A method for modeling contact dynamics of robotic bone resection
Patent Information
- Application Number
- CN202610280979.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-09
- Publication Date
- 2026-09-25
AI Technical Summary
然而，骨组织具有显著的非均匀性与各向异性，皮质骨与松质骨的力学性能差异明显，导致磨削接触力呈现强非线性、时变性及不确定性，传统固定参数解析模型难以刻画实际磨削的复杂力学行为
[0028]本发明的技术方案中，获得骨磨削接触动力学建模方法；依据骨磨削过程中末端钻头与骨组织之间的接触动力学特性，构建一种基于径向基函数神经网络的接触六维力预测模型，用于描述磨削工艺参数、末端位姿及骨密度与接触力/力矩之间的非线性映射关系；依据骺板骨桥磨削临床实验需求，构建覆盖末端位姿、磨削工艺参数以及骨密度参数空间的骨磨削接触力实验数据集，为接触力预测模型训练提供数据支撑；依据构建的骨磨削数据集，通过遗传算法对径向基函数神经网络关键参数进行优化训练，实现末端六维力/力矩的准确预测。
Smart Images
Figure CN122822368A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for modeling contact dynamics in robot bone grinding, belonging to the field of robot bone grinding. Background Technology
[0002] Bone grinding is a common procedure in modern orthopedic surgery. With the development of orthopedic robotic technology, it has shown significant advantages in surgeries such as epiphyseal plate bridge grinding. However, current robotic grinding precision is insufficient, making it difficult to accurately control grinding force. Excessive grinding force can cause mechanical damage to bone tissue. The contact force between the robot's end-effector and bone tissue directly affects the accuracy of bone removal, intraoperative safety, and force control stability. Therefore, accurate modeling and analysis of the dynamic characteristics of the drill-bone contact are crucial theoretical foundations for achieving high-precision and high-safety grinding operations. However, bone tissue exhibits significant non-homogeneity and anisotropy, with marked differences in mechanical properties between cortical and cancellous bone. This results in grinding contact forces exhibiting strong nonlinearity, time-varying characteristics, and uncertainty, making it difficult for traditional fixed-parameter analytical models to characterize the complex mechanical behavior of actual grinding.
[0003] Current research on bone grinding mechanics models mainly falls into two categories: theoretical analysis and data-driven modeling. Theoretical analysis, due to its excessive focus on cutting edge modeling and mechanical behavior analysis, suffers from complex mathematical representations, making it difficult to apply to parameter optimization. While data-driven modeling avoids mechanics simplification, it is limited by the multi-physics coupling characteristics of the grinding process, making it difficult to construct high-quality datasets. Furthermore, accurately representing the complex nonlinear mapping between grinding parameters and contact forces is challenging. Neural networks exhibit significant advantages in fitting nonlinear relationships. Radial basis function neural networks, in particular, possess strong local approximation capabilities and fast convergence. Combining them with genetic algorithms to optimize key parameters can effectively address the shortcomings of traditional models. Therefore, designing a reasonable neural network structure and constructing an accurate contact dynamics prediction model has significant research value and application implications. Summary of the Invention
[0004] In view of this, the present invention provides a robot bone grinding contact dynamics modeling method. This method can automatically evaluate the six-dimensional force / torque of bone grinding contact based on grinding bone density, process parameters and end-effector pose, providing a reliable model basis for subsequent grinding force control strategy design and system dynamics analysis.
[0005] This invention provides a robot control method for bone grinding, comprising:
[0006] Based on the contact dynamics between the end drill bit and bone tissue during bone grinding, a six-dimensional contact force prediction model based on radial basis function neural network is constructed to describe the nonlinear mapping relationship between grinding process parameters, end position and bone density and contact force / torque.
[0007] Based on the clinical trial requirements for epiphyseal plate bone bridge grinding, a bone grinding contact force experimental dataset covering the space of end pose, grinding process parameters, and bone density parameters was constructed to provide data support for the training of contact force prediction models.
[0008] Based on the constructed bone grinding dataset, the key parameters of the radial basis function neural network are optimized and trained using a genetic algorithm to achieve accurate prediction of the six-dimensional force / torque at the end.
