Method for estimating lateral force of patient in operation of surgical robot

Abstract

The invention relates to the technical field of surgical robot control, and discloses a surgical robot intraoperative patient side force estimation method, which comprises the following steps: constructing a kinetic model of a surgical robot patient side mechanical arm in a free space; a patient side mechanical arm is inserted into the abdominal body wall of the patient through a sleeve, joint data are collected, and meanwhile the free space torque of each joint at the current moment is predicted through a kinetic model; taking the collected position and speed data of each joint and the free space torque as input data, taking the residual error of the actually measured torque and the free space torque of each joint at each moment as label data, and separately training a neural network model for each joint; and calculating the contact force between the end effector and the tissue of the patient based on the actually measured torque of each current joint, the free space torque and the predicted extra torque of each joint. According to the method, the interaction force between the end effector of the surgical robot and the tissue can be accurately estimated, and the safety and accuracy of the surgery are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of surgical robot control, and more particularly, to a method for estimating the patient-side force during surgery of a surgical robot. Background Art

[0002] In the field of modern medical technology, surgical robots have been widely used because they can improve the accuracy and safety of surgeries. These robots can assist surgeons in performing complex surgical operations, especially in minimally invasive surgeries, where their advantages are more obvious. However, there are significant technical challenges in the tactile feedback of existing surgical robots, which limit the intuitiveness and safety of surgeries.

[0003] During the assistance of minimally invasive surgeries by surgical robots (such as the da Vinci system), doctors need to accurately sense the interaction forces (such as the resistance during cutting and suturing) between the end effector (including surgical tools) and the patient's tissue (such as the body tissue in the abdominal cavity) to provide tactile feedback, thereby improving the safety and precision of surgeries. In the prior art, the tactile feedback system of surgical robots mainly relies on end effector force sensors or model-based dynamic calculations. Although the force sensor method can directly measure the interaction forces between the instrument and the tissue, it has problems such as high cost, difficulty in sterilization, and difficulty in integration in narrow surgical spaces. The model-based dynamic calculation method estimates the interaction forces by establishing a dynamic model of the robot, but these models usually require a large amount of experimental data for calibration and are sensitive to changes in the robot structure and environment, making it difficult to adapt to different surgical conditions and instruments.

[0004] In addition, the existing patient-side force estimation algorithms have deficiencies in dealing with the interaction forces between the robot and the cannula. The cannula is a necessary channel for the surgical robot to enter the patient's body. Usually, in abdominal surgeries, the cannula is inserted through a small incision into the patient's body wall to provide a channel for surgical instruments to enter the body. During the surgery, the patient-side robotic arm is inserted into the patient's abdominal cavity or other surgical sites through the cannula, and the surgical tool installed at its end performs surgical operations. However, the frictional forces and other interaction forces generated by the contact between the cannula and the patient-side robotic arm will have a significant impact on the estimation of the interaction forces between the end effector and the tissue. For example, the cannula friction and resistance act on the instrument shaft and are transmitted through the joints of the robotic arm, ultimately affecting the motion accuracy and force feedback of the end effector. The prior art is difficult to accurately compensate for these additional torques, resulting in a decrease in the accuracy of the tactile feedback between the end surgical tool and the tissue, and further affecting the precision and safety of the doctor's surgery. Summary of the Invention

[0005] The objective of the present invention is to propose a method for estimating the lateral force of a patient during a surgical operation using a surgical robot, which can accurately estimate the interaction force between the end effector of the surgical robot and the tissue, improve the safety and precision of the operation, and provide more reliable tactile feedback for surgeons.

