surgical robot. Based on the pyramid hierarchical optical flow field registration algorithm, the sub-pixel displacement vector of each pixel in the soft tissue between adjacent frames is calculated to construct an initial nonlinear deformation displacement field. The position of each pixel in this deformation displacement field is the candidate matching point. The displacement field is decomposed in Gaussian scale along the time axis to extract the low-frequency periodic displacement component caused by the breathing rhythm and the high-frequency local mutation component caused by the probe pressure. The target drift trajectory is quantified, and the spatial coordinate mapping relationship of each candidate matching point before and after decomposition is preserved. The decomposed multi-component displacement field is projected onto the base coordinate system of the puncture surgery robot to generate the deformation displacement mapping matrix at the current moment. The positions of each grid node in the matrix are used as the output set of candidate matching points.

4. The ultrasound-guided puncture surgery robot precision positioning system according to claim 2, characterized in that: The fruit fly target tracking unit specifically includes: Each candidate matching point in the deformation displacement field is encoded as the odor concentration sampling location of the fruit fly individual, and the odor concentration fitness function is constructed using the inverse square of the prediction residual of the target drift trajectory. By using the odor-guided random search mechanism of the fruit fly population, dense iterative tracking is performed in the local extreme region of the deformation displacement field, and the system quickly converges to the optimal elastic matching point with the highest fitness at the current moment. Based on the spatial coordinate offset of the optimal matching point, the real-time coordinates of the original planned target point are corrected online, and the dynamic target point position after breathing and pressure compensation is output to the collaborative control center.

5. The ultrasound-guided puncture surgery robot precision positioning system according to claim 2, characterized in that: The needle path adaptive control module includes a drag deflection mapping unit and a yaw dynamic filtering unit. The drag deflection mapping unit is used to learn the mapping relationship between anisotropic drag and needle tip deflection angle using a deep Q-network, and to predict the deflection trend caused by drag. The yaw dynamic filtering unit is used to use Kalman filtering to fuse real-time force sensor feedback and predicted yaw angle, and combined with the predicted yaw trend to dynamically correct the puncture direction and suppress path accumulation error.

6. The ultrasound-guided puncture surgery robot precision positioning system according to claim 5, characterized in that: The drag deflection mapping unit specifically includes: The state space of the deep Q-network is constructed as the real-time needle tip position, needle insertion speed and tissue resistance vector fed back by multi-axis force sensors, and the action space is the predicted compensation amount of the needle tip deflection angle. Through online interactive sampling, the deep Q-network learns the nonlinear mapping relationship between anisotropic drag and needle tip deflection angle, and outputs the deflection trend prediction value under different drag conditions. The predicted deflection trend is injected as a feedforward control quantity into the yaw dynamic filter unit to achieve pre-compensation for the needle tip path deviation caused by drag.

7. The ultrasound-guided puncture surgery robot precision positioning system according to claim 5, characterized in that: The yaw dynamic filtering unit specifically includes: A discrete state-space model of the puncture needle tip is established, and the observation equation of Kalman filter is constructed using the yaw angle observation value fed back by the real-time force sensor and the yaw trend prediction value output by the drag deflection mapping unit. By employing a two-step recursive algorithm of time update and measurement update using Kalman filtering, the observed and predicted values ​​are fused to dynamically estimate the optimal yaw angle and its covariance uncertainty at the current moment. Based on the optimal yaw angle estimation results, the direction fine-tuning command of the puncture robot end effector is generated in real time to dynamically correct the puncture propulsion path and suppress the growth of path accumulation error.

8. The ultrasound-guided puncture surgery robot precision positioning system according to claim 5, characterized in that: The clustering coupled-mode recognition module specifically includes: Collect the target drift residual sequence and the needle tip deflection residual sequence within the real-time and historical time windows, concatenate the two to form a joint pose deviation feature vector, and perform normalization and dimensionality reduction preprocessing. A density-based clustering algorithm is used to perform unsupervised classification of joint pose deviation feature vectors. Based on the deviation amplitude, directional angle, and time delay correlation, different types of coupling mode clusters are identified, including four types: breathing-dominant, pressure drift, drag deflection, and hybrid coupling. Generate a unique mode label for each coupled mode cluster, match the joint deviation at the current moment to the nearest cluster, and output the coupled mode label to the cooperative control center.

9. The ultrasound-guided puncture surgery robot precision positioning system according to claim 8, characterized in that: The particle coupling parameter search module specifically includes: Based on the coupling mode label, retrieve the corresponding decoupling parameter search space, define the decoupling parameter vector of the particle swarm algorithm, including the target-tip weight allocation coefficient and the decoupling time window length; With minimizing the norm of the final integrated residual as the fitness objective, the optimal combination of parameters that minimizes the coupling error is searched in the decoupling parameter space through the velocity-position iterative update formula of the particle swarm. Based on the optimal decoupling parameter combination, the real-time joint pose deviation is decomposed into target-related components and needle-tip-related components that can be processed independently, and then output to the end residual collaborative compensation module.

10. The ultrasound-guided puncture surgery robot precision positioning system according to claim 9, characterized in that: The terminal residual collaborative compensation module specifically includes: Receive the coupling mode label and decoupling component from the output and quickly match the joint pose deviation type to which the current end-effector composite residual belongs; Under the constraint of the deviation type after matching, a cooperative iterative search strategy is adopted to simultaneously optimize the position fine-tuning and attitude fine-tuning of the end effector, and generate a joint pose fine-tuning candidate instruction set. The fine-tuning instruction that minimizes the target-needle tip combined deviation norm is selected from the candidate instruction set and sent to the puncture robot controller to eliminate residual coupling deviation in a closed loop.

Images