[0009] In the above method, based on the contact dynamics characteristics between the end drill bit and bone tissue during bone grinding, a six-dimensional contact force prediction model based on a radial basis function neural network is constructed to describe the nonlinear mapping relationship between grinding process parameters, end drill bit pose, bone density, and contact force / torque, including:
[0010] The contact dynamics between bone grinding drill bits and bone tissue were analyzed, and a six-dimensional force prediction model based on radial basis function neural networks (RBFNN) was constructed. The input layer includes three grinding process parameters (bone density, bone mineral density, and bone tissue density). Spindle speed Feed rate ) and 6 end pose parameters (position) , , and posture , , ). No. The input vector of each sample It can be represented as:
[0011]
[0012] The RBFNN model consists of an input layer, hidden layers, and an output layer. For the input layer... training samples In its first The output of each hidden layer node The calculation is as follows:
[0013]
[0014] in, and Represented as the first Cluster centers and standard deviations of Gaussian activation functions For input samples With the center Euclidean distance.
[0015] The output layer sums and weights the information from the hidden layer neurons and outputs it, which can be represented as:
[0016]
[0017] in, The number of radial basis function units in the hidden layer. For the input vector, For the first The output value of each output neuron For the hidden layer The nth neuron to the output layer The weights of each neuron, For the output layer Bias of each neuron.
[0018] In the above method, the step of constructing a bone grinding contact force experimental dataset covering the end-position, grinding process parameters, and bone density parameter space based on the clinical trial requirements of epiphyseal plate bone bridge grinding, thereby providing data support for neural network model training and the establishment of a bone grinding contact force prediction model, includes:
[0019] To train and validate the grinding force model, real bone grinding experimental data needed to be collected. SAWBONES solid polyurethane foam was selected as the standard artificial bone material, and bone blocks with densities of 800 kg / m³ and 1640 kg / m³ were used to simulate cancellous bone and cortical bone, respectively. Considering the load and safety requirements for grinding pediatric epiphyseal plate bridges, the spindle speed was set to 12,000-30,000 rpm, and the feed rate was controlled within the range of 0.3-1 mm / s. Grinding experiments were conducted within this parameter range.
[0020] To improve the reliability of the six-dimensional force sensor signal at the end point, this invention preprocesses the raw data. First, detrending processing is used to eliminate zero-point drift and baseline offset caused by environmental interference; then, a Butterworth low-pass filter is used to filter out high-frequency noise; finally, a moving average algorithm is used for smoothing to improve signal stability.
[0021] In the above method, the step of optimizing and training the key parameters of the radial basis function neural network using a genetic algorithm based on the constructed bone grinding dataset to achieve accurate prediction of the six-dimensional force / torque at the end includes:
[0022] The prediction accuracy of RBFNN is highly dependent on the number of neurons in the hidden layer. and Gaussian function extension parameters The selection of parameters is crucial. Therefore, a genetic algorithm (GA) is introduced to globally optimize the key parameters of the RBFNN, achieving joint modeling of GA-RBFNN.
[0023] The optimization objective is to minimize the root mean square error of the contact force prediction. The fitness function can be expressed as:
[0024]
[0025] in, For the sample size, and Represented as the first The actual end force of the sample and the predicted value of the network model.
[0026] Genetic algorithms iteratively update the population through selection, crossover, and mutation operations, searching for the optimal population globally. and Parameter combinations to minimize the fitness function The RBFNN model optimized by GA exhibits higher prediction accuracy and generalization ability under conditions of strong nonlinearity and parameter uncertainty, and can accurately learn the nonlinear mapping relationship between contact force and multi-source input variables during bone grinding.
[0027] As can be seen from the above technical solutions, the present invention has the following beneficial effects:
[0028] The technical solution of this invention provides a method for modeling contact dynamics in bone grinding. Based on the contact dynamic characteristics between the end drill bit and bone tissue during bone grinding, a six-dimensional contact force prediction model based on a radial basis function neural network is constructed to describe the nonlinear mapping relationship between grinding process parameters, end drill bit pose, bone density, and contact force / torque. Based on the clinical trial requirements for epiphyseal plate bone bridge grinding, a bone grinding contact force experimental dataset covering the space of end drill bit pose, grinding process parameters, and bone density parameters is constructed to provide data support for training the contact force prediction model. Based on the constructed bone grinding dataset, a genetic algorithm is used to optimize and train the key parameters of the radial basis function neural network to achieve accurate prediction of the six-dimensional end drill bit force / torque. Attached Figure Description
[0029] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort or labor.