[0006] To achieve the above objective, the present invention proposes a method for estimating the lateral force of a patient during a surgical operation using a surgical robot, including: Construct a dynamic model of the robotic arm on the patient side of the surgical robot in free space, where the dynamic model is used to predict the free space torque of the joints based on the position, velocity, and acceleration of the joints. Insert the robotic arm on the patient side into the abdominal wall of the patient through a cannula. During this process, collect joint data including the interaction between the robotic arm on the patient side and the cannula, including the position, velocity, and measured joint torque of each joint; at the same time, predict the free space torque of each joint at the current moment through the dynamic model. Use the collected position and velocity data of each joint and the free space torque of each joint as input data, and use the residual between the measured torque and the free space torque of each joint at each moment as label data to form a training data set. Use the training data set to train a neural network model for each joint separately to predict the additional torque of each joint caused by the interaction between the robotic arm on the patient side and the cannula. During the surgical operation, when the end effector contacts the patient's tissue, calculate the contact force between the end effector and the patient's tissue based on the measured torque of each joint at the current moment, the free space torque of each joint predicted by the dynamic model, and the additional torque of each joint predicted by the neural network model.

[0007] Optionally, the construction of the dynamic model of the robotic arm on the patient side of the surgical robot includes: Establish a kinematic model of the robotic arm on the patient side based on the improved DH parameter method, define the coordinate system of each joint, and use the homogeneous transformation matrix to represent the pose relationship between adjacent joints. Based on the kinematic model, use the Euler-Lagrange equation to construct a dynamic model including link inertia, gravity, joint friction, and spring stiffness. Perform dynamic parameter identification on the dynamic model by combining an optimized excitation trajectory.

[0008] Optionally, the dynamic parameter identification of the dynamic model by combining an optimized excitation trajectory includes: Generate a periodic excitation trajectory based on the Fourier series that can fully stimulate the dynamic characteristics of the robot, and optimize the parameters of the excitation trajectory using a convex optimization method based on the constraint conditions of the upper and lower limits of joint position, velocity, and acceleration to ensure that the robot moves within a safe range. Control the excitation trajectory for optimizing the operation of the robotic arm on the patient side, and collect the position, velocity, acceleration of the joints and the measured joint torque data; Construct a regression matrix based on the collected joint positions, velocities, and accelerations. Each row of the regression matrix corresponds to a sampling point containing joint position, velocity, and acceleration data; Perform QR decomposition on the regression matrix to obtain the base parameter regression matrix; Use the base parameter regression matrix and the measured joint torque data, and solve the base parameters using the weighted least squares method; the base parameters are a linear combination of dynamic parameters, which are the minimum parameter set for describing the dynamics of the robotic arm; Among them, the dynamic parameters include link length, link mass, link center of mass position, link inertia tensor, joint friction coefficient, and spring stiffness.

[0009] Optionally, the objective function for solving the base parameters using the weighted least squares method is:

[0010] Among them, is the base parameter regression matrix, δ b is the base parameter vector, is the measured joint torque at the i-th sampling point, m is the number of sampling points, is the weight coefficient, used to balance the magnitude differences of each joint, .

[0011] Optionally, the neural network model is a multi-layer perceptron independently trained for each joint. The multi-layer perceptron includes an input layer, a hidden layer, and an output layer. The hidden layer includes two fully connected layers, and the activation function is ReLU.

[0012] Optionally, the input data of the neural network model specifically includes: the joint position sequence and velocity sequence within the historical time window, and the free space torque of the current joint predicted by the dynamic model at the current moment; The output of the neural network model is the compensation torque of the current joint.

[0013] Optionally, when training the neural network model, the mean square error is used as the loss function, and the model parameters that minimize the loss function are found as the training objective.

[0014] Optionally, the residual calculation formula for the measured torque and free space torque of each joint is:

[0015] Among them, is the vector of the measured torque of each joint, including the torque actually measured by each joint motor, is the vector of the free-space torque of each joint, including the torque of each joint predicted by the dynamic model during free-space motion, , is the basis parameter regression matrix, is the basis parameter vector.

[0016] Optionally, the method for calculating the contact force between the end effector and the patient tissue includes: Inputting the positions, velocities, and accelerations of each joint collected in real time into the dynamic model, and predicting the free-space torque of each joint through the dynamic model; Inputting the position, velocity data of the joints collected in real time and the free-space torque predicted by the dynamic model into the neural network model and outputting the predicted additional torque of the joints; Compensating the free-space torque of the joints with the additional torque of the joints output by the neural network model; Based on the measured torque of each joint and the compensated joint torque at present, and combining with the inverse matrix of the spatial Jacobian matrix, calculating the Cartesian force of the end effector.

[0017] Optionally, the formula for calculating the Cartesian force of the end effector is:

[0018] where, is the contact force vector between the end effector and the patient tissue, J is the spatial Jacobian matrix at the current joint position, is the vector of the measured torque of each joint at the current moment, is the vector of the free-space torque of each joint predicted by the dynamic model, is the compensation torque vector output by the neural network model.

[0019] The beneficial effects of the present invention are as follows: The present invention first constructs a dynamic model of the patient-side robotic arm of the surgical robot in free space (without contact), uses the dynamic model to predict the free-space torque of the joints, then inserts the patient-side robotic arm into the patient's body through a cannula and collects joint data during the insertion process, including joint positions, velocities, and measured joint torques. By combining the joint data collected during the interaction between the robotic arm and the cannula and the free-space joint torque data calculated by the dynamic model, a neural network model is used to learn the additional torques generated by the interaction between the cannula and the patient-side robotic arm to compensate for the torques predicted by the dynamic model. Then, based on the measured torques of the current joints, the free-space torques of the joints predicted by the dynamic model, and the additional torques of the joints predicted by the neural network model, the contact force between the end effector and the patient's tissue is estimated, thereby achieving an accurate estimation of the interaction force between the end effector of the surgical robot and the tissue, providing precise tactile feedback for the surgical robot, and improving the safety and precision of the surgery. At the same time, the present invention combines the methods of dynamic model and neural network learning, while ensuring the force estimation accuracy, compared with the pure learning method, reduces the requirements for the coverage range of training data, and improves the adaptability to different surgical scenarios and operating conditions.

[0020] Furthermore, the present invention completes the dynamic parameter identification by optimizing the excitation trajectory and extracting the basis parameters through QR decomposition, improving the robustness of the model; adopts a small-scale neural network model trained by joint to compensate for the cannula friction in real time with only a small amount of intraoperative data, realizing lightweight neural network compensation; based on the spatial Jacobian matrix, efficiently converts the joint torque into the end-effector Cartesian force, realizes real-time force mapping, ensures the real-time nature of intraoperative tactile feedback, and improves the intuition and safety of the doctor's surgery.

[0021] The method of the present invention has other characteristics and advantages, which will be obvious from the accompanying drawings incorporated herein and the subsequent detailed description, or will be detailedly described in the accompanying drawings incorporated herein and the subsequent detailed description, and these accompanying drawings and detailed description are jointly used to explain the specific principles of the present invention. Brief Description of the Drawings

[0022] By describing the exemplary embodiments of the present invention in more detail in conjunction with the accompanying drawings, the above-mentioned and other objects, features, and advantages of the present invention will become more obvious. In the exemplary embodiments of the present invention, the same reference numerals generally represent the same components.

[0023] Figure 1 Shows a flowchart of a method for estimating the force on the patient side during surgery of a surgical robot according to the present invention. Detailed Description of the Embodiments

[0024] The present invention will be described in more detail below with reference to the accompanying drawings. Although the preferred embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art. Embodiment

[0025] As Figure 1 shown, this embodiment provides a method for estimating the lateral force of a patient during surgery using a surgical robot, including: S1: Construct a dynamic model of the patient-side robotic arm of the surgical robot in free space, where the dynamic model is used to predict the free-space torque of the joints based on the position, velocity, and acceleration of the joints; This step specifically includes: S101: Establish a kinematic model of the patient-side robotic arm based on the improved DH parameter method, define the coordinate system of each joint, and use homogeneous transformation matrices to represent the pose relationship between adjacent joints; Specifically, the kinematic model describes the relationship between the positions of the joints of the robot and the pose (position and orientation) of the end effector, without involving forces and masses. Its core is coordinate transformation and the Jacobian matrix. Based on the structural characteristics of the patient-side robotic arm, the coordinate system of each joint is defined. This is usually accomplished using Denavit-Hartenberg (DH) parameters. This method uses the improved DH parameters to better adapt to the specific structure of the patient-side robotic arm.