[0030] Figure 1 This is a schematic diagram of the bone grinding contact dynamics modeling method provided in this invention example;
[0031] Figure 2 This is a schematic diagram of the RBFNN network structure provided in the example of this invention;
[0032] Figure 3 This is a schematic diagram of a robot bone grinding experiment provided in this invention example;
[0033] Figure 4These are schematic diagrams of dataset signal preprocessing provided in this invention example: (a) Comparison of end force signal before and after preprocessing, and (b) Comparison of end torque signal before and after preprocessing.
[0034] Figure 5 This is a schematic diagram of the RBFNN parameter optimization process based on genetic algorithm provided in the example of the present invention;
[0035] Figure 6 This is a schematic diagram of fitness changes during the training process provided in this invention example;
[0036] Figure 7 These are schematic diagrams of the test set prediction fitting results provided in the examples of the present invention. Figures (a), (b), and (c) show the prediction fitting effect of the end contact force, and Figures (d), (e), and (f) show the prediction fitting effect of the end contact torque.
[0037] Figure 8 This is a schematic diagram showing the comparison of the prediction results provided by the present invention on different models; Specific Implementation
[0038] To better understand the technical solution of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0039] It should be understood that the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0040] This invention provides a method for modeling contact dynamics in robot bone grinding. Please refer to [the relevant documentation]. Figure 1 This is a schematic diagram of the robot bone grinding contact dynamics modeling method provided by the present invention. Without being limited to the preceding description, the bone grinding contact dynamics modeling method is applicable to the field of orthopedic surgery and provides a reliable basis for robot system dynamics analysis and grinding force control. The method includes the following steps:
[0041] Step 101: Based on the contact dynamics characteristics between the end drill bit and bone tissue during bone grinding, a six-dimensional contact force prediction model based on radial basis function neural network is constructed to describe the nonlinear mapping relationship between grinding process parameters, end position and bone density and contact force / torque.
[0042] The contact dynamics between bone grinding drill bits and bone tissue were analyzed. The end-effector contact force is mainly determined by the working space state, motion state, and grinding process parameters. Considering experimental conditions and modeling feasibility, this invention abstracts the bone grinding contact dynamics into a nonlinear mapping relationship. The input variables include the position and orientation of the end-effector drill bit in the tool coordinate system, the feed rate, and the process parameters, including spindle speed, feed rate, bone density, and grinding depth. The grinding depth is fixed and not used as an adjustable input. The output is the predicted six-dimensional contact force. Therefore, this mapping relationship can be formally described as:
[0043]
[0044] in, Bone density, Main spindle speed This refers to the end feed rate. and These represent the position and attitude of the end-point drill bit, respectively. (Function) As a highly nonlinear mapping, it is affected by the non-uniformity of bone tissue, changes in end-effector state, and coupling of process parameters, making it difficult to accurately model using analytical methods. Therefore, data-driven methods are needed to approximate it, providing support for subsequent high-precision force control and process optimization.
[0045] A six-dimensional force prediction model based on radial basis function neural network (RBFNN) is constructed, where the input layer contains three grinding process parameters (bone mineral density). Spindle speed Feed rate ) and 6 end pose parameters (position) , , and posture , , ). No. The input vector of each sample It can be represented as:
[0046]
[0047] The RBFNN model consists of an input layer, hidden layers, and an output layer, and its model structure is as follows: Figure 2 As shown, the hidden layer achieves nonlinear transformation through the radial basis function (RBF) and uses a Gaussian function as the activation function. For the input layer... training samples In its first The output of each hidden layer node The calculation is as follows:
[0048]
[0049] in, and Represented as the first Cluster centers and standard deviations of Gaussian activation functions For input samples With the center Euclidean distance.
[0050] Regarding cluster centers, the following approach is adopted: The mean clustering algorithm determines the center of the hidden layer neurons, as shown below:
[0051]
[0052] in, For the first A set of samples for each cluster. This represents the number of samples.
[0053] The output layer sums and weights the information from the hidden layer neurons and outputs it, which can be represented as:
[0054]
[0055] in, The number of radial basis function units in the hidden layer. For the input vector, For the first The output value of each output neuron For the hidden layer The nth neuron to the output layer The weights of each neuron, For the output layer Bias of each neuron.
[0056] Step 102: Based on the clinical trial requirements for epiphyseal plate bone bridge grinding, construct a bone grinding contact force experimental dataset covering the space of end pose, grinding process parameters, and bone density parameters to provide data support for training the contact force prediction model.
[0057] To effectively train and validate the constructed RBFNN-GA model, real bone grinding experimental data is required. Given that the mechanical and thermal properties of bone tissue are affected by factors such as age and gender, to ensure the consistency and reproducibility of experimental results, this invention uses SAWBONES solid rigid polyurethane foam as the standard artificial bone material, employing bone blocks with densities of 800 kg / m³ and 1640 kg / m³ to simulate cancellous and cortical bone, respectively. Considering the small size of pediatric epiphyseal plate bone bridge lesions and the extremely high requirements for surgical safety, to avoid irreversible damage to bone tissue caused by excessive grinding load, the spindle speed was limited to 12,000-30,000 rpm, and the feed rate was controlled within the range of 0.3-1 mm / s, with systematic grinding experiments conducted within this parameter range.
[0058] The bone grinding experiment process is as follows Figure 3 As shown. A coordinate system is established for the robot's end effector spherical drill: the X-axis is perpendicular to the surface of the bone being ground, the Y-axis is the feed direction, and the Z-axis is perpendicular to the XY plane. In the experiment, the feed speed of the spherical drill along the Y-axis is... Spindle speed of spherical drill The robot moves the milling unit perpendicular to the bone surface along a straight path at a fixed linear speed, maintaining a grinding depth of 1 mm during the process, while simultaneously acquiring the original force / torque signal through a six-dimensional force sensor.
[0059] To construct the training dataset for the neural network model, this study employed a full-factor experimental design within the range of commonly used clinical parameters, systematically combining the levels of each parameter to form the experimental matrix shown in Table 1. The experimental order was randomized, and each parameter combination was repeated three times to improve the stability and reproducibility of the measurement results. Furthermore, to more objectively evaluate the generalization performance of the model, a separate test experiment was designed, as shown in Table 2, to construct the validation dataset.
[0060] Table 1. Full Factorial Experiment Design Matrix (Training Set)
[0061] Table 2 Test Set Experimental Parameter Design
[0062] Bone mineral density 800 / 1640 kg / m³ Spindle speed (rpm) Feed rate (mm / s) Experiment 1 20000 0.3 Experiment 2 24000 0.4 Experiment 3 18000 0.6 Experiment 4 28000 0.7 Experiment 5 25000 0.9
[0063] To improve the reliability and analysis effect of the signals acquired by the end-effector six-dimensional force sensor, data preprocessing is performed on the raw signals to improve data quality and reduce the impact of measurement noise on subsequent modeling and analysis. First, detrending processing is performed to eliminate baseline offset caused by sensor zero-point drift and environmental interference. Second, a Butterworth low-pass filter is used to filter out high-frequency noise components. Finally, a moving average algorithm is used for smoothing to enhance the temporal stability and readability of the signal. Taking a feed rate of 0.5 mm / s and a spindle speed of 12000 rpm as an example... Figure 4 (a) and Figure 4 (b) shows a comparison of the end force / torque signals before and after preprocessing. This processing flow can accurately reflect the mechanical characteristics of the grinding process, providing reliable data support for the subsequent construction of a high-precision robot bone grinding contact force prediction model.
[0064] Step 103: Based on the constructed bone grinding dataset, the key parameters of the radial basis function neural network are optimized and trained using a genetic algorithm to achieve accurate prediction of the six-dimensional force / torque at the end.
[0065] The prediction accuracy of RBFNN is highly dependent on the number of neurons in the hidden layer. and Gaussian function extension parameters The selection of parameters is crucial. Traditional gradient learning methods are prone to getting stuck in local optima when optimizing these parameters, making it difficult to guarantee a globally optimal solution. Therefore, a genetic algorithm (GA) is introduced to globally optimize the key parameters of the RBFNN, achieving joint modeling of GA-RBFNN. The optimization process is as follows: Figure 5 As shown.
[0066] The optimization objective is to minimize the root mean square error (RMSE) of the contact force prediction. Fitness function It can be represented as:
[0067]
[0068] in, For the sample size, and Represented as the first The true end force and the network model predicted values of the sample, and the squared Euclidean norm of the error vector between the true and predicted values, characterize the sum of squared force errors of the predicted six-dimensional end forces, which can be expressed as:
[0069]
[0070] Genetic algorithms iteratively update the population through selection, crossover, and mutation operations, searching for the optimal population globally. and Parameter combinations to minimize the fitness function The RBFNN model optimized by GA exhibits higher prediction accuracy and generalization ability under conditions of strong nonlinearity and parameter uncertainty, providing reliable model support for subsequent contact dynamics analysis and force control strategy design.