[0026] In one example, using the improved DH parameter method, the coordinate system of each joint i is defined by the following four parameters: The homogeneous transformation matrix of the adjacent joint i −1 to i is: For:

[0027] Where, Rot( z , θ i ) is the rotation matrix of the current joint rotating around the z-axis of the previous joint by the rotation angle θ i , Trans( z , d i ) is the translation matrix of the current joint translating along the z-axis of the previous joint by the translation distance d i , Trans( x , a i ) is the translation matrix of the link translating along the x-axis of the current joint by the translation distance a i , Rot(x , α i ) is the rotation angle of the connecting rod around the current joint x-axis α i rotation matrix.

[0028] The expanded matrix form is:

[0029] Then, based on the homogeneous transformation matrix, the kinematic chain of the robotic arm is constructed. By calculating the homogeneous transformation matrix layer by layer, the exact pose of the end effector is finally obtained. For example, from the base (joint 0) to the end effector (joint n ) the overall transformation matrix is:

[0030] Extract the translation part from the matrix to obtain the end position (such as the fourth column). Extract the rotation matrix (such as the first three columns) from the matrix to obtain the end attitude.

[0031] Then, describe the mapping relationship between joint velocity and end velocity through the Jacobian matrix:

[0032] where, is the end linear velocity and angular velocity vector, is the spatial Jacobian matrix, which is composed of the motion screws of each joint, is the joint position vector, is the joint velocity (angular velocity) vector.

[0033] The improved DH method is more suitable for serial robots, with more intuitive parameter definitions and smaller calculation errors. Through the improved DH parameters and homogeneous transformation matrix, the high-precision calculation of the end pose is ensured.

[0034] S102: Based on the kinematic model, use the Euler-Lagrange equation to construct a dynamic model including link inertia, gravity, joint friction, and spring stiffness; Specifically, in this step, an accurate model of the robot's own dynamics (excluding external disturbances) is established. The dynamic model describes the relationship between joint torque and motion (joint position , joint velocity , joint acceleration ), and involves physical effects such as inertia, gravity, and friction. This method uses the Euler-Lagrange equation to derive the torque equation including inertia, gravity, friction, and spring force.

[0035] The Euler-Lagrange equation is:

[0036] Among them, \(L\) is the Lagrangian, K is the total kinetic energy of the system, P is the total potential energy of the system.

[0037] The dynamic equation is:

[0038] After expansion, the obtained non-linear dynamic equation is:

[0039] Among them, is the joint torque (joint motor torque), is the joint position vector, is the joint velocity vector, is the joint acceleration vector, is the inertia matrix, which describes the mass distribution and inertia characteristics of the robot, is the Coriolis force matrix, which includes Coriolis force and centrifugal force terms; is the gravity term, which is calculated from the gradient of the gravitational potential energy, is the friction term, which is usually modeled as a combination of viscous friction and Coulomb friction.

[0040] The above non-linear dynamic equation can be converted into a linear parameterized form:

[0041] Among them, δ are the dynamic parameters, is the regression matrix, where each row corresponds to the dynamic equation at a time point and each column corresponds to a dynamic parameter.

[0042] The conversion method is as follows: First, define the dynamic parameters δ , which include all physical quantities affecting the dynamics. The dynamic parameters δ include the link mass, center of mass position, inertia tensor, friction coefficient, etc. Then, expand each non-linear term in the non-linear dynamic equation into a linear combination of the dynamic parameters, and extract the coefficients of each expanded term as the columns of the regression matrix W to construct the regression matrix W .

[0043] Reorganize the dynamic equation into a linear combination of the dynamic parameters δ for subsequent parameter identification.

[0044] S103: Perform dynamic parameter identification on the dynamic model by combining the optimized excitation trajectory, specifically as follows: S1031: Generate a periodic excitation trajectory that can fully stimulate the dynamic characteristics of the robot based on Fourier series, and optimize the parameters of the excitation trajectory using a convex optimization method based on the constraint conditions of the upper and lower limits of joint position, velocity, and acceleration to ensure that the robot moves within a safe range; Specifically, use Fourier series to generate a periodic trajectory that can fully stimulate the dynamic characteristics of the robot. Preferably, the goal is to minimize the condition number of the regression matrix W The mathematical expression is:

[0045] where represents j the change of joint coordinates over time, that is, the function of joint position changing over time; represents j the offset or initial position of the joint coordinates, which is the reference starting point of the trajectory; represents the k amplitude coefficient of the sine term of the th harmonic, which determines the contribution of the sine part of this harmonic to the trajectory; k represents the k amplitude coefficient of the cosine term of the t th harmonic, which determines the contribution of the cosine part of this harmonic to the trajectory; nH represents the harmonic order; ω represents the fundamental frequency, which determines the periodic change rate of the trajectory;

[0046] where the fundamental frequency ω and the total harmonic order nH need to be optimized and selected. Preferably, the fundamental frequency is 0.18 Hz and the harmonic order is 6.

[0047] Then, based on the constraint conditions of the upper and lower limits of joint position, velocity, and acceleration, use a convex optimization method (such as using the convex optimization tool CVXPY package combined with the SCS solver) to optimize the amplitude coefficients and of the excitation trajectory to ensure that the robot moves within a safe range.

[0048] S1032: Control the robotic arm on the patient side to run the optimized excitation trajectory in free space (without external contact), and collect the joint position, velocity, acceleration, and measured joint torque data; Among them, the position, velocity, and acceleration of the joints are measured by encoders or other position / velocity sensors built into the robot. The measured torque data of the joints is measured by torque sensors at the robot joints. These data are collected and transmitted through the robot operating system for subsequent dynamic parameter identification and torque estimation processes. In this step, the manipulator is driven to move by an excitation trajectory, and the position, velocity, acceleration, and measured torque of each joint at each moment are recorded.

[0049] S1033: Construct a regression matrix based on the collected joint positions, velocities, and accelerations. Each row of the regression matrix corresponds to a sampling point containing joint position, velocity, and acceleration data; Specifically, construct the regression matrix W , based on the dynamic equation , with each row corresponding to a sampling point, and the elements are calculated from joint positions, velocities, and accelerations, describing the contribution of each dynamic parameter to the torque.

[0050] S1034: Perform QR decomposition on the regression matrix to obtain the basis parameter regression matrix; Specifically, perform column pivoting QR decomposition on the regression matrix W :

[0051] Among them, P is a permutation matrix, Q is an orthogonal matrix, R is the non-zero diagonal elements of the upper triangular matrix indicating independent columns.

[0052] Select the basis parameters, and select the first b linearly independent columns to form the basis parameter regression matrix , among which, P b is the permutation matrix. The basis parameters δ b are linear combinations of the original parameters δ , satisfying , which is the minimum parameter set describing the system, where T represents the transpose.

[0053] Eliminate redundant parameters through QR decomposition to reduce the identification complexity.

[0054] S1035: Use the basis parameter regression matrix and the measured joint torque data, and solve the basis parameters using the weighted least squares method; the basis parameters are linear combinations of dynamic parameters, which are the minimum parameter set describing the dynamics of the manipulator; Among them, the objective function for solving the basis parameters using the weighted least squares method is:

[0055] Among them, is the base parameter regression matrix, δ b is the base parameter vector, is the measured joint torque at the i-th sampling point, m is the number of sampling points, is the weight coefficient, used to balance the magnitude differences of each joint, .