[0071] Based on the constructed dataset, the proposed GA-RBFNN hybrid model was trained and validated. GA was used to globally optimize key structural parameters of the RBFNN, including the number of hidden layer neurons (radial basis centers) and the Gaussian kernel width. The optimization process is detailed below. Figure 5 The key parameters during optimization are set as follows: population size is set to 30-50 individuals, number of generations is 80, crossover probability is 0.85, mutation probability is 0.15; radial basis center number search range is 10-80, Gaussian kernel width optimization interval is [1.0, 50.0]; two optimal individuals are retained in each generation to maintain good genes and accelerate convergence. After GA completes the global search, gradient optimization is used for local fine-tuning to improve the quality of the solution and compensate for the shortcomings of the global search of the genetic algorithm.
[0072] Changes in fitness during training, such as Figure 6 The optimal fitness and average fitness steadily increase with the number of generations, both converging to a high level. The optimal fitness consistently leads the average fitness, indicating that the algorithm has good convergence and optimization efficiency. The optimized parameters obtained are: the optimal value of the Gaussian kernel width of the radial basis function. The optimal number of neurons in the hidden layer The determined RBFNN model structure is as follows: the radial basis center matrix has a dimension of 52×9, the width parameter vector has a dimension of 52×1, and the output layer weight matrix has a dimension of 53×6. The additional 1 dimension of the output layer corresponds to the bias term, and the output is a six-dimensional grinding force / torque vector at the end.
[0073] To verify the model's generalization performance, an independent test dataset not used in the training process was used to evaluate the model's predictions. The model's average coefficient of determination (R²) for the six-dimensional contact force / torque predictions on the test set was 0.962. Comparisons between the predicted and measured values for each component are shown below. Figure 7 As shown, grinding force components , , and torque components , , All results showed high correlation, and the prediction curves were consistent with the overall trend of the experimental measurements, with no obvious systematic deviation, demonstrating excellent cross-condition prediction capabilities.
[0074] To evaluate the effectiveness of the bone grinding mechanics model, three indicators—root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²)—were used for quantitative evaluation to comprehensively measure the consistency between the model's predicted results and the actual measured values. Their calculation can be expressed as follows:
[0075]
[0076]
[0077]
[0078] To comprehensively evaluate the performance of the constructed multi-output bone grinding force prediction model, detailed evaluation metrics for the six output components on the test set are listed. From the univariate prediction results, the model demonstrates good fitting ability and predictive consistency for grinding force and torque components in all directions. Specifically, , , and , , The mean R² values were 0.996, 0.907, 0.970, 0.966, 0.956, and 0.976, respectively, remaining at a high level overall, indicating that the model has a stable ability to capture outputs in different directions. From the error perspective, the RMSE and MAE values corresponding to each component are generally small, especially for the torque component. , , The RMSE values were only 0.012, 0.006, and 0.008, respectively, and the MAE values were as low as 0.008, 0.004, and 0.005, maintaining high absolute accuracy even in predictions of minute magnitudes. , , The RMSEs were 0.362, 0.527, and 0.288, respectively, and the MAEs were 0.267, 0.391, and 0.217, respectively. The error of each component is relatively high, but its R² still reaches 0.907, indicating that the model can reflect the changing trend well. Overall, the average coefficient of determination of each component on the test set reaches 0.962, further verifying the stability and generalization ability of the model in multi-output prediction tasks. In summary, this force prediction model has both high accuracy and good consistency in multi-dimensional output tasks. The prediction results of each component are in high agreement with the measured values, and it has high reliability and engineering application value.
[0079] Table 3. Model evaluation metrics for each component on the test set.
[0080] Evaluation indicators RMSE MAE <![CDATA[R 2 ]]> 0.362 0.267 0.996 0.527 0.391 0.907 0.288 0.217 0.970 0.012 0.008 0.966 0.006 0.004 0.956 0.008 0.005 0.976
[0081] The GA-RBFNN model constructed in this invention is compared with existing bone grinding force modeling methods to evaluate its predictive performance. The comparison results of each model on the three indicators of RMSE, MAE, and R² are as follows: Figure 8 As shown, the results indicate that the method of the present invention outperforms other comparative methods in all three evaluation indicators, exhibiting higher prediction accuracy and stronger fitting ability, with a lower overall error level, thus verifying its superiority and effectiveness in bone grinding contact force modeling tasks.