[0056] The dynamic parameters are fitted by the least squares method δ b , so that the model predicted torque is as close as possible to the measured torque , where .

[0057] The above dynamic modeling process is existing in the art, and the specific details will not be elaborated.

[0058] By this step S1, the robot's own dynamic torque is accurately separated, providing a benchmark for subsequent compensation.

[0059] S2: Insert the patient-side robotic arm into the patient's abdominal body wall through a cannula. During this process, joint data including the interaction between the patient-side robotic arm and the cannula is collected, including the position, velocity, and measured joint torque of each joint; meanwhile, the free-space torque of each joint at the current moment is predicted through the dynamic model; Specifically, when the patient-side robotic arm is inserted into the cannula and passes through the abdominal phantom, the joint state and torque data are collected in real time. During the insertion process, only the patient-side robotic arm contacts the cannula, and the end effector does not contact the tissues in the patient's body, so as to ensure that the collected joint state and torque data only contain the data generated by the interaction between the patient-side robotic arm and the collar. The process of inserting the patient-side robotic arm into the patient's abdominal body wall through the cannula usually takes about 4 minutes (about 48,000 data points), and the preferred sampling rate is 200 Hz (window size 5), which can generate about 9,600 training samples. While collecting the above data, through the dynamic model, according to the measured joint torque, position , velocity and acceleration of each joint collected at the current moment, the free-space torque of each joint at the current moment is predicted in real time for the training of the subsequent neural network model.

[0060] S3: Use the collected position and velocity data of each joint and the free-space torque of each joint as input data, and use the residual between the measured torque of each joint and the free-space torque at each moment as label data to form a training dataset. Use the training dataset to train a neural network model for each joint separately to predict the additional torque of each joint caused by the interaction between the patient-side robotic arm and the cannula; Specifically, in this step, a dynamic model is used to calculate the free-space predicted torque , and the formula for calculating the residual torque between the measured torque of the joint and the free-space torque is:

[0061] where is the vector of the measured torque of the joint, including the torque actually measured by each joint motor, is the vector of the free-space torque of each joint, including the torque of each joint predicted by the dynamic model during free-space movement, , is the base parameter regression matrix, is the base parameter vector.

[0062] The neural network model is a multi-layer perceptron (MLP) independently trained for each joint. The multi-layer perceptron includes an input layer, a hidden layer, and an output layer. The hidden layer includes two fully connected layers, and the activation function is ReLU.

[0063] The input data of the neural network model specifically includes: the joint position sequence and velocity sequence within a historical time window (such as a window size of 5), and the free-space torque of the current joint predicted by the dynamic model at the current moment ; the output of the neural network model is the predicted compensation torque of the current joint .

[0064] When training the neural network model, the mean squared error is used as the loss function: , and the model parameters that minimize the loss function are found as the training objective.

[0065] By training a neural network model for each joint separately, learning the patient-specific cannula friction pattern, and predicting the additional torque of each joint caused by the interaction between the patient-side robotic arm and the cannula, the free-space joint torque predicted by the dynamic model can be compensated, thereby dynamically correcting the prediction error of the dynamic model.

[0066] S4: During the operation, when the end effector contacts the patient's tissue, calculate the contact force between the end effector and the patient's tissue based on the measured torques of the current joints, the free-space torques of the joints predicted by the dynamic model, and the additional torques of the joints predicted by the neural network model.

[0067] In this step, the method for calculating the contact force between the end effector and the patient's tissue includes: Input the positions, velocities, and accelerations of the current joints collected in real time into the dynamic model, and predict the free-space torques of the joints through the dynamic model; Input the position and velocity data of the joints collected in real time and the free-space torques of the joints predicted by the dynamic model into the neural network model and output the predicted additional torques of the joints; Compensate the free-space torques of the joints with the additional torques of the joints output by the neural network model; Based on the measured torques of the current joints and the compensated joint torques, combine with the inverse matrix of the spatial Jacobian matrix to calculate the Cartesian force of the end effector.

[0068] Among them, the formula for calculating the Cartesian force of the end effector is:

[0069] Among them, is the contact force vector between the end effector and the patient's tissue at the current moment (including the forces and torques in the XYZ directions), J is the spatial Jacobian matrix at the current joint position, is the vector of the measured torques of the current joints at the current moment, is the vector of the free-space torques of the joints predicted by the dynamic model, is the vector of the compensated torques output by the neural network model.

[0070] It should be noted that the measured torque is the true torque of each joint at the current moment under the actual surgical action, including the torque required to maintain the basic posture (corresponding to the torque required to maintain the posture without external interference, that is, the free-space torque of the joint ), the cannula interaction torque generated by resisting the cannula friction, etc. (corresponding to the compensated torque predicted by the neural network model), and the torque generated by the interaction when the end effector contacts the patient's tissue. Therefore, is the torque generated by the interaction between the end effector and the patient's tissue, and it can be converted into the force in the Cartesian space through the inverse matrix of the spatial Jacobian matrix.

[0071] Provide intraoperative tactile feedback through the accurately calculated contact force between the end effector and the tissue, thereby improving the intuition and safety of the doctor's surgery.

[0072] In this method, the training time of the small-scale neural network model adopted is usually less than 20 seconds, so it can support rapid intraoperative deployment. The training data collected by this method is collected based on a specific patient, and the trained neural network model can be applied to the torque compensation during the operation of a specific patient. Since there are differences in the physiological structure, tissue hardness, body wall thickness, etc. of different patients, these differences will affect the interaction force between the cannula and the robotic arm. Therefore, for different patients, a short data collection process (such as about 4 minutes) can be carried out before each operation to collect the specific data of the current patient, and the collected data can be used to quickly retrain the neural network model to adapt to the current patient and surgical conditions. During the operation, the retrained model is used to estimate in real time the additional joint torque caused by the interaction force between the robotic arm and the cannula and compensate the free-space joint torque, thereby improving the accuracy of the end-effector force estimation.

[0073] The embodiments of the present invention have been described above. The above description is exemplary and not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations are obvious to those of ordinary skill in the art in the technical field without departing from the scope and spirit of the described embodiments.

Claims

1. A method for estimating the lateral force of a patient during surgery using a surgical robot, characterized in that, Including: Construct a dynamic model of the patient-side robotic arm of the surgical robot in free space, which is used to predict the free-space torque of the joints based on the position, velocity, and acceleration of the joints; Insert the patient-side robotic arm into the patient's abdominal body wall through a cannula. During this process, collect joint data including the interaction between the patient-side robotic arm and the cannula, including the position, velocity, and measured joint torque of each joint; meanwhile, predict the free-space torque of each joint at the current moment through the dynamic model; Use the collected position and velocity data of each joint and the free-space torque of each joint as input data, and use the residual between the measured joint torque and the free-space torque of each joint at each moment as label data to form a training data set. Use the training data set to train a neural network model for each joint separately to predict the additional torque of each joint caused by the interaction between the patient-side robotic arm and the cannula; When the end effector contacts the patient's tissue during the surgical process, calculate the contact force between the end effector and the patient's tissue based on the measured joint torque of each joint at the current moment, the free-space torque of each joint predicted by the dynamic model, and the additional torque of each joint predicted by the neural network model.

2. The intraoperative patient-side force estimation method for a surgical robot according to claim 1, characterized in that The construction of the dynamic model of the patient-side robotic arm of the surgical robot includes: Establish a kinematic model of the patient-side robotic arm based on the improved DH parameter method, define the coordinate system of each joint, and use the homogeneous transformation matrix to represent the pose relationship between adjacent joints; Based on the kinematic model, use the Euler-Lagrange equation to construct a dynamic model including link inertia, gravity, joint friction, and spring stiffness; Perform dynamic parameter identification on the dynamic model by combining an optimized excitation trajectory.