Claims
1. A method for modeling contact dynamics in robot bone grinding, characterized in that, The method includes: Based on the contact dynamics between the end drill bit and bone tissue during bone grinding, a six-dimensional contact force prediction model based on radial basis function neural network is constructed to describe the nonlinear mapping relationship between grinding process parameters, end position and bone density and contact force / torque. Based on the clinical trial requirements for epiphyseal plate bone bridge grinding, a bone grinding contact force experimental dataset covering the space of end pose, grinding process parameters, and bone density parameters was constructed to provide data support for the training of contact force prediction models. Based on the constructed bone grinding dataset, the key parameters of the radial basis function neural network are optimized and trained using a genetic algorithm to achieve accurate prediction of the six-dimensional force / torque at the end.
2. The method according to claim 1, characterized in that, Based on the contact dynamics between the end-point drill bit and bone tissue during bone grinding, a six-dimensional contact force prediction model based on a radial basis function neural network is constructed to describe the nonlinear mapping relationship between grinding process parameters, end-point pose, bone density, and contact force / torque, including: The contact dynamics between bone grinding drill bits and bone tissue were analyzed, and a six-dimensional force prediction model based on radial basis function neural networks (RBFNN) was constructed. The input layer includes three grinding process parameters (bone density, bone mineral density, and bone tissue density). Spindle speed Feed rate ) and 6 end pose parameters (position) , , and posture , , ), No. The input vector of each sample It can be represented as: The RBFNN model consists of an input layer, hidden layers, and an output layer. For the input layer... training samples In its first The output of each hidden layer node The calculation is as follows: in, and Represented as the first Cluster centers and standard deviations of Gaussian activation functions For input samples With the center Euclidean distance; The output layer sums and weights the information from the hidden layer neurons and outputs it, which can be represented as: in, The number of radial basis function units in the hidden layer. For the input vector, For the first The output value of each output neuron For the hidden layer The nth neuron to the output layer The weights of each neuron, For the output layer Bias of each neuron.
3. The method according to claim 1, characterized in that, Based on the clinical trial requirements for epiphyseal plate bone bridge grinding, a bone grinding contact force experimental dataset covering distal end pose, grinding process parameters, and bone mineral density parameter space was constructed to provide data support for training the contact force prediction model, including: To train and validate the grinding force model, real bone grinding experimental data need to be collected. SAWBONES solid polyurethane foam was selected as the standard artificial bone material, and bone blocks with densities of 800 kg / m³ and 1640 kg / m³ were used to simulate cancellous bone and cortical bone, respectively. Considering the requirements of load and safety in grinding pediatric epiphyseal plate bone bridges, the spindle speed was set to 12000-30000 rpm and the feed rate was controlled within the range of 0.3-1 mm / s. Grinding experiments were carried out within this parameter range. To improve the reliability of the six-dimensional force sensor signal at the end, this paper preprocesses the raw data. First, detrending processing is used to eliminate zero drift and baseline offset caused by environmental interference. Then, a Butterworth low-pass filter is used to filter out high-frequency noise. Finally, a moving average algorithm is used for smoothing to improve signal stability.
4. The method according to claim 1, characterized in that, Based on the constructed bone grinding dataset, a genetic algorithm is used to optimize and train the key parameters of the radial basis function neural network, achieving accurate prediction of the six-dimensional force / torque at the distal end, including: The prediction accuracy of RBFNN is highly dependent on the number of neurons in the hidden layer. and Gaussian function extension parameters To select the parameters, a genetic algorithm (GA) is introduced to globally optimize the key parameters of RBFNN, thereby achieving joint modeling of GA-RBFNN. The optimization objective is to minimize the root mean square error of the contact force prediction, and the fitness function can be expressed as: in, For the sample size, and Represented as the first The actual end force of the sample versus the predicted value of the network model; Genetic algorithms iteratively update the population through selection, crossover, and mutation operations, searching for the optimal population globally. and Parameter combinations to minimize the fitness function The RBFNN model optimized by GA exhibits higher prediction accuracy and generalization ability under strong nonlinearity and parameter uncertainty conditions, and can accurately learn the nonlinear mapping relationship between contact force and multi-source input variables during bone grinding.