3. The intraoperative patient-side force estimation method for a surgical robot according to claim 2, wherein, The dynamic parameter identification of the dynamic model by combining an optimized excitation trajectory includes: Generate a periodic excitation trajectory based on the Fourier series that can fully stimulate the dynamic characteristics of the robot, and optimize the parameters of the excitation trajectory using a convex optimization method based on the constraint conditions of the upper and lower limits of joint position, velocity, and acceleration to ensure that the robot moves within a safe range; Control the patient-side robotic arm to run the optimized excitation trajectory in free space, and collect joint position, velocity, acceleration, and measured joint torque data; Construct a regression matrix based on the collected joint position, velocity, and acceleration data. Each row of the regression matrix corresponds to a sampling point containing joint position, velocity, and acceleration data; Perform QR decomposition on the regression matrix to obtain the base parameter regression matrix; Use the base parameter regression matrix and the measured joint torque data, and use the weighted least squares method to solve the base parameters; the base parameters are a linear combination of dynamic parameters, which are the minimum parameter set describing the dynamics of the robotic arm; Among them, the dynamic parameters include link length, link mass, link centroid position, link inertia tensor, joint friction coefficient, and spring stiffness.

4. The intraoperative patient-side force estimation method for a surgical robot according to claim 3, characterized in that The objective function for solving the base parameters using the weighted least squares method is: ， Among them, is the base parameter regression matrix, δ b is the base parameter vector, is the measured joint torque at the i-th sampling point, m is the number of sampling points, is the weight coefficient, used to balance the magnitude differences of each joint, .

5. The intraoperative patient-side force estimation method for a surgical robot according to claim 1, characterized in that The neural network model is a multi-layer perceptron independently trained for each joint. The multi-layer perceptron includes an input layer, a hidden layer, and an output layer. The hidden layer includes two fully connected layers, and the activation function is ReLU.

6. The intraoperative patient-side force estimation method for a surgical robot according to claim 5, characterized in that, The input data of the neural network model specifically includes: the joint position sequence and velocity sequence within the historical time window, and the free-space torque of the current joint predicted by the dynamics model at the current moment; The output of the neural network model is the compensation torque of the current joint.

7. The intraoperative patient-side force estimation method for a surgical robot according to claim 6, wherein When training the neural network model, the mean square error is used as the loss function, and the model parameters that minimize the loss function are found as the training objective.

8. The intraoperative patient-side force estimation method for a surgical robot according to claim 4, wherein The formula for calculating the residual of the measured torque and the free-space torque of each joint is: ， Among them, is the vector of the measured torque of each joint, including the torque actually measured by the motor of each joint, is the vector of the free-space torque of each joint, including the torque of each joint predicted by the dynamic model during free-space motion, , is the base parameter regression matrix, is the base parameter vector.

9. The intraoperative patient-side force estimation method for a surgical robot according to claim 8, wherein The method for calculating the contact force between the end effector and the patient tissue includes: Input the positions, velocities, and accelerations of each joint collected in real time into the dynamics model, and predict the free-space torque of each joint through the dynamics model; Input the position, velocity data of the joint collected in real time and the free-space torque predicted by the dynamics model into the neural network model and output the predicted additional torque of the joint; Compensate the free-space torque of the joint with the additional torque of the joint output by the neural network model; Based on the measured torque of each current joint and the compensated joint torque, combined with the inverse matrix of the spatial Jacobian matrix, calculate the Cartesian force of the end effector.

10. The intraoperative patient-side force estimation method for a surgical robot according to claim 9, characterized in that, The formula for calculating the Cartesian force of the end effector is: ， Among them, is the contact force vector between the end effector and the patient's tissue, J is the spatial Jacobian matrix at the current joint position, T represents the transpose, is the measured torque vector of each joint at the current moment, is the free space torque vector of each joint predicted by the dynamic model, is the compensation torque vector output by the neural network model.

Images